Equipment failure rate evaluation method for reliability calculation of power distribution network

Abstract

The invention discloses an equipment failure rate evaluation method for reliability calculation of a power distribution network, and the method comprises the steps: S1, collecting equipment parameters, constructing a model, recognizing the influence of equipment aging, environmental factors and maintenance conditions on a failure rate, and classifying failure factors; s2, constructing an equipment failure rate evaluation model, and calculating an equipment comprehensive failure rate; s3, constructing a weight decomposition model, performing weight decomposition according to features, and quantifying the influence of each factor on the fault rate; and S4, counting the fault rate of each region, decomposing the region fault rate to each grid based on the network weight coefficient, and decomposing the grid fault rate to each feeder based on the feeder weight coefficient. Through the multi-level weight decomposition method, accurate evaluation of the equipment failure rate is realized, and the technical effects of simplifying the structure, reducing the cost, prolonging the service life, improving the accuracy, realizing real-time response and the like are achieved.

Description

Technical Field

[0001] The invention relates to the field of power system reliability analysis, and in particular to an equipment failure rate evaluation method for distribution network reliability calculation. Background Art

[0002] The health status of power equipment is crucial to the safe operation of the power grid, and the reliability assessment of equipment has always been an important topic in the power industry. Existing assessment methods mainly focus on the evaluation of the current status of the equipment, often through direct phenomena of quantifying the operating status of the equipment or indirect parameters obtained by experimental means, combined with expert experience and standard specifications for evaluation. However, these methods have several shortcomings, including the fact that the deduction value is not detailed and accurate enough to represent the equipment status, the diagnostic threshold is based on expert experience and lacks statistical support, and the status evaluation results based on the deduction value are not equivalent enough. In recent years, researchers have tried to introduce more field data, status parameters and failure factors to improve the accuracy of failure rate assessment, especially through statistical models and probabilistic models to conduct more scientific quantitative analysis of equipment failure rate. However, traditional models often cannot accurately describe the heterogeneity and dynamic risks of equipment.

[0003] In the existing research patents on the determination of distribution network failure rate, although the method has improved the traditional evaluation model to a certain extent, there are still some technical shortcomings:

[0004] (1) Complex structure: Some patented methods involve multi-step calculations and model adjustments, which makes the implementation process complicated and difficult to quickly deploy and operate in practical applications.

[0005] (2) High cost: Since a large number of sensor data collection and real-time monitoring equipment are required, the implementation and maintenance costs are high, especially in large-scale power grid environments.

[0006] (3) Short lifespan: Some methods rely on specific hardware devices or sensors. As technology advances and equipment is updated, the applicable lifespan of the relevant hardware is short and requires frequent replacement and upgrading.

[0007] (4) Strong data dependence: Some patents rely too much on historical data and field data. When the data volume is insufficient or the data quality is low, the accuracy of the evaluation results will be greatly reduced.

[0008] (5) Untimely response: Many existing methods lack a real-time dynamic update mechanism, which results in the inability to reflect changes in equipment status in a timely manner, delaying fault warning and processing opportunities. Summary of the invention

[0009] In order to overcome the deficiencies in the prior art, the present invention provides a method for evaluating equipment failure rate for distribution network reliability calculation, which solves the problems of complex structure, high cost, short life, strong data dependence and untimely response existing in the prior art.

[0010] In order to achieve the above-mentioned invention object, the technical solution adopted to solve the technical problem is as follows:

[0011] A method for evaluating equipment failure rate for distribution network reliability calculation comprises the following steps:

[0012] Step S1: Collect equipment parameters, build a model to identify the impact of equipment aging, environmental factors and maintenance on failure rate, and classify failure factors;

[0013] Step S2: construct an equipment failure rate evaluation model to calculate the comprehensive failure rate of the equipment;

[0014] Step S3: construct a weight decomposition model, perform weight decomposition according to features, and quantify the impact of each factor on the failure rate;

[0015] Step S4: Count the failure rates of each region, decompose the regional failure rates into each grid based on the network weight coefficient, and decompose the grid failure rates into each feeder based on the feeder weight coefficient.

[0016] Furthermore, in step S1, regional equipment inventory data and equipment operation data are collected, and the types of fault factors are analyzed and classified. The collected data include equipment type, historical fault records, years of operation, manufacturers, and geographical area information, and the main factors affecting the equipment failure rate are identified through statistical analysis and data mining technology. The main process of this step is data preprocessing, fault factor identification, fault factor classification and classification result statistics. The main contents of data preprocessing include:

[0017] Comprehensively collect all kinds of data related to medium voltage distribution network equipment to ensure the integrity and accuracy of the data; then perform data preprocessing to ensure the effectiveness and accuracy of subsequent analysis;

[0018] Eliminate duplicate records, fill in or delete missing values, and correct errors and outliers in the data;

[0019] Perform uniform standardization on different types of data to ensure that all variables are compared and analyzed under the same dimension.

