Transformer Maintenance Decision-making Method and System Based on State Evaluation and Failure Rate Correction

Abstract

The present invention relates to a transformer maintenance decision-making method and system based on state evaluation and failure rate correction, belonging to the technical field of data preprocessing and recognition, and solving the problem of how to improve the accuracy of maintenance decision-making. The method includes: collecting transformer fault data and performing preprocessing to obtain characteristic quantities; using historical transformer fault data to divide the transformer fault data types into primary division intervals, and then secondarily dividing the primary division intervals into health assessment intervals; constructing and training multiple neural network models based on the characteristic quantities and health assessment intervals, and combining multiple prediction results to obtain the final predicted interval type; generating a failure rate curve of the transformer based on the final predicted interval type and correcting the failure rate curve based on individual differences and operation state portraits; constructing a maintenance strategy decision-making model based on the corrected failure rate curve and the minimum cost and reliability constraints to obtain the optimal maintenance time. Reduce the fault downtime and improve the accuracy of maintenance decision-making.

Description

Technical Field

[0001] The present invention relates to the technical field of data preprocessing and recognition, and particularly to a transformer maintenance decision-making method and system based on state evaluation and failure rate correction. Background Art

[0002] First of all, with the adjustment of the energy structure and the increase in safety requirements, the asset scale and investment demand of power grid enterprises have risen sharply. To adapt to this change, enterprises need to formulate more refined investment strategies and consider how to more effectively transmit investment and operating costs under electricity price regulation.

[0003] Under such a background, as asset-intensive enterprises, how power grid enterprises formulate effective equipment maintenance strategies has become the focus of attention in the industry. Traditional maintenance modes, whether based on regular maintenance or state evaluation, mainly focus on the overall condition of equipment. However, in the case of limited resources, this overall consideration often cannot meet the operation and maintenance needs of all equipment.

[0004] In fact, even if equipment is in the same health state, its position in the power grid, the importance of the load it serves, and the reliability requirements are different, which means that the maintenance resource requirements and priorities are also different. To use resources more efficiently, it is necessary to comprehensively consider factors such as the health condition of equipment, the topological structure of the power grid, and the power outage losses that may be caused by equipment failures. Different maintenance plans should be formulated for different equipment. With the rapid development of the energy and power markets, power grid enterprises need to continuously innovate and optimize their equipment operation and maintenance strategies to adapt to the new market environment and operation requirements. Summary of the Invention

[0005] In view of the above analysis, the embodiments of the present invention aim to provide a transformer maintenance decision-making method and system based on state evaluation and failure rate correction to solve the problem of how to improve the accuracy of maintenance decision-making.

[0006] On the one hand, an embodiment of the present invention provides a transformer maintenance decision-making method based on state evaluation and failure rate correction, which is characterized by including: collecting transformer fault data and performing preprocessing to obtain characteristic quantities, where the characteristic quantities correspond to oil chromatogram data, oil test data, and electrical test data; using the preprocessed historical transformer fault data to divide the transformer fault data types into a primary division interval, and then based on the primary division interval and set parameters, the primary division interval is secondarily divided into a health assessment interval; constructing and training multiple neural network models based on the characteristic quantities and the health assessment interval to obtain multiple prediction results of the transformer health status, and combining the multiple prediction results to obtain the final prediction interval type; generating a failure rate curve of the transformer based on the final prediction interval type, and correcting the failure rate curve based on individual differences and operation status portraits; constructing a maintenance strategy decision-making model based on the corrected failure rate curve and the minimum cost and reliability constraints to obtain the optimal maintenance time.

[0007] For a further improvement based on the above method, collecting transformer fault data and performing preprocessing to obtain characteristic quantities includes: collecting the transformer fault data, where the transformer fault data includes fault data and its corresponding transformer state data and external environment data in the case of the transformer failure; performing preprocessing on data anomalies, where the preprocessing includes filling in missing data and removing outliers and duplicate values; removing redundant relevant elements from the preprocessed transformer fault data through the entropy weight method to obtain the characteristic quantities.

[0008] For a further improvement based on the above method, further including removing redundant relevant elements from the preprocessed transformer fault data through the entropy weight method to obtain the characteristic quantities: using m to-be-evaluated samples and n indicators in the transformer fault data to form the following original indicator data matrix:

[0009]

[0010] where, x ij represents the j-th indicator value of the i-th sample; performing normalization processing on each indicator value in the original indicator data matrix using the range normalization method and the standard deviation normalization method; calculating the proportion of each indicator value through the following formula:

[0011]

[0012] where, p ij represents the proportion of the j-th indicator value of the i-th sample;

[0013]

[0014] Calculate the index entropy value through the following entropy value calculation formula:

[0015]

[0016] where e j represents the index entropy value of the j-th index value; based on the degree of difference of each index, use the following formula to calculate the redundancy value of each index:

[0017]

[0018] d j = 1 - e j ;

[0019] where d j is the degree of difference of the j-th index; sort the redundancy values of each index from largest to smallest and select the evaluation indexes corresponding to the predetermined number of the largest redundancy values of the indexes as the characteristic quantities.

[0020] Based on a further improvement of the above method, use the preprocessed transformer fault data to divide the transformer fault data type into a primary division interval, and then based on the primary division interval and set parameters, the primary division interval is further divided into a health assessment interval, which further includes: based on the Markov chain, use the preprocessed transformer fault data to divide the transformer into a primary division interval, and the primary division interval includes a no-fault interval [a1, b1], a minor fault interval [a2, b2] and a serious fault interval [a3, b3]; set the point between no-fault and minor fault as c1 and set the point between minor fault and serious fault as c2, and based on the point c1 and the point c2, set two judgment fuzzy regions for no-fault and minor fault, and minor fault and serious fault; calculate the length of the fuzzy region through the following formula:

[0021] L1 = min{b1 - a1, b2 - a2}·k%;

[0022] L2 = min{b2 - a2, b3 - a3}·k%;

[0023] Based on the length of the fuzzy region, determine the fuzzy region as [c1 - L1, c1 + L1],

[0024] [c2 - L2, c2 + L2]; The health status interval is further divided into health assessment intervals based on the length of the fuzzy region and the fuzzy region, where the health assessment intervals include a fault - free interval [a1, c1 - L1], an interval with a possible minor fault [c1 - L1, c1 + L1], a minor - fault interval [c1 + L1, c2 - L2], an interval with a possible severe fault [c2 - L2, c2 + L2], and an interval with a possible severe fault [c2 + L2, b3].

[0025] Based on a further improvement of the above - mentioned method, multiple neural network models are constructed and trained based on the characteristic quantity and the health assessment intervals to obtain multiple prediction results of the transformer health status. Combining the multiple prediction results to obtain the final prediction interval type further includes: normalizing the characteristic quantity:

[0026]

[0027] where x gy is the normalized value, x is the value of the characteristic quantity in the dataset, x min and x max are the maximum and minimum values of the characteristic quantity respectively; three neural network models corresponding to the characteristic quantities corresponding to the oil chromatographic data, the oil test data, and the electrical test data after normalization are established respectively, and the three neural network models are trained, where the output quantity of each neural network model is determined based on the health assessment interval; the final prediction interval of the transformer is predicted using the trained neural network models, where the final prediction interval includes the fault - free interval, the interval with a possible minor fault, the minor - fault interval, the interval with a possible severe fault, and the interval with a possible severe fault.

[0028] Based on a further improvement of the above - mentioned method, generating a failure rate curve of the transformer based on the final prediction interval type, and correcting the failure rate curve based on individual differences and operation status portraits further includes: obtaining the shape parameter and scale parameter of each stage using the transformer fault data based on the final prediction interval type to establish the failure rate curve of the transformer based on the shape parameter and scale parameter; according to the actual operation situation of the transformer, linearly correcting the failure rate curve of the transformers produced by the same manufacturer to obtain a corrected failure rate function considering familial defects, and then obtaining a corrected failure rate curve considering individual defects based on the corrected failure rate curve considering familial defects; constructing an operation status portrait and correcting the failure rate curve of the transformer based on the operation status portrait to obtain a corrected failure rate curve considering the operation portrait.

