Converter transformer key performance evaluation method based on multi-source heterogeneous data

Abstract

The invention discloses a multi-source heterogeneous data-based converter transformer key performance evaluation method, which comprises the steps of data acquisition, BP neural network-based multi-source heterogeneous data cleaning, SVM-based multi-source heterogeneous data fusion, performance evaluation index determination, evaluation system construction and converter transformer key performance evaluation method determination. Comprising the following steps: determining a method fused by an analytic hierarchy process and a fuzzy comprehensive evaluation model method according to an evaluation system to evaluate the key performance of the converter transformer, and analyzing multi-source heterogeneous data obtained by the converter transformer by adopting an AHP and FCA fusion mode so as to comprehensively evaluate the key performance of the converter transformer.

Description

Technical Field

[0001] The present invention relates to the field of data processing, and in particular to a method for evaluating key performance of a converter transformer based on multi-source heterogeneous data. Background Art

[0002] With the continuous growth of global electricity demand and the rapid development of renewable energy, the power system is evolving towards a more complex and intelligent direction. High-voltage direct current (HVDC) transmission technology is widely used due to its advantages in long-distance transmission, among which converter transformers (converter transformers) play a core role. Converter transformers not only realize the power conversion between AC and DC, but also play an important role in voltage stabilization, phase modulation and power quality. The operating environment of converter transformers is complex, and their performance is affected by multiple factors, including thermal load, mechanical stress, environmental conditions, etc. Traditional performance monitoring methods often rely on a single data source or simple data analysis, which cannot fully reflect the actual operating status of the converter transformer. The limitations of this single method make fault prediction and preventive maintenance difficult, increasing the risk of equipment downtime and maintenance costs. In modern power systems, with the development of sensor technology, the Internet of Things (IoT) and big data analysis, it has become feasible to collect and integrate heterogeneous data from multiple sources. Multi-source heterogeneous data can include sensor data, real-time monitoring of temperature, pressure, vibration and other information. Operation records, equipment operation history, fault cases, maintenance logs, etc., environmental data, external environmental factors such as climate change, temperature and humidity fluctuations, simulation and emulation results, and various data to verify equipment performance through models. By combining these diverse data, a comprehensive performance diagnosis of commutation changes can be formed. Summary of the invention

[0003] The purpose of the present invention is to overcome the existing state evaluation of a single type of data, lack of support from multi-source data, inability to fully perceive the operation of the converter transformer, and no research has been conducted on the evaluation methods of the thermal distribution and mechanical stress performance of the converter transformer. Therefore, a key performance evaluation method of the converter transformer based on multi-source heterogeneous data is proposed to comprehensively evaluate the performance of the converter transformer.

[0004] To achieve the above object, the present invention provides the following technical solution: a method for evaluating key performance of a converter transformer based on multi-source heterogeneous data, comprising the following steps:

[0005] Step 1: Data collection, including collecting converter performance evaluation index data through various inspection methods;

[0006] Step 2: Multi-source heterogeneous data cleaning based on BP neural network, including obtaining similarity relationship constraints of multi-source heterogeneous data, building a BPNN data cleaning model based on the relationship constraints, and completing the cleaning of multi-source heterogeneous data through the data cleaning model;

[0007] Step 3: Multi-source heterogeneous data fusion based on SVM, including extracting and processing the features of multi-source heterogeneous data according to the cleaning results of multi-source heterogeneous data, using SVM algorithm to obtain the optimal classification hyperplane of the features, building a data fusion model, and realizing the fusion of multi-source heterogeneous data through the established data fusion model;

[0008] Step 4: Determine the performance evaluation index, including determining the thermal distribution performance index and mechanical stress performance evaluation index based on the fusion results of the multi-source heterogeneous data of the converter transformer, and constructing an evaluation system;

[0009] Step 5: Determine the evaluation method for the key performance of the converter transformer, including evaluating the key performance of the converter transformer by integrating the hierarchical analysis method and the fuzzy comprehensive evaluation model method according to the evaluation system.

[0010] Preferably, in step 2, the method for obtaining similarity relationship constraints of multi-source heterogeneous data includes the following steps: the original multi-source heterogeneous data exists in the form of a data table, the data table is set to b, the total amount of data is set to p, and a data table set B = {b1, b2, ..., b p}, the data attribute set is The Manhattan distance algorithm is used to obtain the similarity between multi-source heterogeneous data. The process is shown in the following formula:

[0011]

[0012] In the formula, the similarity between multi-source heterogeneous data i and data j is expressed as Indicates; sample data attributes are expressed as Represents; attribute feature mark of data i Form; attribute feature tag of data j Format; data dimension mark S format.