[0020] Furthermore, in step S1, fault factors are identified based on the collected data, and the main contents include:

[0021] Based on the collected data, the main factors affecting the equipment failure rate are identified and classified, mainly based on correlation analysis and regression analysis. The analysis results serve as the basis for subsequent classification:

[0022] Correlation analysis: The Pearson correlation coefficient is used to measure the linear correlation between each factor and the failure rate. By calculating the correlation coefficient between each factor and the failure rate, the factors significantly correlated with the failure rate are preliminarily screened out:

[0023]

[0024] Among them, X and Y represent the research variables; r XY is the correlation coefficient between two variables, ranging from -1 to 1;

[0025] Regression analysis: Use multiple linear regression or logistic regression models to further analyze the combined impact of multiple factors on the failure rate and determine the degree and significance of each factor's impact on the failure rate:

[0026] Y=β 0 +β 1 X 1 +β 2 X 2 +···+β n X n +∈

[0027] Among them, Y is the failure rate, X is 1 ,X 2 ,···,X n is the influencing factor, β 0 is a constant term, β 1 ,β 2 ,···,β n is the regression coefficient; ∈ is the error term, which obeys the normal distribution.

[0028] Furthermore, in step S1, fault factor classification and cluster analysis are performed on the collected data, and the main contents include:

[0029] After identifying the main failure factors, they are classified according to different dimensions to facilitate the subsequent model construction and failure rate calculation;

[0030] Frequency analysis: Count the occurrence frequencies of various fault factors in historical data, identify high-frequency fault factors, and give priority to fault factors with higher occurrence frequencies as the main influencing factors;

[0031]

[0032] Among them, f i is the frequency of the ith fault factor, N i is the number of occurrences of the i-th fault factor, N total is the total number of occurrences of all fault factors;

[0033] Equipment type: For medium voltage distribution network, the main equipment involved includes distribution transformers, overhead lines, cables, upstream power supplies, circuit breakers, load switches, disconnectors, etc. The reliability of these equipment directly affects the overall performance of the distribution network;

[0034] Operation life: refers to the time that various types of equipment have been in use since they were put into operation. The operation life is divided into four time periods: 0-5 years, 5-10 years, 10-15 years, and more than 15 years. Equipment in different time periods has different failure risks;

[0035] Manufacturer: The manufacturer of the equipment is another important factor affecting the failure rate. If there are too many manufacturers in a certain area, the manufacturers that account for more than 10% will be mainly considered, and other more scattered manufacturers will be classified as "others";

[0036] Geographical area: The geographical distribution of equipment refers to the distribution of each type of equipment in a region. Geographical areas are divided into three levels: feeder, grid, and region. Several feeders form a grid, and several grids form a region. Therefore, by counting the distribution of equipment in feeders, the distribution of equipment in grids and regions can be derived.

[0037] Cluster analysis: After the classification is completed, cluster analysis is used to refine the classified equipment, and K-means clustering is used to classify fault factors with similar characteristics into the same category to facilitate subsequent classification and analysis.

[0038] Furthermore, the cluster analysis includes the following contents:

[0039] (1) Initialization: Select k initial cluster centers;

[0040] (2) Assignment: Assign each data point to the cluster to which the nearest cluster center belongs;

[0041] (3) Update: recalculate the cluster center of each cluster;

[0042] (4) Iteration: Repeat steps (2) and (3) until the cluster center no longer changes significantly or the preset number of iterations is reached.

[0043] Furthermore, in step S2, based on the Bayesian neural network, a device failure rate prediction model is constructed to make a preliminary prediction of the failure rate of each type of equipment, including the following contents;

[0044] The structure of the Bayesian neural network consists of an input layer, a hidden layer, and an output layer. The hidden layer in the Bayesian neural network is a probability layer, which introduces uncertainty into the weight of the Bayesian layer of the i+1th layer by sampling from the distribution of the Bayesian layer of the ith layer.

[0045] In the parameter sampling process of Bayesian neural network, the weight calculation formula is:

[0046]

[0047] The formula for calculating the bias is:

[0048]

[0049] in, The weights and biases of the linear transformation of the Bayesian layer corresponding to the i-th layer with n samples, where ρ i is the standard deviation of the sample, μ i is the average value of the sample, as standard;

[0050] By constructing different Bayesian linear layers, the posterior distribution p(ω|X,Y) on the model parameters ω of the training set data Y and x is inferred, and the formula is:

[0051]

[0052] After completing the fitting of the training set, the posterior distribution p(ω|X,Y) exists, and there is a new data point X pre When the input model is integrated over ω, we can get X pre The prediction distribution of is as follows:

[0053] p(Y pre |X pre ,X,Y)=∫p(Y pre |X pre ,ω)p(ω|X,Y)dω

[0054] In the prediction process, a sample n is drawn from the posterior distribution p(ω|X,Y) each time, and a new data point X is obtained for a given n. pre Point estimate Y pre The above operation is repeated many times, thus obtaining a new data point X pre Estimation of the distribution of .

[0055] Furthermore, in step S2, the Bayesian neural network is trained and parameter optimized by setting a loss function and an optimizer, including the following:

[0056] The loss function of BNN not only needs to measure the difference between the predicted failure rate and the actual failure, but also needs to consider the modeling of parameter uncertainty in Bayesian reasoning, using the negative log-likelihood loss function:

[0057] L(θ)=-∑logP(Y|X,θ)

[0058] Where Y is the actual fault label (0 or 1); X is the input feature of the device; θ is the parameter weight and bias of the network; P(Y|X,θ) is the probability distribution predicted by the model given the parameter θ;

[0059] In order to minimize the loss function, the Adam optimizer is used. As a gradient descent optimization method, it can adaptively adjust the learning rate and is suitable for handling the complexity of Bayesian neural networks. The weight update formula of Adam is:

[0060] m t =β 1 m t-1 +(1-β 1 ) t

[0061]

[0062]

[0063] Among them, g t is the gradient of the loss function L(θ) under the current weight; β 1 Momentum factor, used to update the smooth gradient to prevent the gradient from fluctuating too much. The common default value is β 1 = 0.9, that is, 90% of the new gradient information and 10% of the past information are weighted averaged; m t is the first-order momentum of the gradient at the current moment, similar to the weighted average; v t It is a second-order moment estimation. Similar to the first-order moment, Adam uses the square of the gradient to calculate the second-order moment, which is the moving average of the square of the gradient to balance the uneven size of the gradient update. It is for m t and v t Adam combines the corrected gradient momentum and gradient variance estimate to update the weight W t ; η is the learning rate, which determines the size of the weight adjustment at each update; ∈ is a very small constant used to avoid division by zero errors.