[0029] Based on the further improvement of the above method, the corrected failure rate curve after considering familial defects is expressed as:

[0030]

[0031] The corrected failure rate curve after considering individual defects is expressed as;

[0032]

[0033] Where, gt(t) is the corrected failure rate function after considering individual defects, wu is the average trouble-free time of this transformer, t2 is the operating years of this transformer, and n2 is the number of failures occurred during the operation of this transformer; qx(t) is the corrected failure rate function after considering familial defects, jc(t) is the failure rate function fitted according to the historical data of similar transformers, wf is the average trouble-free time of similar transformers, wt is the average trouble-free time of transformers produced by the same manufacturer, t 1_i is the operating years of the i-th transformer among similar transformers, n1 is the total number of failures occurred during the operation of the statistically similar transformers, t 0_i is the operating years of the i-th transformer among the transformers produced by the same manufacturer, and n0 is the total number of failures occurred during the operation of the transformers produced by the same manufacturer.

[0034] Based on the further improvement of the above method, constructing an operation status portrait and correcting the failure rate curve of the transformer based on the operation status portrait to obtain the following corrected failure rate curve further includes: objectively extracting indicators by using correlation analysis and Laplace scoring method to select a predetermined number of indicators with different association strengths with other indicators and the largest Laplace score as the main feature indicators; generating indicator labels based on the main feature indicators, where the indicator labels include indicators corresponding to the regional attribute, natural attribute, and operation attribute respectively; constructing an operation status portrait based on the indicator labels, where the operation status portrait includes the regional attribute, the natural attribute, and the operation attribute; calculating the weight values of the indicator labels under each attribute dimension by the entropy weight method, and performing weighted processing and sorting on the indicator labels based on the weight values of the indicator labels to obtain the corresponding label levels; performing correlation analysis on different operation status portraits and the transformer failure rate to obtain an adjustment coefficient g; generating the following corrected failure rate curve after considering the operation portrait based on the adjustment coefficient, the failure rate curve of the transformer, the corrected failure rate curve after considering familial defects, and the corrected failure rate curve after considering individual defects:

[0035]

[0036] Where g is the adjustment coefficient, HX1 is the failure rate of the corresponding portrait, and HX2 is the average failure rate.

[0037] Based on the further improvement of the above method, based on the corrected failure rate curve and the minimum cost and reliability constraints, constructing a maintenance strategy decision model to obtain the final maintenance plan further includes: constructing the maintenance strategy decision model and its constraint conditions based on the maintenance risk and the failure risk, wherein, the maintenance risk is determined based on the random load shedding loss and the planned load shedding loss of the power grid in the n-period maintenance mode m; and the failure risk is determined based on the random load shedding loss of the power grid and the self-loss of the transformer in the n-period maintenance mode m; the constraint conditions include simultaneous maintenance constraint, mutually exclusive maintenance constraint, maintenance resource constraint and power grid security constraint, wherein, the simultaneous maintenance constraint refers to the constraint of simultaneously maintaining the transformer to avoid repeated power outages caused by transformer maintenance; the mutually exclusive maintenance constraint refers to the constraint of not maintaining the transformer at the same time to avoid power outages; the maintenance resource constraint refers to that the total number of transformers to be maintained cannot exceed the maintenance capacity of the maintenance personnel; and the power grid security constraint refers to the constraint of performing safety inspection through power flow calculation.

[0038] On the other hand, the embodiment of the present invention provides a transformer maintenance decision system based on state evaluation and failure rate correction, which is characterized by including: a data collection and processing module, configured to collect transformer fault data and perform preprocessing to obtain characteristic quantities, wherein, the characteristic quantities correspond to oil chromatogram data, oil test data and electrical test data; a health assessment interval acquisition module, configured to divide the transformer fault data type into a primary division interval by using the preprocessed historical transformer fault data, and then secondarily divide the primary division interval into a health assessment interval based on the primary division interval and set parameters; an interval prediction module, configured to construct and train multiple neural network models based on the characteristic quantities and the health assessment interval to obtain multiple prediction results, and combine the multiple prediction results to obtain a final prediction interval type; a failure curve generation module, configured to generate a failure rate curve of the transformer based on the final prediction interval type, and correct the failure rate curve based on individual differences and operation state portraits; an optimal maintenance time determination module, configured to construct a maintenance strategy decision model based on the corrected failure rate curve and the minimum cost and reliability constraints to obtain the optimal maintenance time.

[0039] Compared with the prior art, the present invention can at least achieve one of the following beneficial effects:

[0040] 1. Through preprocessing, the influence of incorrect data on the analysis result can be reduced to improve data reliability. Detailed fault classification helps to more accurately identify fault modes, provides accurate information for subsequent steps to improve fault identification accuracy, and time series analysis can improve the prediction accuracy of future fault occurrences.

[0041] 2. Provide a more accurate health status through an optimized neural network structure. Multi-source information fusion can provide more comprehensive equipment status information, which helps improve the accuracy of judgment. Enhancing the model adaptability enables it to maintain a high judgment accuracy when facing new data.

[0042] 3. Considering more influencing factors can improve the accuracy of the failure rate curve. Dynamic correction can ensure that the failure rate curve always reflects the latest status of the equipment. Uncertainty analysis helps evaluate the risk and reliability of the prediction results.

[0043] 4. Cost-benefit analysis can ensure that the maintenance strategy is optimal in terms of economy and long-term benefits. Multi-objective optimization can help find the best balance among multiple objectives. Algorithm optimization can improve the solution efficiency and accuracy of the maintenance strategy decision-making model.

[0044] In the present invention, the above technical solutions can also be combined with each other to achieve more preferred combination schemes. Other features and advantages of the present invention will be described in the subsequent specification, and some advantages can be made obvious from the specification or understood by implementing the present invention. The objectives and other advantages of the present invention can be achieved and obtained through the content specifically pointed out in the specification and the drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] The drawings are only for the purpose of showing specific embodiments and are not considered as a limitation to the present invention. Throughout the drawings, the same reference signs represent the same components.

[0046] Figure 1 is a flowchart of a transformer maintenance decision-making method based on state assessment and failure rate correction according to an embodiment of the present invention;

[0047] Figure 2 is a topological structure diagram of a neural network model according to an embodiment of the present invention;

[0048] Figure 3 is a working flowchart of a neural network model according to an embodiment of the present invention;

[0049] Figure 4 is a flowchart of optimizing a neural network model through a genetic algorithm according to an embodiment of the present invention;

[0050] Figure 5 is a block diagram of a transformer maintenance decision-making system based on state assessment and failure rate correction according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0051] The following will specifically describe the preferred embodiments of the present invention with reference to the drawings. The drawings form a part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, rather than to limit the scope of the present invention.

[0052] A specific embodiment of the present invention discloses a transformer maintenance decision-making method based on state evaluation and failure rate correction, as Figure 1 shown, which includes the following specific steps.

[0053] In step S101, transformer fault data is collected and preprocessed to obtain characteristic quantities, where the characteristic quantities correspond to oil chromatogram data, oil test data, and electrical test data.

[0054] Collecting transformer fault data and preprocessing to obtain characteristic quantities includes: collecting transformer fault data, and the transformer fault data (i.e., the original data set) includes fault data and its corresponding transformer state data and external environment data in the case of transformer failure.

[0055] For example, the fault data includes fault types: such as winding short circuit, insulation aging, cooling system failure, etc. The fault occurrence time: specific to year, month, day, hour, and minute to understand the frequency and time period of fault occurrence. The fault severity: such as minor, medium, severe, or evaluated according to power outage time, repair difficulty, etc. The equipment state data includes oil temperature: monitored in real time through an oil temperature sensor, the normal oil temperature range, and the possible faults indicated by abnormal oil temperature rise. Load condition: including the load rate and three-phase load balance of the transformer, reflecting the operating state of the transformer. Oil level: The height of the oil level is directly related to the heat dissipation and insulation performance of the transformer. Winding temperature: monitored through a winding temperature sensor, and too high temperature may indicate winding faults. Insulation resistance: Regularly measuring the insulation resistance to evaluate the insulation state of the transformer. External environment data: Temperature: The ambient temperature has a direct impact on the heat dissipation and performance of the transformer. Humidity: High humidity may cause the insulation performance of the transformer to decline. Pollution degree: Such as pollutants like dust and salt spray may affect the insulation and heat dissipation of the transformer. Wind speed: Affects the natural cooling effect of the transformer. Lightning activity: Lightning may trigger lightning strikes on the transformer. Other relevant data: Historical maintenance records: including information such as maintenance time, maintenance content, and maintenance personnel, which helps analyze the cause of the fault and formulate a maintenance plan. Equipment service life: The service life of the transformer is closely related to its operating state, and equipment with a long service life may be more prone to failures. Operating log: Records the daily operating data of the transformer, such as voltage, current, power, etc., which helps detect potential faults.