[0013] The correlation coefficient between the data is set to Formulate the parameter threshold λ of the data and use it as the constraint condition of the data cleaning model. The process is shown in the following formula:

[0014]

[0015] Based on the constraints constructed above, a data cleaning model is constructed.

[0016] Preferably, in the step 2, a BPNN data cleaning model is constructed based on relational constraints, and the method for cleaning multi-source heterogeneous data through the data cleaning model includes the following steps: setting the multi-source heterogeneous data cleaning model to T, and the model contains N layers of nodes,

[0017] In the cleaning model of multi-source heterogeneous data, each layer of the RBM structure has visible layer and hidden layer data. The input layer data nodes are set to c = (α1, α2, ..., α c}, the number of hidden layer nodes is d = {β1, β2, …, β d}, in order to construct the node weight matrix between the levels of the cleaning model of multi-source heterogeneous data, the process is shown in the following formula:

[0018]

[0019] In the formula, the constructed node weight matrix is ​​described in the form of ξ; the matrix vector is labeled ω cd In this form, based on the constructed node weight matrix combined with the RBM energy function, the node energy function of multi-source heterogeneous data is obtained. The process is shown in the following formula:

[0020]

[0021] In the formula, the node data α i The bias coefficient is δ i Represents; node data β j The bias coefficient is γ j Represents; data weight mark ω ij The form; constants are represented by i and j; the number of nodes of the input layer data in the model is represented by c; the energy function of the acquired data node is described as A(α i ,β j )form;

[0022] The data distribution probability of the model input layer nodes and hidden layer nodes is tested based on the constructed node energy function. The test results are shown in the following formula:

[0023]

[0024] In the formula, the normalized coefficient of the data energy function is described as form; the distribution factor is described in the form of ε; the data distribution probability is expressed as Q(α i ,β j )express;

[0025] Based on the above-obtained data distribution probability, the activity of the model node variables can be effectively obtained to optimize the cleaning model; multi-source heterogeneous data samples are input into the cleaning model, and the states of the data nodes in the middle layer of the model are independent of each other. Assuming that the hidden layer nodes β j When it is 1, the constraint conditions for obtaining the data distribution probability are as follows:

[0026]

[0027] In the formula, the constraints established are Representation; Binary Sigmoid function notation η form; Data state coefficient is expressed as Based on the above calculation results, it can be seen that when the value interval of the Sigmoid function independent variable is [0, 1], the probability distribution function of the data will present a single continuous change state;

[0028] After activating the hidden layer nodes of the neural network model using the above calculation process, input β j Back propagation is performed on the model to realize model training. The node energy anomalies and discrete distribution data captured in the training results are removed to complete the cleaning of multi-source heterogeneous data of the converter transformer.

[0029] Preferably, in step 3, the multi-source heterogeneous data fusion algorithm of SVM includes the following steps: based on the above data cleaning results, the multi-source heterogeneous data set is updated to Y ij (t), the sample data volume is M, the data state is t, and the mean standardization method is used to standardize the data. The result is shown in the following formula:

[0030]

[0031] In the formula, the standardized result of the data is marked Form; Data mean labeling Based on the above data, the standardized processing results are obtained, and the feature vector of multi-source heterogeneous data i at the jth attribute is extracted. The calculation results are shown in the following formula:

[0032]

[0033] In the formula, the feature vector of multi-source heterogeneous data is expressed as Indicates; the data attribute recording period is expressed by T.

[0034] Preferably, in step 3, the multi-source heterogeneous data fusion algorithm of SVM further comprises the following steps:

[0035] Set the acquired data feature vector as the input vector y i , the kernel function is K(·), the feature vector set is y, and the output vector is u i , in order to obtain the optimal hyperplane of the data, the process is as follows:

[0036] u=τ i u i K(y i ,y)+θ

[0037] In the formula, the optimal hyperplane is represented by u; the weighting coefficient is τ iIndicated by; the bias term is denoted by θ; the number of eigenvectors is denoted by m; the constant is marked in the form of i;

[0038] Based on the obtained optimal super-classification surface, the multi-source heterogeneous data is divided into h groups of data, and the data of different groups come from different sensors. The actual value of the data is set as u, the training output is U, and the generalization parameter is χ, so as to establish the fusion model of multi-source heterogeneous data of SVM, as shown in the following formula:

[0039]

[0040] In the formula, the constructed data fusion model is represented by J; the optimal solution of data i and j is marked by l i , l j Format; optimal solution mean labeling Form; Constraint Tagging form; the number of data sources is represented by g;

[0041] Based on the determined optimal solution of the data, the decision function of the data is established, the model is solved, and the data fusion is realized. The process is shown in the following formula:

[0042]

[0043] In the formula, the constructed decision function is represented by f(x); the data mean set of data source g is labeled form; data optimal solution mark l g form.