[0064] Furthermore, in step S3, the Bayesian neural network is used to decompose the equipment failure rate into weights of different failure factors. By decomposing the weights of various equipment failure factors, the influence of each factor on the equipment failure rate can be accurately identified;

[0065] The equipment failure rate taking failure factors into consideration can be expressed as:

[0066] R=∑W i ·f i

[0067] Among them, W i is the corresponding fault factor fi The weight of

[0068] For the weight W, the fault factor is taken as input and the weight distribution of each fault factor is taken as output to construct a BNN weight decomposition model. At this time, the simple structure of the model is: input layer-Bayesian layer-output layer.

[0069] Furthermore, in step S3, the equipment failure rate is decomposed into weights of different failure factors using a Bayesian neural network. The training steps of the BNN weight decomposition model are as follows:

[0070] (1) Determine the likelihood function, given the training data Define the likelihood function p(D|ω):

[0071]

[0072] Among them, x i is the input feature vector, including the equipment operation years, equipment region, and equipment manufacturer, y i is the output failure rate, ω is the network weight;

[0073] (2) Determine the posterior distribution and calculate the posterior distribution p(ω|D):

[0074]

[0075] Since p(D) is a constant, it can be ignored and simplified to:

[0076] p(ω|D)∝p(D|ω)p(ω)

[0077] Variational inference approximates the posterior distribution p(ω|D) to a tractable distribution q(ω) and finds the optimal approximate solution by minimizing the difference between the two distributions. The specific steps are:

[0078] Choose an approximate distribution: Choose an easy-to-handle distribution q(ω) to approximate the posterior distribution p(ω|D). Use a Gaussian distribution:

[0079] q(ω)=N(μ,σ 2 )

[0080] Among them, μ is the mean, σ 2 is the variance;

[0081] Define the variational objective: Use the evidence lower bound (ELBO) to approximate the optimization objective;

[0082] ELBO=E q(ω) [LogP(D|ω)]-KL(q(ω)||P(ω))

[0083] Among them, the first term is the expected likelihood function, that is, the degree of fit of the expected model to the observed data; the second term is the Kullback-Leibler divergence, which is used to measure the difference between the approximate distribution q(ω) and the prior distribution P(ω);

[0084] Gradient optimization: Use the gradient descent method to update the parameters of q(ω) to maximize the ELBO;

[0085] When ELBO converges or the optimization reaches the preset number of iterations, the variational inference ends, and q(ω) at this time is considered to be the optimal approximation of the posterior distribution;

[0086] Iterate prediction and output weight result W i .

[0087] Furthermore, in step S4, the weight distribution obtained by the BNN model is applied to the actual failure rate data, and the failure rate is decomposed layer by layer. The calculation formula is as follows:

[0088] R i =W age ·f age +W reg ·f reg +W maf ·f maf

[0089] Among them, R i is the failure rate of type i equipment, W age is the weight of the equipment’s operating life, W reg is the weight of the area to which the device belongs, W maf is the weight of the equipment manufacturer; f age is the failure rate level of the equipment during its service life, f reg is the failure rate level of the area to which the equipment belongs, f maf is the failure rate level of the equipment manufacturer, f age ,f reg ,f maf Based on the calculation in step S2, W age ,W reg ,W maf Calculated based on step S3;

[0090] Based on the grid weight coefficient, the regional failure rate is decomposed into each grid:

[0091] λ grid =λ region ·W grid

[0092] Among them, W grid is the grid weight coefficient:

[0093]

[0094] In the above formula, N grid is the number of devices in the grid, f BNN is the weight decomposition model, x i is the feature vector of the device in the grid;

[0095] Based on the feeder weight coefficient, the grid failure rate is decomposed into each feeder:

[0096] λ freeder =λ grid ·W freeder

[0097] Among them, W freeder The calculation method for the grid weight coefficient is:

[0098]

[0099] In the above formula, N freeder is the number of devices in the feeder, f BNN is the weight decomposition model, x i is the eigenvector of the device in the feeder.

[0100] Due to the adoption of the above technical solution, the present invention has the following advantages and positive effects compared with the prior art:

[0101] 1. The present invention utilizes the adaptive learning capability of machine learning to simplify the complex calculation steps in the traditional model, thereby reducing the difficulty of implementation and the complexity of operation.

[0102] 2. The present invention reduces the reliance on a large number of sensors and real-time monitoring equipment by optimizing the data collection and processing process, thereby reducing the implementation and maintenance costs.

[0103] 3. The present invention adopts a more general software algorithm to reduce the dependence on specific hardware devices, improves the adaptability and sustainability of the method, and extends the service life of the overall solution.

[0104] 4. The present invention dynamically updates and optimizes the equipment failure rate, thereby improving the accuracy of the evaluation results, especially when the data volume is insufficient or the quality is not high.