[0056] Preprocess the data anomalies, where the preprocessing includes filling in the missing data and removing the outliers and duplicate values. First, construct a data missing value and outlier processing model in combination with relevant actual requirements; second, carry out reasonable and effective processing on the data through integration and transformation means; third, adopt the missing value processing method based on repeated simulation for the missing value processing, and generate a set of complete data sets from a data set containing missing values. The missing data in each data set is filled in by the Monte Carlo method.

[0057] Missing value processing: Use the variable mean vector and variance-covariance matrix as prior information to construct a Markov chain, ensure that the distribution of its elements can converge to a stationary distribution, repeatedly simulate this Markov chain through sampling, obtain the stationary posterior distribution, and generate an estimate of the missing data. Its steps can be sorted out as follows.

[0058] Receive a continuous data vector set (i.e., the original data set) Y c =[Y1, Y2,...., Y n , where the i-th data vector is Y(i)=[y i (1), y i (2),....., y i (D)], i = 1, 2,....., N, where Y c includes the observed data Y wz and the missing data Y qs .

[0059] Set a Gaussian model according to the i-th item of data, where the parameter space of the Gaussian model is θ, and calculate the probability p(Y g |Y qs , θ wz ) of the occurrence of the missing data according to the estimated value θ g of the space θ, and calculate the occurrence probability of the parameter space θ according to the current complete data and the estimated value of the missing data and update the estimated value of the parameter space θ of the Gaussian model. Until the obtained Markov chain converges, estimate the missing data.

[0060] The final missing data calculation formula is:

[0061]

[0062] where N sample is the total number of samples, and N Burn-in is the number of missing samples.

[0063] Adopt the method of deletion for both outliers and duplicate values, and finally obtain the processed data domain set. Divide the fault types according to experience (refer to Table 1 below).

[0064] Table 1

[0065]

[0066] Redundant relevant elements are removed from the preprocessed transformer fault data by the entropy weight method to obtain characteristic quantities (refer to Table 2 below). The main purpose of the entropy value method is to assign weights to the index system. The larger the entropy, the more chaotic the system, the less information it carries, and the smaller the weight; the smaller the entropy, the more orderly the system, the more information it carries, and the larger the weight. The entropy value method is an objective weight assignment method. It draws on the idea of information entropy. By calculating the information entropy of the indexes and determining the weights of the indexes according to the impact of the relative change degree of the indexes on the overall system, that is, assigning weights according to the difference degree of the flag values of each index, so as to obtain the corresponding weights of each index. The indexes with a large relative change degree have larger weights.

[0067] Table 2

[0068]

[0069] Removing redundant relevant elements from the preprocessed transformer fault data by the entropy weight method to obtain characteristic quantities further includes the following steps.

[0070] Original data collection and rectification: Using m samples to be evaluated and n indexes in the transformer fault data, the following original index data matrix is formed:

[0071]

[0072] Among them, x ij represents the j-th index value of the i-th sample. For a certain index X j , the greater the dispersion degree of the samples, the greater the role of this index in the comprehensive evaluation. If the flag values of this index are all equal, it means that this index plays no role in the comprehensive evaluation.

[0073] Data processing: The normalization processing is carried out on each index value in the original index data matrix by using the range normalization method and the standard deviation normalization method. To eliminate the influence of different dimensions on the evaluation results, it is necessary to normalize or standardize each index. Currently, there are the range normalization method and the Z-score normalization method.

[0074] Range normalization: Select the maximum and minimum values of the sample data, perform standardization processing on the original data, and finally map all data to the interval [0,1].

[0075]

[0076] In the formula, X′ and X represent the sample data after and before standardization respectively; Xmax , X min respectively represent the maximum and minimum values in the sample data.

[0077] Z - score normalization method: When the maximum and minimum values of a certain indicator are unknown, or there are outlier values outside the value range, the range normalization method is no longer applicable. Another commonly used method for data normalization can be adopted, namely Z - score normalization, also called standard deviation normalization method. Based on the mean μ and standard deviation σ of the original data, the data is normalized. The original value X of A is normalized to X′ using the Z - score method. The specific calculation formula

[0078]

[0079] Calculate the proportion: Calculate the proportion of each indicator value through the following formula:

[0080]

[0081] where p ij represents the proportion of the j - th indicator value of the i - th sample. Based on this, the ratio matrix of the data can be constructed as follows:

[0082]

[0083] Calculate the indicator entropy value: Calculate the indicator entropy value through the following entropy value calculation formula:

[0084]

[0085] where e j represents the indicator entropy value of the j - th indicator value;

[0086] Define the indicator redundancy value: Based on the degree of difference of each indicator, calculate the redundancy value of each indicator using the following formula:

[0087]

[0088] where d j is the degree of difference of the j - th indicator;

[0089] Sort the redundancy values of each indicator from largest to smallest and select the evaluation indicators corresponding to a predetermined number of the largest indicator redundancy values as characteristic quantities. Combining the calculation results of the entropy weight method, select 40% of the largest calculation results as the subsequent typical characteristic elements.

[0090] The volume fractions of H2 and the total C2H2 are important characteristic quantities for monitoring the occurrence of faults. H2 is a typical representative gas for partial discharge type faults and can best reflect whether there is a partial discharge fault. The total hydrocarbons, as the sum of hydrocarbon gases related to faults, can reflect whether there are overheating faults and discharge faults. C2H2 plays a very obvious role in reflecting discharge faults. CO is one of the decomposition products of insulating paper and can reflect the insulation aging condition of the transformer, especially the overheating fault condition. The furfural content in the oil chemical test data is also one of the insulation decomposition products and can reflect the insulation aging condition of the transformer. The micro water content in the oil is an important parameter for monitoring insulation moisture, which is of great significance for identifying insulation moisture faults. The oil breakdown voltage is similar to the aldehyde content and the CO gas production rate, and can all reflect the insulation aging condition to a certain extent. The oil dielectric loss can reflect the insulation condition of the oil. In the electrical test data, the absorption ratio can reflect the degree of insulation moisture, insulation breakdown or severe overheating aging. The core insulation resistance is of great significance in judging core-related faults such as core grounding. The difference in the DC resistance of the winding can reflect thermal faults, and the dielectric loss of the winding insulation can also reflect the insulation condition of the winding.

[0091] In step S102, the transformer fault data types are divided into primary division intervals by using the preprocessed historical transformer fault data, and then the primary division intervals are secondarily divided into health assessment intervals based on the primary division intervals and the set parameters.

[0092] Dividing the transformer fault data types into primary division intervals by using the preprocessed transformer fault data, and then secondarily dividing the primary division intervals into health assessment intervals further includes: based on the Markov chain, dividing the transformer into primary division intervals by using the preprocessed transformer fault data. The primary division intervals include a fault-free interval [a1, b1], a minor fault interval [a2, b2], and a severe fault interval [a3, b3]. Set the point between the fault-free and minor faults as c1 and the point between the minor and severe faults as c2, and set two judgment fuzzy regions for the fault-free and minor faults, and the minor and severe faults based on the points c1 and c2. Calculate the lengths of the fuzzy regions through the following formula:

[0093] L1 = min{b1 - a1, b2 - a2}·k%; L2 = min{b2 - a2, b3 - a3}·k%;

[0094] Determine the fuzzy regions as [c1 - L1, c1 + L1], [c2 - L2, c2 + L2] based on the lengths of the fuzzy regions.

[0095] Based on the length of the fuzzy region and the fuzzy region, the health state interval is divided into health assessment intervals for the second time. Among them, the health assessment intervals include a fault-free interval [a1, c1 - L1], an interval with a possible slight fault [c1 - L1, c1 + L1], a slight fault interval [c1 + L1, c2 - L2], an interval with a possible severe fault [c2 - L2, c2 + L2], and a severe fault interval [c2 + L2, b3].

[0096] Preliminary division of the health state interval based on Markov chain and Gibbs sampling: A Markov chain is a stochastic process in probability theory and mathematical statistics that has the Markov property and exists in a discrete index set and state space. A Markov chain applicable to a continuous index set is called a Markov process, but sometimes it is also regarded as a subset of a Markov chain, namely a continuous-time Markov chain, corresponding to a discrete-time Markov chain. Therefore, a Markov chain is a relatively broad concept. A Markov chain can be defined by a transition matrix and a transition graph. In addition to the Markov property, a Markov chain may have irreducibility, recurrence, periodicity, and ergodicity. An irreducible and positive-recurrent Markov chain is a strictly stationary Markov chain with a unique stationary distribution. The limiting distribution of an ergodic Markov chain converges to its stationary distribution.