[0044] Preferably, in step four, the thermal distribution performance indicators include temperature distribution uniformity, thermal response time, cooling efficiency, heat dissipation area and heat loss rate; the mechanical stress performance evaluation indicators include stress, deformation, fatigue life, seismic resistance, connection strength and material yield strength.

[0045] Preferably, in step 5, the maximum eigenvalue λ is calculated. max When the eigenvector W is approximated, the accuracy requirement is not high, and the required parameters can be obtained by using an approximate method. The present invention uses the square root method;

[0046] First do the square root calculation:

[0047]

[0048] Then normalize it:

[0049]

[0050] Let the approximate eigenvector W be:

[0051] W=[w1 w2 … w n ]T

[0052] Calculate the maximum eigenvalue λ of the judgment matrix A max :

[0053]

[0054] For the evaluation of the key performance of the converter transformer, the evaluation matrix A of the R layer to the G layer is calculated; the evaluation matrix of the index layer P relative to the criterion layer R is calculated:

[0055] In addition, the matrix A must be subjected to a consistency check. First, the consistency index is calculated:

[0056]

[0057] The larger the C1 value, the more it deviates from consistency, and the smaller the C1 value, the closer it is to consistency. As shown in Table 7, the R 1； One-time ratio C R , if C R <0.1, then A meets the consistency condition. If C R ≥0.1, then A does not meet the consistency condition; for those that do not meet the requirements, a ij Revalue and recalculate so that C R <0.1, so that A meets the consistency condition;

[0058] Calculate the composite weight of the factors of each layer on R and then sort them.

[0059]

[0060] In the formula, is the total ranking weight of R, and m is the number of factors of R;

[0061] The total ordering vector of P relative to G can be found The consistency ratio is:

[0062]

[0063] For the evaluation of the key performance of the converter transformer, six evaluation criteria are used: {excellent, good, medium, poor, relatively poor, extremely poor}, and the corresponding score set is {6, 5, 4, 3, 2, 1};

[0064] Through actual testing of the commutation transformer, the obtained indicators are used to evaluate the performance of the commutation transformer, and the fuzzy evaluation matrix is ​​obtained based on the comprehensive evaluation of the test results and the scoring statistics of each indicator:

[0065]

[0066] In the formula, m is the number of indicators in the indicator layer; n is the number of comment sets; i is the serial number of a single factor, indicating a certain stage of the test design;

[0067] therefore:

[0068]

[0069] By calculating the fuzzy evaluation comprehensive index, the comprehensive evaluation value B is obtained:

[0070]

[0071] Compared with the prior art, the present invention has the following beneficial effects:

[0072] 1. The multi-source heterogeneous data cleaning algorithm based on BP uses the BP neural network in the machine learning algorithm to construct a multi-source heterogeneous data cleaning processing model to reduce data redundancy. Its mechanism is to first set the similarity constraint relationship of multi-source heterogeneous data to achieve the correlation of the internal structure of the data, which can not only improve the speed of the later BP neural network internal data cleaning operation, but also improve the efficiency of data cleaning. The most basic classifier in the BP neural network is the classification regression tree, which mainly integrates and classifies the multi-source heterogeneous data results under the constraints of the above-mentioned correlation, obtains more accurate multi-source heterogeneous data relationships, and eliminates useless data relationships, thereby completing data cleaning.

[0073] Second, the multi-source heterogeneous data fusion algorithm of SVM extracts and processes the features of multi-source heterogeneous data based on the data cleaning results. SVM is a binary classification method based on statistics. The basic idea of ​​this method is: in the training set, a minimum boundary function is obtained through training so that the shortest distance between each data is maximized. It has the principle of structural risk minimization, has good generalization ability, and is a convex optimization problem. As long as the local optimal solution is determined, it must be the global optimal solution, avoiding the problem of falling into the local optimal solution. Use the SVM algorithm to obtain the optimal classification hyperplane of the feature, build a data fusion model, and finally realize the fusion of multi-source heterogeneous data through the established data fusion model.

[0074] Third, the evaluation method of thermal distribution and mechanical stress performance of converter transformer based on fuzzy analytic hierarchy process. The present invention adopts a method combining analytic hierarchy process and fuzzy comprehensive evaluation model method to evaluate the thermal distribution and mechanical stress performance of converter transformer. AHP adopts a multi-criteria decision-making method. It expresses a complex problem as an ordered hierarchical structure through the factors and related conditions contained in the system. The relative importance of each factor in the hierarchy is determined by mutual comparison, and then the order of relative importance of the factors is determined by comprehensive human judgment. The multi-source heterogeneous data obtained from the converter transformer is analyzed by the fusion of AHP and FCA, so as to comprehensively evaluate the key performance of the converter transformer. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] Figure 1 The figure is a flow chart of the evaluation method of the present invention.