[0105] 5. The present invention introduces a real-time data dynamic update mechanism, so that the model can reflect changes in equipment status in a timely manner, thereby enhancing the timeliness of fault warning and processing. BRIEF DESCRIPTION OF THE DRAWINGS

[0106] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative work. In the drawings:

[0107] Figure 1 It is an overall flow chart of a method for evaluating equipment failure rate for distribution network reliability calculation of the present invention;

[0108] Figure 2 It is a basic framework flow chart of the BNN failure rate prediction model of the present invention;

[0109] Figure 3 It is a structural schematic diagram of the BNN model established by the present invention;

[0110] Figure 4 It is a basic framework flow chart of the equipment failure rate weight decomposition model of the present invention. DETAILED DESCRIPTION

[0111] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0112] like Figure 1 As shown, this embodiment discloses a method for evaluating equipment failure rate for distribution network reliability calculation, comprising the following steps:

[0113] Step S1: Collect equipment parameters, build a model to identify the impact of equipment aging, environmental factors and maintenance on failure rate, and classify failure factors;

[0114] Specifically, in step S1, regional equipment ledger data and equipment operation data are collected, and the types of fault factors are analyzed and classified. The collected data include equipment type, historical fault records, years of operation, manufacturers, and geographical area (feeder, grid, region) information, and the main factors affecting the equipment failure rate are identified through statistical analysis and data mining technology. The main process of this step is data preprocessing, fault factor identification, fault factor classification and classification result statistics:

[0115] (1) Data preprocessing:

[0116] Comprehensively collect all kinds of data related to medium voltage distribution network equipment to ensure the integrity and accuracy of the data; then perform data preprocessing to ensure the effectiveness and accuracy of subsequent analysis;

[0117] Data cleaning: remove duplicate records, fill or delete missing values, and correct errors and outliers in the data;

[0118] Data standardization: Perform unified standardization on different types of data to ensure that all variables are compared and analyzed under the same dimension;

[0119] (2) Fault factor identification:

[0120] Based on the collected data, the main factors affecting the equipment failure rate are identified and classified, mainly based on correlation analysis and regression analysis. The analysis results serve as the basis for subsequent classification:

[0121] Correlation analysis: The Pearson Correlation Coefficient is used to measure the linear correlation between each factor and the failure rate. By calculating the correlation coefficient between each factor and the failure rate, the factors significantly correlated with the failure rate are preliminarily screened out:

[0122]

[0123] Among them, X and Y represent research variables (such as years of operation and failure rate); r XY is the correlation coefficient between two variables, ranging from -1 to 1;

[0124] Regression analysis: Use multiple linear regression or logistic regression models to further analyze the combined impact of multiple factors on the failure rate and determine the degree and significance of each factor's impact on the failure rate:

[0125] Y=β 0 +β 1 X 1 +β 2 X 2 +···+β n X n +∈

[0126] Among them, Y is the failure rate, X is 1 ,X 2 ,···,X n is the influencing factor (such as years of operation, manufacturer), β 0 is a constant term, β 1 ,β 2 ,···,β n is the regression coefficient;∈ is the error term, which usually obeys the normal distribution;

[0127] (3) Fault factor classification and cluster analysis:

[0128] After identifying the main failure factors, they are classified according to different dimensions to facilitate the subsequent model construction and failure rate calculation;

[0129] Frequency analysis: Count the occurrence frequencies of various fault factors in historical data, identify high-frequency fault factors, and give priority to fault factors with higher occurrence frequencies as the main influencing factors;

[0130]

[0131] Among them, f i is the frequency of the ith fault factor, N i is the number of occurrences of the i-th fault factor, N total is the total number of occurrences of all fault factors;

[0132] Equipment type: For medium voltage distribution network, the main equipment involved includes distribution transformers, overhead lines, cables, upstream power supplies, circuit breakers, load switches, disconnectors, etc. The reliability of these equipment directly affects the overall performance of the distribution network;

[0133] Operation life: refers to the time that various types of equipment have been in use since they were put into operation. The operation life is divided into four time periods: 0-5 years, 5-10 years, 10-15 years, and more than 15 years. Equipment in different time periods may have different failure risks;

[0134] Manufacturer: The manufacturer of the equipment is another important factor affecting the failure rate. If there are too many manufacturers in a certain area, the manufacturers that account for more than 10% will be mainly considered, and other more scattered manufacturers will be classified as "others";

[0135] Geographical area: The geographical distribution of equipment refers to the distribution of each type of equipment in a region. The geographical area in this embodiment is divided into three levels: feeder, grid, and region. Several feeders form a grid, and several grids form a region. Therefore, by counting the distribution of equipment in feeders, the distribution of equipment in grids and regions can be derived.

[0136] Cluster analysis: After classification, cluster analysis is used to refine the classified equipment in order to further optimize the model performance. Clustering is used to discover the feature similarities of equipment in each category, so as to facilitate the mining of more potential relationships within the same category and simplify the subsequent model processing. K-means clustering is used to classify fault factors with similar features into the same category for subsequent classification and analysis.

[0137] Furthermore, the cluster analysis includes the following contents:

[0138] (1) Initialization: Select k initial cluster centers;

[0139] (2) Assignment: Assign each data point to the cluster to which the nearest cluster center belongs;

[0140] (3) Update: recalculate the cluster center of each cluster;

[0141] (4) Iteration: Repeat steps (2) and (3) until the cluster center no longer changes significantly or the preset number of iterations is reached.