[0097] Let X t represent the value of the random variable X at discrete time t. If the transition probability of this variable changing over time depends only on its current value, that is

[0098] P(X t+1 = s j |X0 = s0, X1 = s1,..., X t = s t ) = P(X t+1 = s j |X t = s t );

[0099] means that the state transition probability depends only on the previous state, then this variable is called a Markov variable, where s0, s1,..., s i , s j ∈Ω are the possible states of the random variable X. This property is called the Markov property, and a stochastic process with the Markov property is called a Markov process. A Markov chain refers to the sequence of values of the random variable X over a period of time (X0, X1,..., X m ). A Markov chain is defined by the corresponding transition probability. The transition probability refers to the probability that the random variable transfers from one state s i to another state s j at the next moment, that is:

[0100] P(i → j) = P i,j = P(X t+1 = s j |X t = s i );

[0101] denotes the probability that the random variable X takes the value s at time t. Then the probability that the random variable X takes the value s at time t + 1 is: k For the probability that the value of the random variable X at time t + 1 is s i is:

[0102]

[0103] Assume the number of states is n, then there is:

[0104]

[0105] P = (P i,j ) n×n is the transition probability matrix. The Markov chain has periodicity, that is, after a finite number of state transitions, it will return to itself. It also has irreducibility, that is, the states can transfer between each other. If a Markov process has neither periodicity nor irreducibility, it is called ergodic. For an ergodic Markov process, regardless of the initial value π (0) how it takes values, as the number of transitions increases, the value distribution of the random variable will eventually converge to a unique stationary distribution π * . The data for establishing the transformer health assessment system conforms to the ergodic situation.

[0106] The principle of the Monte Carlo method is to simulate the system through a large number of random sample samplings, so as to obtain the parameters to be calculated. Based on the Markov chain method, combined with the Monte Carlo method, sampling for the initial division interval of the transformer can be realized. The data for sampling are multi-dimensional feature quantities, and there is a certain correlation between them. When directly using all the data for calculating the probability distribution of the feature quantities, both the complexity and the time consumption will be extremely large. Therefore, Gibbs sampling in the Markov chain Monte Carlo method can be used to sample the multi-dimensional feature quantities and obtain the distribution probabilities of each feature quantity accordingly.

[0107] Gibbs sampling is applicable to the situation where the joint distribution is not known explicitly or is difficult to sample directly, but the conditional distribution of each variable is known and easy to sample. Compared with other MCMC methods with an acceptance rate less than 1, the acceptance rate of Gibbs sampling is 1, and the sampling efficiency is higher, and it can meet the sampling problem of multi-dimensional data. The principle of Gibbs sampling is as follows. If the state transition matrix P and the probability distribution π * of the aperiodic Markov chain satisfy for all i, j:

[0108] π(i)P(i, j) = π(j)P(j, i);

[0109] Then the probability distribution π * is the stationary distribution of the state transition matrix P. Suppose there is a two-dimensional probability distribution π(i, j), for two points A(x1, y1) and B(x2, y2) with the same x-coordinate, there is:

[0110] π(x1, y1)π(y2|x1) = π(x1)π(y1|x1)π(y2|x1);

[0111] π(x1, y2)π(y1|x1) = π(x1)π(y2|x1)π(y1|x1);

[0112] It can be seen that on the line x = x1, using the conditional probability distribution π(y|x) as the state transition probability of the Markov chain, the transition between any two points satisfies the detailed balance condition. The same is true for the points C(x2, y1) on the line y = y1. Therefore, construct the state transition probability matrix P between any two points on the plane:

[0113] P(A→B) = π(y B |x1) if x A = x B = x1;

[0114] P(A→C) = π(x c |y1) if y A = y B = y1;

[0115] P(A→D) = 0 else;

[0116] For the state transition probability matrix P, it can be seen that for any two points X and Y on the plane, the detailed balance condition is satisfied:

[0117] π(X)P(X→Y) = π(Y)P(Y→X);

[0118] Perform Gibbs sampling on the data in each partition interval according to the feature quantity category to obtain the distribution probability of each feature quantity. Obtain the sampling results of no-fault, minor fault, and severe fault for each feature quantity as the health assessment interval for the initial partition.

[0119] Precise partitioning of the health state interval based on normal distribution and statistical superposition: The normal distribution, also known as the "normal distribution", also known as the Gaussian distribution, is a very important probability distribution in the fields of mathematics, physics, and engineering, and has a major influence in many aspects of statistics. The normal curve is bell-shaped, low at both ends, high in the middle, and symmetric about the y-axis. Because its curve is bell-shaped, it is often called the bell curve.

[0120] Three normal distribution curves. Assume that the intervals corresponding to the normal curves of no failure, minor failure, and severe failure are [a1, b1], [a2, b2], and [a3, b3] respectively, and the points of no failure and minor failure, minor failure and severe failure are c1 and c2. Then, two judgment fuzzy regions of no failure and minor failure, minor failure and severe failure are set as new division intervals to warn of the transition from no failure to minor failure and from minor failure to severe failure. The lengths L1 and L2 of the corresponding fuzzy regions are respectively:

[0121] L1 = min{b1 - a1, b2 - a2}·k%; L2 = min{b2 - a2, b3 - a3}·k%;

[0122] The corresponding fuzzy intervals are:

[0123] [c1 - L1, c1 + L1]; [c2 - L2, c2 + L2];

[0124] For other parameters, the normal distribution function of the probability distribution can also be determined by sampling within the three initially divided intervals through this method, and then the interval secondary division can be achieved by setting parameters, dividing the final interval into five. This way of dividing intervals can realize the fault warning function, and the boundaries of the interval division can be adaptively adjusted according to the actual data set and requirements. The processed intervals are referred to Table 3 below.

[0125] Table 3

[0126] Divided interval Number Interval value No fault 1 <![CDATA[[a1,c1-L1]]]> Possible occurrence of minor fault 2 <![CDATA[[c1-L1,c1+L1]]]> Occurrence of minor fault 3 <![CDATA[[c1+L1,c2-L2]]]> Possible occurrence of serious fault 4 <![CDATA[[c2-L2,c2+L2]]]> Possible occurrence of serious fault 5 <![CDATA[[c2+L2,b3]]]>

[0127] The improvements in this step include: (1) Data cleaning and preprocessing: Introduce more advanced data cleaning techniques to remove outliers, missing values, and duplicate values to improve data quality. (2) Refinement of fault classification: Make a more detailed division of fault types to more accurately identify fault patterns. (3) Time series analysis: Utilize time series analysis techniques to capture the trends and periodicities of fault occurrences.

[0128] Technical effects: (1) Improve data reliability: Data cleaning and preprocessing can reduce the impact of incorrect data on the analysis results. (2) Accuracy of fault identification: Detailed fault classification helps to more accurately identify fault patterns and provides accurate information for subsequent steps. (3) Prediction accuracy: Time series analysis can improve the prediction accuracy of future fault occurrences.

[0129] In step S103, multiple neural network models are constructed and trained based on the characteristic quantities and the health assessment intervals to obtain multiple prediction results of the transformer health condition, and the multiple prediction results are combined to obtain the final prediction interval type. Constructing and training multiple neural network models based on the characteristic quantities and the health assessment intervals to obtain multiple prediction results of the transformer health condition, and combining the multiple prediction results to obtain the final prediction interval type further includes:

[0130] (1) Normalize the fault-related elements: Normalize the characteristic quantities:

[0131]

[0132] where x gy is the normalized value, x is the value of the characteristic quantity in the dataset, x min , x max are the maximum and minimum values of the characteristic quantity respectively.

[0133] (2) Construct multiple neural network models: Three neural network models corresponding to the characteristic quantities of the oil chromatogram data, the oil test data, and the electrical test data after normalization are established respectively, and the three neural network models are trained. Among them, the output quantity of each neural network model is determined based on the health assessment interval;

[0134] Using the feedforward neural network method, multiple deep neural networks are combined as sub-models, and the final output result is coupled by the results of each sub-model. For different types of characteristic quantities, their sensitivities to different faults or health states of the transformer are also different. Therefore, establishing models for different types of characteristic quantities separately can highlight their advantages in classifying different health states. The combined model can evaluate and screen the results according to the prediction situations of each sub-model, and different sub-models can correct each other to improve the final prediction accuracy.

[0135] Obtain the training accuracy. Neural network models are established for the training data of the oil chromatogram data, the oil test data, and the electrical test data respectively to obtain 3 models, namely model 1 to model 3. Then when the best training accuracy is obtained, the training model is stored, and 5 division intervals are output, corresponding to the above five processed intervals. Let X1, X2, …, X n be the input vectors of the BP neural network, Y1, Y2, …, Y m be the output values, w ij and w jk be the weights. The typical topological structure diagram of the BP neural network is as Figure 2 shown.