[0076] Figure 2 It is the key performance evaluation diagram.

[0077] Figure 3 It is a neural network model for multi-source heterogeneous data.

[0078] Figure 4 It is a multi-source heterogeneous data fusion process.

[0079] Figure 5 A hierarchical analysis model for commutation transformer performance evaluation. DETAILED DESCRIPTION

[0080] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. 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.

[0081] S1 Step 1: Data collection, including collecting the performance evaluation index data of the converter transformer through a variety of inspection methods; the performance evaluation index data of the converter transformer can be obtained through a variety of inspection methods, including online detection, robot inspection and manual inspection. These methods provide multi-source heterogeneous data, which can fully grasp the thermal distribution and mechanical stress performance of the converter transformer.

[0082] Online detection monitors the operating status of the converter transformer in real time by installing sensors, mainly including the following aspects:

[0083] Temperature distribution monitoring: Use distributed fiber optic temperature sensors to collect temperature data at different locations in real time and generate temperature distribution diagrams to evaluate the temperature uniformity of the equipment. Thermal response time: Monitor the temperature response after load changes, record its change curve, and calculate the time required to reach stability. Cooling efficiency evaluation: Combine flow meters and temperature sensors to calculate the heat exchange efficiency of the cooling system to determine its operating effect.

[0084] Through robot inspection, equipment data can be efficiently obtained in complex or dangerous environments:

[0085] Stress and deformation monitoring: Using inspection robots equipped with strain gauges, real-time monitoring of stress and deformation of key parts. Fatigue life analysis: The robot collects load data during operation and combines it with historical data to evaluate the fatigue life of the material. Seismic resistance test: In a simulated vibration environment, the seismic resistance of the converter transformer is evaluated by monitoring the dynamic response.

[0086] Manual inspections mainly focus on intuitive inspections of equipment and some complex assessments:

[0087] Temperature detection: Use an infrared thermometer or thermal imager to check the temperature distribution of the transformer. Connector inspection: Manually visually inspect the connectors to evaluate their tightness and wear, and perform torque tests when necessary. Material yield strength assessment: Take samples for laboratory testing to obtain the yield strength of the material.

[0088] S2 Step 2: Multi-source heterogeneous data cleaning based on BP neural network, including obtaining similarity relationship constraints of multi-source heterogeneous data, constructing a BPNN data cleaning model based on the relationship constraints, and completing the cleaning of multi-source heterogeneous data through the data cleaning model; after obtaining the multi-source heterogeneous data, it is also necessary to complete data cleaning and data fusion, and finally form a unified multi-source heterogeneous data set. To this end, the present invention proposes a multi-source heterogeneous data fusion algorithm of BP neural network (backpropagation neural network, BPNN) and support vector machine (support vector machine, SVM). Utilize BP neural network to clean data, preliminarily eliminate noise interference, and eliminate redundant data; Utilize SVM algorithm to perform data fusion operation, and with the advantage of its strong generalization ability and stability, find and train the optimal classification surface, avoid falling into local optimality, etc., ensure the effectiveness of classification, and make the constructed multi-source heterogeneous data fusion result more reliable. The specific method is as follows:

[0089] (1) Multi-source heterogeneous data cleaning algorithm based on BP

[0090] ① Obtain similarity relationship constraints for multi-source heterogeneous data

[0091] The original multi-source heterogeneous data exists in the form of data tables. Let the data table be b and the total amount of data be p. Then, we can establish a data table set B = {b1, b2, …, b p}, the data attribute set is The Manhattan distance algorithm is used to obtain the similarity between multi-source heterogeneous data. The process is shown in the following formula:

[0092]

[0093] In the formula, the similarity between multi-source heterogeneous data i and data j is expressed as Indicates; sample data attributes are expressed as Represents; attribute feature mark of data i Form; attribute feature tag of data j Format; data dimension mark S format.

[0094] The correlation coefficient between the data is set to Formulate the parameter threshold λ of the data and use it as the constraint condition of the data cleaning model. The process is shown in the following formula:

[0095]

[0096] Based on the constraints constructed above, a data cleaning model is constructed.

[0097] ②Data cleaning model based on BPNN

[0098] Set the multi-source heterogeneous data cleaning model to T, which contains N layers of nodes. The specific model structure is as follows: Figure 3 shown.

[0099] When the BPNN data cleaning model is back-propagating, the unsupervised Boltzmann machine (RBM) as the main component of the model is the key to solving the redundant data removal and feature extraction of multi-source heterogeneous data.