[0142] Step S2: construct an equipment failure rate evaluation model to calculate the comprehensive failure rate of the equipment;

[0143] Specifically, in step S2, based on the Bayesian Neural Network (BNN), a device failure rate prediction model is constructed to make a preliminary prediction of the failure rate of various types of equipment. The Bayesian Neural Network is one of the important results of the combination of the Bayesian method and the deep learning method, and has been widely used. The process of the present invention for predicting the device failure rate is as follows: Figure 2 As shown, including the following:

[0144] (1) Constructing a Bayesian neural network

[0145] The structure of the Bayesian neural network is similar to that of a deep neural network, consisting of an input layer, a hidden layer, and an output layer. Unlike ordinary deep neural networks, the hidden layer (Bayesian layer) in the Bayesian neural network is a probability layer. The Bayesian layer (Bayesian layer) is sampled from the distribution of the Bayesian layer of the i-th layer, thereby introducing uncertainty in the weight of the Bayesian layer of the i+1th layer.

[0146] In the parameter sampling process of Bayesian neural network, the weight calculation formula is:

[0147]

[0148] The formula for calculating the bias is:

[0149]

[0150] in, The weights and biases of the linear transformation of the Bayesian layer corresponding to the i-th layer with n samples, where ρ i is the standard deviation of the sample, μ i is the average value of the sample, as standard;

[0151] By constructing different Bayesian linear layers, the posterior distribution p(ω|X,Y) on the model parameters ω of the training set data Y and X is inferred, and the formula is:

[0152]

[0153] After completing the fitting of the training set, the posterior distribution p(ω|X,Y) exists, and there is a new data point X pre When the input model is integrated over ω, we can get X pre The prediction distribution of is as follows:

[0154] p(Y pre |X pre ,X,Y)=∫p(Y pre |X pre ,ω)p(ω|X,Y)dω

[0155] In the prediction process, a sample n is drawn from the posterior distribution p(ω|X,Y) each time, and a new data point X is obtained for a given n. pre Point estimate Y pre The above operation is repeated many times, thus obtaining a new data point X pre Estimation of the distribution of ;.

[0156] In this embodiment, the structural design of the equipment failure rate prediction model has two depths, one input layer is the equipment failure factor, and the other is the Bayesian layer. The rectified linear unit (ReLU) is used as the activation function between the input layer and the Bayesian layer network. Finally, the output layer is the structure of the equipment failure rate prediction. The structural diagram is shown in FIG. Figure 3 shown.

[0157] (2) Setting the loss function and optimizer

[0158] The loss function of BNN not only needs to measure the difference between the predicted failure rate and the actual failure, but also needs to consider the modeling of parameter uncertainty in Bayesian reasoning, using the negative log-likelihood loss function:

[0159] L(θ)=-∑logP(Y|X,θ)

[0160] Where Y is the actual fault label (0 or 1); X is the input feature of the device; θ is the parameter weight and bias of the network; P(Y|X,θ) is the probability distribution predicted by the model given the parameter θ;

[0161] In order to minimize the loss function, the Adam optimizer is used. As a gradient descent optimization method, it can adaptively adjust the learning rate and is suitable for handling the complexity of Bayesian neural networks. The weight update formula of Adam is:

[0162] m t =β 1 m t-1 +(1-β 1 ) t

[0163]

[0164]

[0165] The Adam optimizer is an optimization algorithm that combines momentum method and adaptive learning rate, which is specially designed to efficiently update the weights of neural networks. t is the gradient of the loss function L(θ) under the current weight; β 1 Momentum factor, used to update the smooth gradient to prevent the gradient from fluctuating too much. The common default value is β 1 = 0.9, that is, 90% of the new gradient information and 10% of the past information are weighted averaged; m t is the first-order momentum of the gradient at the current moment, similar to the weighted average; v t It is a second-order moment estimation. Similar to the first-order moment, Adam uses the square of the gradient to calculate the second-order moment, which is the moving average of the square of the gradient to balance the uneven size of the gradient update. It is for m t and v t Adam combines the corrected gradient momentum and gradient variance estimate to update the weight W t ; η is the learning rate, which determines the size of the weight adjustment at each update; ∈ is a very small constant (such as 10 -8 ) to avoid division by zero errors.

[0166] Step S3: construct a weight decomposition model, perform weight decomposition according to features, and quantify the impact of each factor on the failure rate;

[0167] Specifically, in step S3, the Bayesian neural network (BNN) is used to decompose the equipment failure rate into the weights of different failure factors. By decomposing the weights of various equipment failure factors, the impact of each factor on the equipment failure rate can be accurately identified. The modeling process of the equipment failure rate weight decomposition model is as follows: Figure 4 ;

[0168] The equipment failure rate taking failure factors into consideration can be expressed as:

[0169] R=∑W i ·f i

[0170] Among them, W i is the corresponding fault factor f i The weight of

[0171] For the weight W, the fault factor is taken as input and the weight distribution of each fault factor is taken as output to construct a BNN weight decomposition model. At this time, the simple structure of the model is: input layer (including various fault factor characteristics of the equipment, such as equipment aging, environmental factors, maintenance conditions, etc.) - Bayesian layer - output layer (outputs the weight distribution of each fault factor).