[0136] When the number of input and output nodes of the BP neural network is n and m respectively, it reflects the mapping relationship between n independent variables and m dependent variables. The BP neural network modeling prediction includes three steps: network structure construction, training, and prediction. The basic work process is as Figure 3 shown.

[0137] Assume that the activation function of each layer of nodes in the network is the S-shaped function, and the input of the i-th node in the first layer of the network is denoted as net i , the output is denoted as o i , and the output of the k-th node in the output layer is y k . Then the input of the j-th node in the middle layer is:

[0138] o j = f(net j );

[0139] Define the error of the network as the difference between the expected output and the actual output. Then there is If there are i neurons in the output layer, define the square error between the actual output and the expected output as:

[0140] Since the BP algorithm modifies the weights according to the negative gradient of the error E, the modification of the weights can be expressed as: W m+1 = w m + Δw m = w m - λg m ;

[0141] where λ is the learning step size, m represents the number of iterations,

[0142] Because it is the output layer, at this time is the actual output value. According to the definition of e k and the square error, we can get:

[0143] According to the definition of e k , we can get:

[0144] According to the above formula , we can get:

[0145] According to the above formula , we can get:

[0146] Finally, we get:

[0147] Now let the learning error of the output layer: σk = e k f′(net k );

[0148]

[0149] The modification amount Δw of the weights of the hidden layer neural units kj :

[0150]

[0151] According to the above formula and o j = f(net j ) we can get:

[0152]

[0153] Because we are looking for the change in the weights of the hidden layer. At this time, the effect of the upper layer on it should be considered, so we have:

[0154]

[0155] According to it can be known that

[0156]

[0157] Also according to we can get:

[0158]

[0159] Substitute into formula and deduce:

[0160]

[0161] Let the learning error of the hidden layer be:

[0162]

[0163] Then optimize the model through the genetic algorithm (refer to Figure 4 ).

[0164] Use multiple trained neural network models to predict the final prediction interval of the transformer, where the final prediction interval includes a fault-free interval, a possible interval of slight fault occurrence, a slight fault occurrence interval, a possible interval of severe fault occurrence, and a severe fault occurrence interval.

[0165] Build a combined model: Combine all the prediction results of Model 1 to Model 3 to obtain the final predicted state type. The rules are as follows: Among the prediction results of Model 1 to Model 3, the division interval with the most occurrences is the final prediction interval. If there are multiple division intervals with the same number of occurrences, then weight these division intervals according to the best training accuracy of the model, and the division interval with the highest weight is the final prediction interval.

[0166] Test the combined model: When the fault diagnosis is no fault, input the test data into the combined model to obtain the final predicted interval type; when the fault diagnosis is a fault, the cases where the test results of Model 1 to 3 are no fault can be excluded to improve the prediction accuracy.

[0167] The improvements in this step include: (1) Model optimization: Adopt a more advanced neural network structure, such as a deep learning model, to improve the accuracy of health state judgment. (2) Multi-source information fusion: Combine other sensor data (such as temperature, vibration, etc.) and environmental data to provide richer input information for the neural network. (3) Model adaptability: Enhance the adaptability of the model to new data and new fault modes to avoid overfitting.

[0168] Technical effects: (1) Improve judgment accuracy: The optimized neural network structure can provide more accurate health state judgment. (2) More comprehensive information: Multi-source information fusion can provide more comprehensive equipment state information, which helps to improve the accuracy of judgment. (3) Better generalization ability: Enhancing the model adaptability enables it to maintain a high judgment accuracy when facing new data.

[0169] In step S104, generate the failure rate curve of the transformer based on the final predicted interval type, and correct the failure rate curve based on individual differences and operation state portraits. Generating the failure rate curve of the transformer based on the final predicted interval type, and correcting the failure rate curve based on individual differences and operation state portraits further includes the following steps.

[0170] Calculate the conventional failure curve: Based on the Weibull distribution, use the transformer failure data to obtain the shape parameter and scale parameter at each stage, and establish the failure rate curve of the transformer based on the shape parameter and scale parameter.

[0171] The Weibull distribution is the theoretical basis for reliability analysis and life testing. The Weibull distribution is widely used in reliability engineering, especially suitable for the distribution form of cumulative wear failure of electromechanical products. Since it can easily infer its distribution parameters using probability values, it is widely used in the data processing of various life tests.

[0172]

[0173] where λ(t) is the transformer life, t is the time, β is the scale parameter, and α is the shape parameter. By selecting different scale parameters, the changing trend of the failure rate in different periods can be described: when β < 1, the failure rate shows a downward trend, corresponding to the early failure period of the bathtub curve; when β > 1, the failure rate shows an upward trend, corresponding to the wear-out failure period of the bathtub curve; when β = 1, the failure rate is a constant, corresponding to the accidental failure period of the bathtub curve. Based on the historical equipment failure data, by piecewise fitting the Weibull distribution corresponding to the equipment failure rate, the shape parameter α and the scale parameter β are obtained to establish the basic equipment failure rate model.

[0174] According to the actual operation conditions of the transformers, the failure rate curve of the transformers produced by the same manufacturer is linearly corrected to obtain the corrected failure rate function considering the familial defects, and then the corrected failure rate curve considering the individual defects is obtained based on the corrected failure rate curve considering the familial defects.

[0175] Correction considering individual differences: Since the manufacturing processes of different manufacturers are different, the operation conditions and aging failure processes of the equipment produced are very different. In addition, when calculating the equipment failure rate curve, due to the limited number of equipment produced by a certain manufacturer, it is difficult to accurately obtain the shape parameter and scale parameter of its Weibull distribution. Therefore, it is necessary to correct the basic failure rate curve according to the statistical data of the equipment produced by different manufacturers, so as to reflect the influence of familial defects on the equipment failure rate. In engineering applications, the actual mean time between failures of the equipment produced by the same manufacturer can be compared with the overall mean time between failures of the same type of equipment, and the ratio of the two can be used to linearly correct the basic failure rate of the same type of products.

[0176] The corrected failure rate curve considering the familial defects is expressed as:

[0177]

[0178] Similar to the familial defects, to reflect the individual differences between different equipment produced by the same manufacturer, it is necessary to linearly correct the failure rate curve of the equipment produced by the same manufacturer according to the actual operation conditions of the equipment, so as to reflect the influence of individual defects on the equipment failure rate.

[0179] The corrected failure rate curve considering the individual defects is expressed as;

[0180]

[0181] Among them, gt(t) is the corrected failure rate function after considering individual defects, wu is the average failure-free time of the individual transformer, t2 is the operating life of the transformer, and n2 is the number of failures of the transformer during operation; qx(t) is the corrected failure rate function after considering family defects, jc(t) is the failure rate function fitted according to the historical data of similar transformers, wf is the average failure-free time of similar transformers, wt is the average failure-free time of transformers produced by the same manufacturer, and t 1_i is the operating life of the i-th transformer among the same type of transformers, n1 is the total number of failures of the same type of transformers during operation, t 0_i is the operating life of the i-th transformer produced by the same manufacturer, and n0 is the total number of failures that occurred during the operation of transformers produced by the same manufacturer.

[0182] Considering the correction of the operation profile: constructing the operation status profile and correcting the failure rate curve of the transformer based on the operation status profile to obtain the corrected failure rate curve after considering the operation profile. Constructing the operation status profile and correcting the failure rate curve of the transformer based on the operation status profile to obtain the following corrected failure rate curve further includes the following steps.

[0183] 1. Construction of operation portrait system: Use correlation analysis and Laplace scoring method to objectively extract indicators to select a predetermined number of indicators with different correlation strengths with other indicators and the largest Laplace scores as main feature indicators.

[0184] Feature extraction is the most critical link in the process of building a grid portrait. It is the integration of characteristics or commonalities. By calculating the similarity and modeling the fact label system under each attribute, the main information is extracted, the dimension is reduced, and the calculation model is simplified. The idea of ​​extraction is generally to remove or merge features with high correlation, requiring each feature to be able to independently and clearly describe the grid. If two variables with high correlation are added to the model as features at the same time, it will cause overfitting and increase the complexity of the model. The method of feature extraction also affects the accuracy of the results. Commonly used methods include mapping method, selection method, objective extraction method, etc. This paper adopts "correlation analysis + Laplace scoring method" to objectively extract indicators. The basic analysis process is as follows:

[0185] (1) Data standardization. Each indicator extracted from the fact label library contains an indicator value and a graded semantic label value. The distribution of each indicator value is different. In order to reduce the error in the calculation of the correlation coefficient, it is necessary to undergo a skewness-kurtosis test or approximate normal distribution processing. After the variable x is normalized, it is converted to X:

[0186]

[0187] Where: is the mean of variable X; σ is the standard deviation of variable X.