[0100] In the cleaning model of multi-source heterogeneous data, there are visible layer and hidden layer data in each layer of the RBM structure. The input layer data nodes are set to c = (α1, α2, ..., α c}, the number of hidden layer nodes is d = {β1, β2, …, β d}, in order to construct the node weight matrix between the levels of the cleaning model of multi-source heterogeneous data, the process is shown in the following formula:

[0101]

[0102] In the formula, the constructed node weight matrix is ​​described in the form of ξ; the matrix vector is labeled ω cdBased on the constructed node weight matrix combined with the RBM energy function, the node energy function of multi-source heterogeneous data is obtained. The process is shown in the following formula:

[0103]

[0104] In the formula, the node data α i The bias coefficient is δ i Represents; node data β j The bias coefficient is γ j Represents; data weight mark ω ij The form; constants are represented by i and j; the number of nodes of the input layer data in the model is represented by c; the energy function of the acquired data node is described as A(α i ,β j )form.

[0105] The data distribution probability of the model input layer nodes and hidden layer nodes is tested based on the constructed node energy function. The test results are shown in the following formula:

[0106]

[0107] In the formula, the normalized coefficient of the data energy function is described as form; the distribution factor is described in the form of ε; the data distribution probability is expressed as Q(α i ,β j )express.

[0108] Based on the above-obtained data distribution probability, the activity of the model node variables can be effectively obtained and the cleaning model can be optimized. Generally speaking, when multi-source heterogeneous data samples are input into the cleaning model, the states of the data nodes in the middle layer of the model are independent of each other. Assuming that the hidden layer nodes β j When it is 1, the constraint conditions for obtaining the data distribution probability are as follows:

[0109]

[0110] In the formula, the constraints established are Representation; Binary Sigmoid function notation η form; Data state coefficient is expressed as Based on the above calculation results, it can be seen that when the Sigmoid function independent variable takes a value interval of [0, 1], the probability distribution function of the data will present a single continuous change state.

[0111] After activating the hidden layer nodes of the neural network model using the above calculation process, input β j Back propagation is performed on the model to achieve model training. According to the node energy anomalies and discrete distribution data captured in the training results, they are removed to complete the cleaning of the multi-source heterogeneous data of the converter transformer.

[0112] S3 Step 3: Multi-source heterogeneous data fusion based on SVM, including extracting and processing the features of multi-source heterogeneous data according to the cleaning results of multi-source heterogeneous data, using SVM algorithm to obtain the optimal classification hyperplane of the features, building a data fusion model, and realizing the fusion of multi-source heterogeneous data through the established data fusion model;

[0113] Based on the above data cleaning results, the features of multi-source heterogeneous data are extracted and processed. SVM is a binary classification method based on statistics. The basic idea of ​​this method is: in the training set, a minimum boundary function is obtained through training to maximize the shortest distance between each data. It has the principle of structural risk minimization, has good generalization ability, and is a convex optimization problem. As long as the local optimal solution is determined, it must be the global optimal solution, avoiding the problem of falling into the local optimal solution. The SVM algorithm is used to obtain the optimal classification hyperplane of the feature, build a data fusion model, and finally realize the fusion of multi-source heterogeneous data through the established data fusion model.

[0114] ① Extract features of multi-source heterogeneous data

[0115] Based on the above data cleaning results, update the multi-source heterogeneous data set to Y ij (t), the sample data volume is M, the data state is t, and the mean standardization method is used to standardize the data. The result is shown in the following formula:

[0116]

[0117] In the formula, the standardized result of the data is marked Form; Data mean labeling Based on the above data, the standardized processing results are obtained, and the feature vector of multi-source heterogeneous data i at the jth attribute is extracted. The calculation results are shown in the following formula:

[0118]

[0119] In the formula, the feature vector of multi-source heterogeneous data is expressed as Indicates; the data attribute recording period is expressed by T.

[0120] ②Multi-source heterogeneous data fusion algorithm

[0121] The SVM algorithm is used to accelerate the convergence speed through the kernel function. The dimension of the space is controlled, the nonlinear vector space of data features is mapped to the linear separable space, the optimal classification hyperplane is obtained, and a data fusion model of multi-source heterogeneous data is constructed.

[0122] Set the acquired data feature vector as the input vector y i, the kernel function is K(·), the feature vector set is y, and the output vector is u i , in order to obtain the optimal hyperplane of the data, the process is as follows:

[0123] u=τ i u i K(y i ,y)+θ

[0124] In the formula, the optimal hyperplane is represented by u; the weighting coefficient is τ i The bias term is represented by θ; the number of eigenvectors is represented by m; and the constant is marked in the form of i.

[0125] Based on the obtained optimal super-classification surface, the multi-source heterogeneous data is divided into h groups of data, and the data of different groups come from different sensors. The actual value of the data is set as u, the training output is U, and the generalization parameter is χ, so as to establish the fusion model of multi-source heterogeneous data of SVM, as shown in the following formula:

[0126]

[0127] In the formula, the constructed data fusion model is represented by J; the optimal solution of data i and j is marked by l i , l j Format; optimal solution mean labeling Form; Constraint Tagging format; the number of data sources is represented by g.