[0172] The training steps of the BNN weight decomposition model are as follows:

[0173] (1) Determine the likelihood function, given the training data Define the likelihood function p(D|ω):

[0174]

[0175] Among them, x i is the input feature vector, including the equipment operation years, equipment region, and equipment manufacturer, y i is the output failure rate, ω is the network weight;

[0176] (2) Determine the posterior distribution and calculate the posterior distribution p(ω|D):

[0177]

[0178] Since p(D) is a constant, it can be ignored and simplified to:

[0179] p(ω|D)∝p(D|ω)p(ω)

[0180] (3) Variational inference: approximate the posterior distribution p(ω|D) to a tractable distribution q(ω) and find the optimal approximate solution by minimizing the difference between the two distributions. The specific steps are:

[0181] Choose an approximate distribution: Choose an easy-to-handle distribution q(ω) to approximate the posterior distribution P(ω|D). Use a Gaussian distribution:

[0182] q(ω)=N(μ,σ 2 )

[0183] Among them, μ is the mean, σ 2 is the variance, which is the parameter that needs to be optimized.

[0184] Define the variational objective: Use the evidence lower bound (ELBO) to approximate the optimization objective;

[0185] ELBO=E q(ω) [LogP(D|ω)]-KL(q(ω)||P(ω))

[0186] Among them, the first term is the expected likelihood function, that is, the degree of fit of the expected model to the observed data; the second term is the Kullback-Leibler (KL) divergence, which is used to measure the difference between the approximate distribution q(ω) and the prior distribution P(ω);

[0187] Gradient optimization: Use the gradient descent method to update the parameters of q(ω) to maximize the ELBO;

[0188] When ELBO converges or the optimization reaches the preset number of iterations, the variational inference ends, and q(ω) at this time is considered to be the optimal approximation of the posterior distribution;

[0189] Iterate prediction and output weight result W i .

[0190] Step S4: Count the failure rates of each region, decompose the regional failure rates into each grid based on the network weight coefficient, and decompose the grid failure rates into each feeder based on the feeder weight coefficient.

[0191] Specifically, in step S4, the weight distribution obtained by the BNN model is applied to the actual failure rate data, and the failure rate is decomposed layer by layer. The calculation formula is as follows:

[0192] R i =W age ·f age +W reg ·f reg +W maf ·f maf

[0193] Among them, R i is the failure rate of type i equipment, W age is the weight of the equipment’s operating life, W reg is the weight of the area to which the device belongs, W maf is the weight of the equipment manufacturer; f age is the failure rate level of the equipment during its service life, f reg is the failure rate level of the area to which the equipment belongs, f maf is the failure rate level of the equipment manufacturer, f age ,f reg ,f maf Based on the calculation in step S2, W age ,W reg ,W maf Calculated based on step S3;

[0194] Based on the grid weight coefficient, the regional failure rate is decomposed into each grid:

[0195] λ grid =λ region ·W grid

[0196] Among them, W grid is the grid weight coefficient:

[0197]

[0198] In the above formula, N grid is the number of devices in the grid, f BNN is the weight decomposition model, x i is the feature vector of the device in the grid;

[0199] Based on the feeder weight coefficient, the grid failure rate is decomposed into each feeder:

[0200] λ freeder =λ grid ·W freeder

[0201] Among them, W freeder The calculation method for the grid weight coefficient is:

[0202]

[0203] In the above formula, N freeder is the number of devices in the feeder, f BNN is the weight decomposition model, x i is the eigenvector of the device in the feeder.

[0204] This embodiment proposes a step-by-step weight decomposition method for evaluating the failure rate of distribution network equipment. This method collects historical equipment failure data, operating parameters, and environmental information, identifies and classifies failure factors, quantifies the impact weights of each factor, constructs a failure rate evaluation model, and verifies and optimizes it by comparing it with actual data. By decomposing step by step, from regions to grids, and then to feeders, the failure rate calculation is gradually refined, overcoming the problems of strong data dependence and incomplete consideration of failure factors in the prior art, making the evaluation results more scientific and reliable. Compared with the prior art, it has the following advantages:

[0205] 1. Multi-level weight decomposition: The step-by-step weight decomposition of the failure rate can more accurately quantify the impact of different failure factors on the equipment failure rate, avoiding the limitations of rough division and empirical parameters in traditional methods.

[0206] 2. Dynamic update and optimization: By comparing with actual operation data, the model is continuously optimized and adjusted to improve the real-time and accuracy of failure rate assessment.

[0207] 3. Comprehensive failure rate model: By integrating statistical methods and machine learning technology, a comprehensive failure rate model is established, which can more comprehensively reflect the failure risk of equipment under different conditions.

[0208] 4. Strong adaptability: This patented method is not only applicable to failure rate assessment at the equipment level, but can also be applied to failure rate assessment at the grid and regional levels by deducing step by step through the decomposition method, and has wider adaptability.

[0209] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed by the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A method for evaluating equipment failure rate for distribution network reliability calculation, characterized in that: The following steps are involved: Step S1: Collect equipment parameters, build a model to identify the impact of equipment aging, environmental factors and maintenance on failure rate, and classify failure factors; Step S2: construct an equipment failure rate evaluation model to calculate the comprehensive failure rate of the equipment; Step S3: construct a weight decomposition model, perform weight decomposition according to features, and quantify the impact of each factor on the failure rate; Step S4: Count the failure rates of each region, decompose the regional failure rates into each grid based on the network weight coefficient, and decompose the grid failure rates into each feeder based on the feeder weight coefficient.