[0188] (2) Calculation of the correlation coefficient between indicators. The Pearson correlation coefficient is a calculation method for measuring the linear correlation of two sets of discrete variables. The Pearson correlation coefficient of two sets of variables x and y is defined as:

[0189]

[0190] where: σ xy is the covariance of variables x and y; σ(x) and σ(y) are the standard deviations of variables x and y, respectively.

[0191] (3) Laplace score calculation. First, construct the weight matrix Z, and let LS r be the Laplace score of the r-th feature, and f ri represent the i-th sample (i = 1, 2,..., m) of the r-th feature. For each feature, construct an m×m adjacency matrix Z, and the values of the elements in Z are as follows:

[0192]

[0193] where, x i , x j are the values of the i-th and j-th samples of this feature, and t is an appropriate constant.

[0194] Finally, calculate the Laplace score:

[0195]

[0196] where, f r = [f r1 , f r2 , …, f rm T .

[0197] (4) Selection of characteristic indicators. Based on the Laplace score and the results of correlation analysis, select several indicators with significantly different association strengths with other indicators and relatively high scores as the main characteristic indicators.

[0198] (5) Model label generation. The model label is the labeled text further refined through methods such as abstract clustering based on the extracted main characteristic indicators. It can accurately express the grid characteristics concisely, is easy to understand and apply, and is a further induction and summary of the operation characteristics. The model label not only identifies the operation characteristics but also provides convenience for the management and optimization services of equipment failures.

[0199] Generate index labels based on the main characteristic indicators. The index labels include indicators corresponding to regional attributes, natural attributes, and operation attributes respectively.​

[0200] (6) Image construction. Based on the metric labels, construct the operation status image, which includes regional attributes, natural attributes, and operation attributes (refer to Table 4 below).

[0201] Table 4

[0202]

[0203]

[0204] Calculate the weight values of the metric labels under each attribute dimension through the entropy weight method, and perform weighted processing and sorting on the metric labels based on the weight values of the metric labels to obtain the corresponding label levels.

[0205] In summary, through modeling analysis, the operation status image finally has the model label attributes of three dimensions: regional attributes, natural attributes, and operation attributes. On the basis of the model labels, the entropy weight method is introduced to obtain the weight values of the model labels under each attribute dimension. The calculation steps of the entropy weight method are as follows:

[0206] The calculation steps of the entropy weight method are roughly divided into the following three steps: 1) Determine whether there are negative numbers in the input matrix. If so, re-normalize it to the non-negative interval. 2) Calculate the proportion of the i-th sample under the j-th index, and regard it as the probability used in the relative entropy calculation. 3) Calculate the information entropy of each index, calculate the information utility value, and normalize it to obtain the entropy weight of each index. Sort according to the percentile after weighted scoring to obtain the corresponding label level (refer to Table 5 below).

[0207] Table 5

[0208]

[0209]

[0210] 2. Analysis of the correlation between the operation image and the failure rate: The main purpose of correlation analysis is to study the degree of closeness of the relationship between variables, such as the relationship between height and weight, the amount of wire and the amount of tower materials, the capacity of the main transformer and the distribution device, etc. In statistical analysis, correlation generally refers to "linear correlation", and its degree of closeness is represented by the correlation coefficient. The correlation coefficient is usually denoted as r, and its value ranges from -1 to +1. The closer the absolute value is to 1, the closer the relationship between the variables. When the absolute value is equal to 1, it means that the two variables are completely correlated, and the value of variable B can be obtained from the value of variable A. When the correlation coefficient is positive, it means that when variable A increases, variable B also increases, and the two are in a positive correlation relationship; conversely, it means that when variable A increases, variable B decreases, and the two are in a negative correlation relationship.

[0211] The Pearson simple correlation coefficient is used to measure the linear correlation of interval variables and is the most widely used in calculating the correlation coefficient. Its calculation formula is as follows:

[0212]

[0213] where n is the number of samples, x i and y i are the values of the two variables in different samples respectively, and are the means of variables x i and y i respectively. Since the calculation formula of the Pearson simple correlation coefficient is exactly in the form of matrix product, it is also called the product-moment correlation coefficient. After transforming the formula, it is found that the correlation coefficient can be expressed as the product of the standardized x i and y i respectively, and then the average of n products is calculated.

[0214] 3. Modified failure curve: Conduct a correlation analysis on the portraits of different operating states and the failure rate of the transformer to obtain the adjustment coefficient g; Based on the adjustment coefficient, the failure rate curve of the transformer, the modified failure rate curve considering familial defects, and the modified failure rate curve considering individual defects, generate the following modified failure rate curve considering the operating portrait:

[0215]

[0216] where g is the adjustment coefficient, HX1 is the failure rate corresponding to the portrait, and HX2 is the average failure rate.

[0217] Improvements in this step: (1) Consider more influencing factors: In addition to individual defects, maintenance behaviors, and operating states, other factors affecting the failure rate can also be considered, such as environmental factors, load changes, etc. (2) Dynamic correction: As the operating time of the equipment increases, regularly update and correct the failure rate curve to reflect the changes in the equipment state. (3) Introduce uncertainty analysis: Conduct uncertainty analysis on the failure rate curve to evaluate the reliability and stability of the prediction results.

[0218] Technical effects: (1) More accurate prediction: Considering more influencing factors can improve the accuracy of the failure rate curve. (2) Real-time update: Dynamic correction can ensure that the failure rate curve always reflects the latest state of the equipment. (3) Risk assessment: Uncertainty analysis helps to evaluate the risk and reliability of the prediction results.

[0219] In step S105, based on the corrected failure rate curve and the minimum cost and reliability constraints, a maintenance strategy decision model is constructed to obtain the optimal maintenance time. Constructing a maintenance strategy decision model based on the corrected failure rate curve and the minimum cost and reliability constraints to obtain the final maintenance time plan (constructing an optimized maintenance strategy decision model based on the corrected failure rate curve and the dual considerations of cost and reliability) further includes: constructing a maintenance strategy decision model and its constraints based on the maintenance risk and the failure risk, where the maintenance risk is determined based on the random load loss and the planned load loss of the power grid under the maintenance mode m in the nth time period; and the failure risk is determined based on the random load loss of the power grid and the self-loss of the transformer under the maintenance mode m in the nth time period; the constraints include the simultaneous maintenance constraint, the mutually exclusive maintenance constraint, the maintenance resource constraint, and the power grid security constraint, where the simultaneous maintenance constraint refers to the constraint of simultaneously maintaining transformers to avoid repeated power outages caused by transformer maintenance; the mutually exclusive maintenance constraint refers to the constraint of not maintaining transformers at the same time to avoid power outages; the maintenance resource constraint refers to the total number of transformers to be maintained not exceeding the maintenance capacity of the maintenance personnel; and the power grid security constraint refers to the constraint of performing safety inspection through power flow calculation. By solving the model, the maintenance time point that can ensure the safe and stable operation of the transformer and achieve the optimal economic cost is determined.

[0220] (1) Grid maintenance risk: Condition-based maintenance is a preventive maintenance, which is carried out before equipment failure. Therefore, while reducing the loss of equipment failure, it causes another loss, that is, the maintenance risk of the power grid.

[0221]

[0222] O M1 =p m P M1,m T m c m1 ;

[0223]

[0224] In the formula, O M is the maintenance risk; O M (n) is the maintenance risk in the nth time period; T is the number of time periods in a cycle; M n is the set of maintenance modes in the tth time period; O M1,m and O M2,m are the random load loss and the planned load loss of the power grid under the maintenance mode m in the nth time period; p m is the probability of the occurrence of the maintenance mode m, P M1,m is the planned load loss caused by the maintenance mode m in the tth time period, T m is the duration of the power outage caused by the maintenance; c m1Compensation for the units that lost load on the demand side plan; F n is the maintenance equipment set under maintenance mode m; C k is the maintenance cost of equipment k under maintenance mode m.

[0225] (2) Grid failure risk: In addition to the equipment under maintenance, other equipment may fail at any time, causing grid failure losses.

[0226]

[0227] In the formula, O F is the risk of failure; F (n) is the grid failure risk in period t; in period t, under maintenance mode m, O F1,f is the random load loss of the power grid; O F2,f is the equipment loss itself; GZX(t) is the equipment failure rate, P f,m is the random load loss of the power grid, N F is the set of faulty devices. μ is the repair probability; C mj is the fault repair cost of equipment j; C Oj is the replacement cost of equipment j; T f The power outage time caused by fault f; c m2 Compensation for units that lose load on the demand side plan.