[0128] Based on the determined optimal solution of the data, the decision function of the data is established, the model is solved, and the data fusion is realized. The process is shown in the following formula:

[0129]

[0130] In the formula, the constructed decision function is represented by f(x); the data mean set of data source g is labeled form; data optimal solution mark l g The multi-source heterogeneous data fusion process is as follows: Figure 4 shown.

[0131] Finally, based on the above data fusion process, the trained data is entered into the data fusion model, and the constructed data fusion model is verified according to the kernel function, loss function and generalization parameter, and the fusion result is output. When performing multi-source data fusion, it is necessary to encapsulate, store and query the data of the relevant database at the same time, so as to perform business processing, retrieval query and online optimization at any time. The multi-source heterogeneous data that has completed the above processing is classified according to the characteristics of the data. In this way, data cleaning processing of multi-source data characteristics can be realized, duplicate records can be reduced, and vacant records can be quickly processed for multi-source heterogeneous data, effectively improving the quality of data, completing efficient data cleaning, and realizing the fusion of multi-source heterogeneous data.

[0132] S4 Step 4: Determine the performance evaluation index, including determining the thermal distribution performance index and the mechanical stress performance evaluation index according to the fusion result of multi-source heterogeneous data of the converter transformer, and constructing an evaluation system; the present invention selects the following thermal distribution performance indicators for the key performance evaluation indicators of the converter transformer: temperature distribution uniformity, thermal response time, cooling efficiency, heat dissipation area and heat loss rate.

[0133] Temperature distribution uniformity refers to the uniformity of the internal or surface temperature of the converter transformer, reflecting the thermal management capability of the equipment during operation. Thermal response time refers to the time required for the internal temperature of the converter transformer to reach a stable state when the external environment or load changes, which can evaluate the adaptability of the equipment to environmental changes. Cooling efficiency indicates the effectiveness of the cooling system in reducing the temperature of the converter transformer, which directly affects the safety and reliability of the equipment. The heat dissipation area refers to the area of ​​heat exchange between the surface of the converter transformer and the surrounding environment. The larger the heat dissipation area, the stronger the heat dissipation capacity. The heat loss rate refers to the proportion of heat lost due to heat dissipation during the operation of the converter transformer to the total input heat, reflecting the energy efficiency of the equipment.

[0134] In addition, the present invention also selects the following mechanical stress performance evaluation indicators: stress, deformation, fatigue life, earthquake resistance, connection strength and material yield strength.

[0135] Stress is the force per unit area and is an important parameter that describes the internal stress state of a material. Deformation refers to the change in shape or size of a material under the action of a force, which may appear as elastic deformation (recoverable after the external force is removed) or plastic deformation (permanent deformation). The degree of deformation is usually characterized by strain (the ratio of the change in unit length to the original length). Fatigue life refers to the number of load cycles that a transformer can withstand under repeated or cyclical loads until fatigue failure occurs, reflecting the durability of the material. Seismic resistance refers to the ability of a transformer to resist damage under earthquakes or other dynamic loads, ensuring the safety of equipment under extreme conditions. Connection strength refers to the maximum stress that can be borne at the connection between two materials or components, and is usually evaluated at the connector to ensure the overall reliability of the structure. The yield strength of a material refers to the maximum stress that a material can withstand before plastic deformation occurs, and is an important indicator for evaluating material performance.

[0136] The above indicators together constitute a comprehensive evaluation system for the thermal distribution and mechanical stress performance of converter transformers, providing an important basis for the design, optimization and safe operation of equipment.

[0137] S5 Step 5: Determine the evaluation method of the key performance of the converter transformer, including determining the method of evaluating the key performance of the converter transformer by integrating the analytic hierarchy process and the fuzzy comprehensive evaluation model method according to the evaluation system. The present invention adopts the method of integrating the analytic hierarchy process (AHP) and the fuzzy comprehensive assessment model method (FCA) to evaluate the key performance of the converter transformer. AHP adopts a multi-criteria decision-making method. It expresses a complex problem as an ordered hierarchical structure through the factors and related conditions contained in the system, determines the relative importance of each factor in the hierarchy through mutual comparison, and then integrates human judgment to determine the order of relative importance of the factors. The indicators obtained from a converter transformer are discussed in a fusion method of AHP and FCA, so as to comprehensively evaluate the key performance of the converter transformer.

[0138] like Figure 5 As shown in the figure, the target layer is to evaluate the key performance of the commutation rheology; the criterion layer is the scheme adopted to achieve the evaluation; and the indicator layer is the corresponding evaluation index obtained through the scheme.