2. The method for evaluating equipment failure rate for distribution network reliability calculation according to claim 1, characterized in that: In step S1, regional equipment ledger data and equipment operation data are collected, and the types of fault factors are analyzed and classified. The collected data include equipment type, historical fault records, years of operation, manufacturers, and geographical area information. The main factors affecting the equipment failure rate are identified through statistical analysis and data mining technology. The main process of this step is data preprocessing, fault factor identification, fault factor classification and classification result statistics. The main contents of data preprocessing include: Comprehensively collect all kinds of data related to medium voltage distribution network equipment to ensure the integrity and accuracy of the data; then perform data preprocessing to ensure the effectiveness and accuracy of subsequent analysis; Eliminate duplicate records, fill in or delete missing values, and correct errors and outliers in the data; Perform uniform standardization on different types of data to ensure that all variables are compared and analyzed under the same dimension.

3. The device failure rate evaluation method for distribution network reliability calculation according to claim 2 is characterized in that: In step S1, fault factors are identified based on the collected data, and the main contents include: Based on the collected data, the main factors affecting the equipment failure rate are identified and classified, mainly based on correlation analysis and regression analysis. The analysis results serve as the basis for subsequent classification: Correlation analysis: The Pearson correlation coefficient is used to measure the linear correlation between each factor and the failure rate. By calculating the correlation coefficient between each factor and the failure rate, the factors significantly correlated with the failure rate are preliminarily screened out: Among them, X and Y represent the research variables; r XY is the correlation coefficient between two variables, ranging from -1 to 1; Regression analysis: Use multiple linear regression or logistic regression models to further analyze the combined impact of multiple factors on the failure rate and determine the degree and significance of each factor's impact on the failure rate: Y=β0+β1X1+β2X2+…+β n X n +∈ Among them, Y is the failure rate, X1, X2, ···, X n is the influencing factor, β0 is a constant term, β1, β2, ···, β n is the regression coefficient; ∈ is the error term, which obeys the normal distribution.

4. The device failure rate evaluation method for distribution network reliability calculation according to claim 2 is characterized in that: In step S1, fault factor classification and cluster analysis are performed on the collected data, and the main contents include: After identifying the main failure factors, they are classified according to different dimensions to facilitate the subsequent model construction and failure rate calculation; Frequency analysis: Count the occurrence frequencies of various fault factors in historical data, identify high-frequency fault factors, and give priority to fault factors with higher occurrence frequencies as the main influencing factors; Among them, f i is the frequency of the ith fault factor, NNN i is the number of occurrences of the i-th fault factor, N total is the total number of occurrences of all fault factors; Equipment type: For medium voltage distribution network, the main equipment involved includes distribution transformers, overhead lines, cables, upstream power supplies, circuit breakers, load switches, disconnectors, etc. The reliability of these equipment directly affects the overall performance of the distribution network; Operation life: refers to the time that various types of equipment have been in use since they were put into operation. The operation life is divided into four time periods: 0-5 years, 5-10 years, 10-15 years, and more than 15 years. Equipment in different time periods has different failure risks; Manufacturer: The manufacturer of the equipment is another important factor affecting the failure rate. If there are too many manufacturers in a certain area, the manufacturers that account for more than 10% will be mainly considered, and other more scattered manufacturers will be classified as "others"; Geographical area: The geographical distribution of equipment refers to the distribution of each type of equipment in a region. Geographical areas are divided into three levels: feeder, grid, and region. Several feeders form a grid, and several grids form a region. Therefore, by counting the distribution of equipment in feeders, the distribution of equipment in grids and regions can be derived. Cluster analysis: After the classification is completed, cluster analysis is used to refine the classified equipment, and K-means clustering is used to classify fault factors with similar characteristics into the same category to facilitate subsequent classification and analysis.

5. The method for evaluating equipment failure rate for distribution network reliability calculation according to claim 4, characterized in that: The cluster analysis includes the following contents: (1) Initialization: Select k initial cluster centers; (2) Assignment: Assign each data point to the cluster to which the nearest cluster center belongs; (3) Update: recalculate the cluster center of each cluster; (4) Iteration: Repeat steps (2) and (3) until the cluster center no longer changes significantly or the preset number of iterations is reached.

6. The device failure rate evaluation method for distribution network reliability calculation according to claim 1 is characterized in that: In step S2, based on the Bayesian neural network, a device failure rate prediction model is constructed to make a preliminary prediction of the failure rate of each type of equipment, including the following contents; The structure of the Bayesian neural network consists of an input layer, a hidden layer, and an output layer. The hidden layer in the Bayesian neural network is a probability layer, which introduces uncertainty into the weight of the Bayesian layer of the i+1th layer by sampling from the distribution of the Bayesian layer of the ith layer. In the parameter sampling process of Bayesian neural network, the weight calculation formula is: The formula for calculating the bias is: in, The weights and biases of the linear transformation of the Bayesian layer corresponding to the i-th layer with n samples, where ρ i is the standard deviation of the sample, μ i is the average value of the sample, as standard; By constructing different Bayesian linear layers, the posterior distribution p(ω│X,Y) on the model parameters ω of the training set data Y and X is inferred, and the formula is: After completing the fitting of the training set, the posterior distribution p(ω│X,Y) exists, and there is a new data point X pre When the input model is integrated over ω, we can get X pre The prediction distribution of is as follows: p(Y pre │X pre ,X,Y)=p(Y pre │X pre ,ω)p(ω│X,Y)dω In the prediction process, a sample n is drawn from the posterior distribution p(ω│X,Y) each time, and a new data point X is obtained for the given n. pre Point estimate Y pre The above operation is repeated many times, thus obtaining a new data point X pre Estimation of the distribution of .