[0228] (3) Distribution network maintenance plan optimization model: The goal of distribution transformer maintenance optimization is to minimize the risk during maintenance. When a distribution transformer is maintained, there is a maintenance risk. M and failure risk O F , reducing the risk of failure will inevitably increase the risk of maintenance, that is, the two are contradictory. However, the two are unified in the safe and economical operation of the distribution network. Therefore, the objective function expression is as follows:

[0229] F = min(O F +O M );

[0230] The constraints are as follows:

[0231] 1) Simultaneous maintenance constraints. Avoid repeated power outages caused by equipment maintenance to improve power supply reliability. Maintenance plans should try to avoid repeated maintenance operations. For problems that can be solved by a single power outage, repeated power outages are not allowed due to inconsiderate considerations. Therefore, some equipment must be maintained at the same time.

[0232] tt x =tt y ;

[0233] 2) Mutually exclusive maintenance constraints. To avoid unnecessary power outages during maintenance, the maintenance of some equipment should not be scheduled at the same time.

[0234] tt y >tt x +T x +1;

[0235] 3) Maintenance resource constraints. Maintenance resource constraints refer to constraints such as the number of maintenance personnel and their technical capabilities. The equipment to be maintained simultaneously cannot exceed the maintenance capacity of the maintenance personnel.

[0236]

[0237] In the formula, m is the total number of equipment, u sn is the equipment maintenance status variable within time period t. When the equipment is in the maintenance state, its value is 1; otherwise, it is 0. M is the upper limit of the number of equipment to be maintained in each time period.

[0238] 4) Power grid security constraints. When equipment is taken out of operation for maintenance, it will cause changes in the power flow, which may lead to overload of some lines and over-limit of node voltages. Therefore, safety inspections must be carried out through power flow calculations.

[0239] P l ≤p lmax ;

[0240] U qmin ≤U q ≤U qmax ;

[0241] In the formula, P l is the power of transformer l; p lmax is the maximum power that transformer l is allowed to pass through; U q , U qmin and U qmax are the voltage of node q and the upper and lower limit values of the voltage of node q, respectively.

[0242] (4) Solution: The Marine Predators Algorithm is a new meta-heuristic optimization algorithm. By simulating the natural law of survival of the fittest in the ocean, various marine organisms continuously switch their identities as predators and prey, and change their foraging strategies according to different situations, thus carrying out the optimization process.

[0243] The Marine Predators Algorithm believes that the foraging strategy of marine predators changes between flight and Brownian walk, and selects between the two strategies according to different scenarios, so as to obtain the optimal foraging strategy.

[0244] First, initialize the positions of predators and prey. Construct an elite matrix with the predator with the optimal fitness, and construct a prey matrix with uniformly distributed prey, as shown in the following formulas respectively.

[0245] According to different speed ratios, the optimization process of MPA is divided into three stages.

[0246] Phase 1: Survey phase: This phase is also called the high-speed ratio phase. In this phase, the prey is much faster than the predator. The predator adopts a static strategy, while the prey performs Brownian motion. This phase often occurs at the beginning of the algorithm optimization iteration, and is a survey of the global position information. The mathematical model of this phase is expressed as follows:

[0247]

[0248] in is the moving step length; is a random vector based on Brownian walk normal distribution; is a term-by-term multiplication operation; P is a constant, equal to 0.5; is a uniform random vector in [0,1]; Iter is the current number of iterations; Max_Iter is the maximum number of iterations; N is the population size.

[0249] Phase 2: This phase is also called the medium speed ratio phase. In this phase, the speeds of the predator and prey are similar, and both are looking for their own prey. This phase generally occurs in the middle of the algorithm iteration. The population is divided into two parts, where the prey performs Lévy flight and is responsible for the development of the space, and the predator performs Brownian motion and is responsible for the exploration of the space. The mathematical model expression of this phase is as follows:

[0250]

[0251] in, represents the Lévy motion random vector, and CF represents the adaptive parameter that controls the predator's moving step size.

[0252] Phase 3: Development phase: This phase is also called the low-speed phase. In this phase, the speed of the predator is greater than that of the prey. This mainly occurs in the late stage of algorithm iteration. The predator adopts the Levy flight strategy and pays more attention to the development of local areas. The mathematical model expression of this phase is as follows:

[0253]

[0254] In addition, the algorithm also takes into account external environmental factors such as fish aggregation devices and eddy effects, and changes the foraging strategy of predators to jump out of local extremes and avoid premature convergence. The mathematical model expression is as follows:

[0255]

[0256] Where FADs is the impact probability, which is generally taken as 0.2; is a binary vector; r is a random number within [0, 1]; r1 and r2 are respectively random indices of the prey matrix. Finally, the optimal overhaul plan is obtained according to the solution.

[0257] The improvements in this step include: (1) Cost-benefit analysis: In addition to considering the overhaul cost, the long-term benefits brought by the overhaul can also be considered, such as reducing the fault downtime and improving the equipment reliability. (2) Multi-objective optimization: Incorporate multiple objectives (such as cost, reliability, safety, etc.) into the optimization model to seek the overall optimal solution. (3) Algorithm optimization: Adopt more advanced optimization algorithms (such as genetic algorithm, simulated annealing algorithm, etc.) to improve the solution efficiency and accuracy.

[0258] Technical effects: (1) Comprehensive optimization: The cost-benefit analysis can ensure that the overhaul strategy reaches the optimal in terms of economy and long-term benefits. (2) Multi-objective balance: The multi-objective optimization can help find the best balance point among multiple objectives. (3) Efficient solution: The algorithm optimization can improve the solution efficiency and accuracy of the overhaul strategy decision model.

[0259] Reference Figure 5 , a specific embodiment of the present invention discloses a transformer overhaul decision-making system based on state evaluation and failure rate correction, including: a data acquisition and processing module 501, configured to collect transformer fault data and perform preprocessing to obtain characteristic quantities, where the characteristic quantities correspond to oil chromatogram data, oil test data, and electrical test data; a health assessment interval acquisition module 502, configured to divide the transformer fault data types into primary division intervals by using the preprocessed historical transformer fault data, and then secondary divide the primary division intervals into health assessment intervals based on the primary division intervals and set parameters; an interval prediction module 503, configured to construct and train multiple neural network models based on the characteristic quantities and the health assessment intervals to obtain multiple prediction results of the transformer health status, and combine the multiple prediction results to obtain the final prediction interval type of the transformer health status; a fault curve generation module 504, configured to generate a failure rate curve of the transformer based on the final prediction interval type, and correct the failure rate curve based on individual differences and operation status portraits; an optimal overhaul time determination module 505, configured to construct an overhaul strategy decision model based on the corrected failure rate curve and the minimum cost and reliability constraints to obtain the optimal overhaul time.

[0260] Those skilled in the art can understand that all or part of the processes of implementing the above embodiment methods can be completed by instructing relevant hardware through a computer program, and the program can be stored in a computer-readable storage medium. Among them, the computer-readable storage medium is a disk, an optical disk, a read-only memory, or a random access memory, etc.

[0261] The above are only the preferred specific embodiments 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 those skilled in the art within the technical scope disclosed by the present invention should be covered by the protection scope of the present invention.