[0139] By establishing a judgment matrix for mutual comparison and giving a 1-9 scale, as shown in Table 1:

[0140] Table 1 Meaning of scale

[0141]

[0142] In R, a ijIt is the ratio of the importance of i to j.

[0143] In the AHP, the maximum eigenvalue λ is calculated max When the eigenvector W is approximated, the accuracy requirement is not high and an approximate method can be used to obtain the required parameters. The present invention adopts the square root method.

[0144] First do the square root calculation:

[0145]

[0146] Then normalize it:

[0147]

[0148] Let the approximate eigenvector W be:

[0149] W=[w1 w2 … w n ] T

[0150] Calculate the maximum eigenvalue λ of the judgment matrix A max :

[0151]

[0152] For the evaluation of the key performance of the converter transformer, the judgment matrix A of the R layer to the G layer is calculated, then:

[0153] Table 2 Evaluation matrix of R layer to G layer

[0154]

[0155] Calculate the evaluation matrix of the indicator layer P relative to the criterion layer R:

[0156] Table 3 Evaluation matrix of indicator layer P relative to criterion layer R1

[0157]

[0158] Table 4 Evaluation matrix of indicator layer P relative to criterion layer R2

[0159]

[0160]

[0161] Table 5 Evaluation matrix of indicator layer P relative to criterion layer R3

[0162]

[0163] Table 6 Evaluation matrix of indicator layer P relative to criterion layer R4

[0164]

[0165] In addition, the matrix A must be subjected to a consistency check. First, the consistency index is calculated:

[0166]

[0167] The larger the C1 value, the more it deviates from the consistency, and the smaller the C1 value, the closer it is to the consistency. As shown in Table 7, the R1 obtained by calculating the positive reciprocal matrix 1000 times includes 1-10 orders.

[0168] Table 7 Average Stochastic Index R1

[0169]

[0170] One-time ratio C R , if C R <0.1, then A meets the consistency condition. If C R ≥0.1, then A does not meet the consistency condition. For cases that do not meet the requirements, a ij Revalue and recalculate so that C R <0.1, so that A meets the consistency condition.

[0171] Calculate the composite weight of the factors of each layer on R and then sort them.

[0172]

[0173] In the formula, is the total ranking weight of R, and m is the number of factors of R.

[0174] Therefore, the total sorting vector of P relative to G can be found The consistency ratio is:

[0175]

[0176] For the evaluation criteria for the key performance of converter transformers, six criteria are used: {excellent, good, medium, poor, relatively poor, extremely poor}, and the corresponding score set is {6, 5, 4, 3, 2, 1}.

[0177] Through actual testing of the commutation transformer, the obtained indicators are used to evaluate the performance of the commutation transformer. According to the testability, the scores of each indicator are comprehensively evaluated and the fuzzy evaluation matrix is ​​obtained:

[0178]

[0179] In the formula, m is the number of indicators in the indicator layer; n is the number of comment sets; i is the serial number of the single factor, indicating a certain stage of the test design.

[0180] therefore:

[0181]

[0182] By calculating the fuzzy evaluation comprehensive index, the comprehensive evaluation value B is obtained:

[0183]

[0184] The scores are calculated according to the principle of maximum membership, and the key performance evaluation level of the converter transformer is obtained based on the score set of the aforementioned comments.

[0185] The present invention combines multi-source heterogeneous data, proposes a data fusion algorithm, and realizes comprehensive perception of the thermal distribution and mechanical stress performance of the converter transformer. The existing technology has a relatively single data source for the evaluation of the performance of the converter transformer, and the data fusion technology has a poor effect. The present invention proposes a multi-source heterogeneous data fusion algorithm of BP neural network and support vector machine. The BP neural network is used to clean the data, preliminarily eliminate noise interference, and eliminate redundant data; the SVM algorithm is used to perform data fusion operations, and with its strong generalization ability and stable advantages, the optimal classification surface is found and trained to avoid falling into the local optimum and other situations, thereby ensuring the effectiveness of the classification, so that the constructed multi-source heterogeneous data fusion results are more reliable. The present invention focuses more on the study of the local thermodynamic and mechanical properties of the converter transformer during operation, and proposes a method for evaluating the thermal distribution and mechanical stress performance of the converter transformer. At present, there are few studies on the key performance evaluation methods of converter transformers. This invention selects temperature distribution uniformity, thermal response time, cooling efficiency, heat dissipation area and heat loss rate as thermal distribution performance indicators for the thermal distribution and mechanical stress performance evaluation of converter transformers, and selects stress, deformation, fatigue life, seismic resistance, connection strength and material yield strength as mechanical stress performance evaluation indicators. Compared with the traditional converter transformer performance evaluation, the data is more diversified. In addition, a method combining the hierarchical analysis method and the fuzzy comprehensive evaluation model method is proposed to realize the key performance evaluation of converter transformers.