7. The device failure rate evaluation method for distribution network reliability calculation according to claim 1 is characterized in that: In step S2, the Bayesian neural network is trained and parameter optimized by setting the loss function and optimizer, including the following: The loss function of BNN not only needs to measure the difference between the predicted failure rate and the actual failure, but also needs to consider the modeling of parameter uncertainty in Bayesian reasoning, using the negative log-likelihood loss function: L(θ)=-logP(Y|X,θ) Where Y is the actual fault label (0 or 1); X is the input feature of the device; θ is the parameter weight and bias of the network; P(Y|X,θ) is the probability distribution predicted by the model given the parameter θ; In order to minimize the loss function, the Adam optimizer is used. As a gradient descent optimization method, it can adaptively adjust the learning rate and is suitable for handling the complexity of Bayesian neural networks. The weight update formula of Adam is: m t =β1m t-1 +(1-β1)g t Among them, g t is the gradient of the loss function L(θ) under the current weight; β1 is the momentum factor, which is used to update the smooth gradient to prevent the gradient from fluctuating too much. The common default value is β1=0.9, which is a weighted average of 90% of the new gradient information and 10% of the past information; m t is the first-order momentum of the gradient at the current moment, similar to the weighted average; v t It is a second-order moment estimation. Similar to the first-order moment, Adam uses the square of the gradient to calculate the second-order moment, which is the moving average of the square of the gradient to balance the uneven size of the gradient update. It is for m t and v t Adam combines the corrected gradient momentum and gradient variance estimate to update the weight W t ; η is the learning rate, which determines the size of the weight adjustment at each update; it is a very small constant used to avoid division by zero errors.

8. The device failure rate evaluation method for distribution network reliability calculation according to claim 1 is characterized in that: In step S3, the Bayesian neural network is used to decompose the equipment failure rate into weights of different failure factors. By decomposing the weights of various equipment failure factors, the influence of each factor on the equipment failure rate can be accurately identified; The equipment failure rate taking failure factors into consideration can be expressed as: R=W i ·f i Among them, W i is the corresponding fault factor f i The weight of For the weight W, the fault factor is taken as input and the weight distribution of each fault factor is taken as output to construct a BNN weight decomposition model. At this time, the simple structure of the model is: input layer-Bayesian layer-output layer.

9. The device failure rate evaluation method for distribution network reliability calculation according to claim 1 is characterized in that: In step S3, the equipment failure rate is decomposed into weights of different failure factors using a Bayesian neural network. The training steps of the BNN weight decomposition model are as follows: (1) Determine the likelihood function, given the training data Define the likelihood function p(D|ω): Among them, x i is the input feature vector, including the equipment operation years, equipment region, and equipment manufacturer, y i is the output failure rate, ω is the network weight; (2) Determine the posterior distribution and calculate the posterior distribution p(ω|D): Since p(D) is a constant, it can be ignored and simplified to: p(ω|D)∝p(D|ω)p(ω) (3) Variational inference: approximate the posterior distribution p(ω|D) to a tractable distribution q(ω) and find the optimal approximate solution by minimizing the difference between the two distributions. The specific steps are: Choose an approximate distribution: Choose an easy-to-handle distribution q(ω) to approximate the posterior distribution P(ω|D). Use a Gaussian distribution: q(ω)=N(μ,σ 2 ) Among them, μ is the mean, σ 2 is the variance; Define the variational objective: Use the evidence lower bound (ELBO) to approximate the optimization objective; ELBO=E q(ω) [LogP(D|ω)]-KL(q(ω)||P(ω)) Among them, the first term is the expected likelihood function, that is, the degree of fit of the expected model to the observed data; the second term is the Kullback-Leibler divergence, which is used to measure the difference between the approximate distribution q(ω) and the prior distribution P(ω); Gradient optimization: Use the gradient descent method to update the parameters of q(ω) to maximize the ELBO; When ELBO converges or the optimization reaches the preset number of iterations, the variational inference ends, and q(ω) at this time is considered to be the optimal approximation of the posterior distribution; Iterate prediction and output weight result W i .

10. The device failure rate evaluation method for distribution network reliability calculation according to claim 1, characterized in that: In step S4, the weight distribution obtained by the BNN model is applied to the actual failure rate data, and the failure rate is decomposed layer by layer. The calculation formula is as follows: R i =W age ·f age +W reg ·f reg +W maf ·f maf Among them, R i is the failure rate of type i equipment, W age is the weight of the equipment’s operating life, W reg is the weight of the area to which the device belongs, W maf is the weight of the equipment manufacturer; f age is the failure rate level of the equipment during its service life, f reg is the failure rate level of the area to which the equipment belongs, f maf is the failure rate level of the equipment manufacturer, f age ,f reg ,f maf Based on the calculation in step S2, W age ,W reg ,W maf Calculated based on step S3; Based on the grid weight coefficient, the regional failure rate is decomposed into each grid: l grid =λ region ·W grid Among them, W grid is the grid weight coefficient: In the above formula, N grid is the number of devices in the grid, f BNN is the weight decomposition model, x i is the feature vector of the device in the grid; Based on the feeder weight coefficient, the grid failure rate is decomposed into each feeder: l freeder =λ grid ·W freeder Among them, W freeder The calculation method for the grid weight coefficient is: In the above formula, N freeder is the number of devices in the feeder, f BNN is the weight decomposition model, x i is the eigenvector of the device in the feeder.

Images