Claims

1. A transformer maintenance decision-making method based on state assessment and fault rate correction, characterized in that: include: Collecting transformer fault data and preprocessing it to obtain characteristic quantities, wherein the characteristic quantities correspond to oil chromatographic data, oil test data and electrical test data; Using the preprocessed historical transformer fault data, the transformer fault data type is divided into primary division intervals, and then based on the primary division intervals and setting parameters, the primary division intervals are secondary divided into health assessment intervals; Constructing and training multiple neural network models based on the characteristic quantity and the health assessment interval to obtain multiple prediction results of the transformer health status, and combining the multiple prediction results to obtain a final prediction interval type; Generating a failure rate curve of the transformer based on the final prediction interval type, and correcting the failure rate curve based on individual differences and operating status portraits; Based on the corrected failure rate curve and the minimum cost and reliability constraints, a maintenance strategy decision model is constructed to obtain the optimal maintenance time. Among them, using the pre-processed transformer fault data to divide the transformer fault data type into primary division intervals, and then based on the primary division intervals and setting parameters, the primary division intervals are divided into health assessment intervals for the second time, further comprising: based on the Markov chain, using the pre-processed transformer fault data to divide the transformer into primary division intervals, the primary division intervals include a no-fault interval [a1, b1], a minor fault interval [a2, b2], and a major fault interval [a3, b3]; setting the point between no-fault and minor fault to c1 and the point between minor fault and major fault to c2, and setting two judgment fuzzy regions of no-fault and minor fault, and minor fault and major fault based on the point c1 and the point c2; calculating the length of the fuzzy region by the following formula: L1=min{b1-a1,b2-a2}·k%; L2=min{b2-a2,b3-a3}·k%; The blur region is determined as [c1-L1, c1+L1] based on the length of the blur region, [c2-L2, c2+L2]; Based on the length of the fuzzy area and the fuzzy area, the health status interval is divided into health assessment intervals for secondary use, wherein the health assessment interval includes a no-fault interval [a1, c1-L1], an interval with a possible minor fault [c1-L1, c1+L1], an interval with a minor fault [c1+L1, c2-L2], an interval with a possible serious fault [c2-L2, c2+L2], and an interval with a possible serious fault [c2+L2, b3]; Among them, generating a failure rate curve of the transformer based on the final prediction interval type, and correcting the failure rate curve based on individual differences and an operating status portrait further include: obtaining shape parameters and scale parameters of each stage based on the final prediction interval type using transformer fault data to establish the failure rate curve of the transformer based on the shape parameters and scale parameters; performing linear correction on the failure rate curves of transformers produced by the same manufacturer according to the actual operating conditions of the transformer to obtain a corrected failure rate function after considering family defects, and then obtaining a corrected failure rate curve after considering individual defects based on the corrected failure rate curve after considering family defects; constructing an operating status portrait and correcting the failure rate curve of the transformer based on the operating status portrait to obtain a corrected failure rate curve after considering the operating portrait; Among them, the modified failure rate curve after considering familial defects is expressed as: The modified failure rate curve after considering individual defects is expressed as; Among them, gt(t) is the corrected failure rate function after considering individual defects, wu is the average failure-free time of the individual transformer, t2 is the operating life of the transformer, and n2 is the number of failures of the transformer during operation; qx(t) is the corrected failure rate function after considering family defects, jc(t) is the failure rate function fitted according to the historical data of similar transformers, wf is the average failure-free time of similar transformers, wt is the average failure-free time of transformers produced by the same manufacturer, and t 1_i is the operating life of the i-th transformer among the same type of transformers, n1 is the total number of failures of the same type of transformers during operation, t 0_i is the operating life of the i-th transformer produced by the same manufacturer, and n0 is the total number of failures that occurred during the operation of transformers produced by the same manufacturer.

2. The transformer maintenance decision-making method based on state assessment and fault rate correction according to claim 1 is characterized in that: Collecting transformer fault data and preprocessing it to obtain characteristic quantities includes: Collecting the transformer fault data, wherein the transformer fault data includes fault data when the transformer fails and its corresponding transformer state data and external environment data; Preprocessing data anomalies, wherein the preprocessing includes filling missing data and removing outliers and duplicate values; The redundant related elements are removed from the preprocessed transformer fault data by using the entropy weight method to obtain the characteristic quantity.

3. The transformer maintenance decision-making method based on state assessment and fault rate correction according to claim 2 is characterized in that: Removing redundant related elements from the preprocessed transformer fault data by the entropy weight method to obtain the characteristic quantity further includes: Using the m samples to be evaluated and n indicators in the transformer fault data, the following original indicator data matrix is ​​formed: Among them, x ij represents the jth index value of the i-th sample; Normalizing each indicator value in the original indicator data matrix using the range normalization method and the standard deviation normalization method; The weight of each indicator value is calculated by the following formula: Among them, p ij Indicates the proportion of the jth index value of the i-th sample; The entropy value of the indicator is calculated by the following entropy value calculation formula: Among them, e j The index entropy value representing the j-th index value; Based on the difference degree of each indicator, the redundancy value of each indicator is calculated using the following formula: d j =1-e j ; Among them, d j is the difference degree of the jth indicator; The index redundancy values ​​are sorted from large to small and evaluation indexes corresponding to a predetermined number of index redundancy values ​​with the largest index redundancy values ​​are selected as the feature quantities.

4. The transformer maintenance decision-making method based on state assessment and fault rate correction according to claim 1 is characterized in that: Constructing and training a plurality of neural network models based on the characteristic quantity and the health assessment interval to obtain a plurality of prediction results of the transformer health status, and combining the plurality of prediction results to obtain a final prediction interval type further comprises: The feature quantity is normalized: Among them, x gy is the normalized value, x is the value of the feature in the data set, and x min 、x max They are the maximum and minimum values ​​of the characteristic quantity respectively; Establishing three neural network models of characteristic quantities corresponding to the oil chromatographic data, the oil test data and the electrical test data after normalization respectively, and training the three neural network models, wherein the output quantity of each neural network model is determined based on the health assessment interval; The trained neural network model is used to predict the final prediction interval of the transformer, wherein the final prediction interval includes the fault-free interval, the interval with possible minor faults, the interval with minor faults, the interval with possible serious faults, and the interval with possible serious faults.

5. The transformer maintenance decision-making method based on state assessment and fault rate correction according to claim 1 is characterized in that: Constructing an operating status profile and correcting the failure rate curve of the transformer based on the operating status profile to obtain the following corrected failure rate curve further includes: The indicators are objectively extracted by using correlation analysis and Laplace scoring method to select a predetermined number of indicators with different correlation strengths with other indicators and the largest Laplace scores as the main characteristic indicators; Generate an indicator label based on the main characteristic indicator, wherein the indicator label includes indicators corresponding to regional attributes, natural attributes, and operation attributes respectively; Building an operation status portrait based on the indicator label, the operation status portrait including the regional attributes, the natural attributes and the operation attributes; The weight of the indicator label under each attribute dimension is calculated by the entropy weight method, and the indicator label is weighted and sorted based on the indicator label weight to obtain the corresponding label level; Conduct correlation analysis between different operating status portraits and transformer failure rate to obtain the adjustment coefficient g; Based on the adjustment coefficient, the failure rate curve of the transformer, the modified failure rate curve after considering family defects and the modified failure rate curve after considering individual defects, the following modified failure rate curve after considering the operation profile is generated: Where g is the adjustment coefficient, HX1 is the failure rate of the corresponding image, and HX2 is the average failure rate.

6. The transformer maintenance decision-making method based on state assessment and fault rate correction according to claim 1 is characterized in that: Based on the corrected failure rate curve and the minimum cost and reliability constraints, constructing a maintenance strategy decision model to obtain the final maintenance plan further includes: constructing the maintenance strategy decision model and its constraints based on maintenance risk and fault risk, wherein the maintenance risk is determined based on the random load loss and planned load loss of the power grid under the maintenance mode m in the n-period; and the fault risk is determined based on the random load loss of the power grid and the transformer's own loss under the maintenance mode m in the n-period; the constraints include simultaneous maintenance constraints, mutually exclusive maintenance constraints, maintenance resource constraints and power grid safety constraints, wherein the simultaneous maintenance constraint refers to the constraint of simultaneously repairing transformers to avoid repeated power outages caused by transformer maintenance; the mutually exclusive maintenance constraint refers to the constraint of not repairing transformers at the same time to avoid power outages; the maintenance resource constraint means that the total number of transformers to be repaired cannot exceed the maintenance capacity of the maintenance personnel; and the power grid safety constraint refers to the constraint of safety inspection through flow calculation.

7. A transformer maintenance decision system based on state assessment and fault rate correction, characterized in that: The transformer maintenance decision method based on state assessment and fault rate correction for implementing any one of claims 1 to 6 comprises: A data acquisition and processing module, used for acquiring transformer fault data and preprocessing to obtain characteristic quantities, wherein the characteristic quantities correspond to oil chromatographic data, oil test data and electrical test data; A health assessment interval acquisition module is used to divide the transformer fault data type into primary division intervals using the preprocessed historical transformer fault data, and then divide the primary division interval into health assessment intervals based on the primary division intervals and setting parameters; An interval prediction module, used to construct and train multiple neural network models based on the characteristic quantity and the health assessment interval to obtain multiple prediction results, and combine the multiple prediction results to obtain a final prediction interval type; A fault curve generating module, used for generating a fault rate curve of the transformer based on the final prediction interval type, and correcting the fault rate curve based on individual differences and operating status portraits; The optimal maintenance time determination module is used to construct a maintenance strategy decision model to obtain the optimal maintenance time based on the corrected failure rate curve and the minimum cost and reliability constraints.

Images