[0186] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for evaluating key performance of converter transformers based on multi-source heterogeneous data, characterized by: The following steps are involved: Step 1: Data collection, including collecting converter performance evaluation index data through various inspection methods; Step 2: Multi-source heterogeneous data cleaning based on BP neural network, including obtaining similarity relationship constraints of multi-source heterogeneous data, building a BPNN data cleaning model based on the relationship constraints, and completing the cleaning of multi-source heterogeneous data through the data cleaning model; Step 3: Multi-source heterogeneous data fusion based on SVM, including extracting and processing the features of multi-source heterogeneous data according to the cleaning results of multi-source heterogeneous data, using SVM algorithm to obtain the optimal classification hyperplane of the features, building a data fusion model, and realizing the fusion of multi-source heterogeneous data through the established data fusion model; Step 4: Determine the performance evaluation index, including determining the thermal distribution performance index and mechanical stress performance evaluation index based on the fusion results of the multi-source heterogeneous data of the converter transformer, and constructing an evaluation system; Step 5: Determine the evaluation method for the key performance of the converter transformer, including evaluating the key performance of the converter transformer by integrating the hierarchical analysis method and the fuzzy comprehensive evaluation model method according to the evaluation system.

2. The method for evaluating key performance of converter transformers based on multi-source heterogeneous data according to claim 1 is characterized in that: In the step 2, the method for obtaining similarity relationship constraints of multi-source heterogeneous data includes the following steps: the original multi-source heterogeneous data exists in the form of a data table, the data table is set to b, the total amount of data is set to p, and the data table set B = {b1, b2, ..., b p }, the data attribute set is The Manhattan distance algorithm is used to obtain the similarity between multi-source heterogeneous data. The process is shown in the following formula: In the formula, the similarity between multi-source heterogeneous data i and data j is expressed as Indicates; sample data attributes are expressed as Represents; attribute feature mark of data i Form; attribute feature tag of data j Format; data dimension mark S format. The correlation coefficient between the data is set to Formulate the parameter threshold λ of the data and use it as the constraint condition of the data cleaning model. The process is shown in the following formula: Based on the constraints constructed above, a data cleaning model is constructed.

3. The method for evaluating key performance of converter transformers based on multi-source heterogeneous data according to claim 2 is characterized in that: In the step 2, a BPNN data cleaning model is constructed based on relational constraints. The method for cleaning multi-source heterogeneous data through the data cleaning model includes the following steps: setting the multi-source heterogeneous data cleaning model to T, and the model contains N layers of nodes, In the cleaning model of multi-source heterogeneous data, each layer of the RBM structure has visible layer and hidden layer data. The input layer data nodes are set to c = (α1, α2, ..., α c }, the number of hidden layer nodes is d = {β1, β2, …, β d }, in order to construct the node weight matrix between the levels of the cleaning model of multi-source heterogeneous data, the process is shown in the following formula: In the formula, the constructed node weight matrix is ​​described in the form of ξ; the matrix vector is labeled ω cd In this form, based on the constructed node weight matrix combined with the RBM energy function, the node energy function of multi-source heterogeneous data is obtained. The process is shown in the following formula: In the formula, the node data α i The bias coefficient is δ i Represents; node data β j The bias coefficient is γ j Represents; data weight mark ω ij The form; constants are represented by i and j; the number of nodes of the input layer data in the model is represented by c; the energy function of the acquired data node is described as A(α i ,β j )form; The data distribution probability of the model input layer nodes and hidden layer nodes is tested based on the constructed node energy function. The test results are shown in the following formula: In the formula, the normalized coefficient of the data energy function is described as form; the distribution factor is described in the form of ε; the data distribution probability is expressed as Q(α i ,β j )express; Based on the above-obtained data distribution probability, the activity of the model node variables can be effectively obtained to optimize the cleaning model; multi-source heterogeneous data samples are input into the cleaning model, and the states of the data nodes in the middle layer of the model are independent of each other. Assuming that the hidden layer nodes β j When it is 1, the constraint conditions for obtaining the data distribution probability are as follows: In the formula, the constraints established are Representation; Binary Sigmoid function notation η form; Data state coefficient is expressed as Based on the above calculation results, it can be seen that when the value interval of the Sigmoid function independent variable is [0, 1], the probability distribution function of the data will present a single continuous change state; After activating the hidden layer nodes of the neural network model using the above calculation process, input β j Back propagation is performed on the model to realize model training. The node energy anomalies and discrete distribution data captured in the training results are removed to complete the cleaning of multi-source heterogeneous data of the converter transformer.

4. The method for evaluating key performance of converter transformers based on multi-source heterogeneous data according to claim 1 is characterized in that: In step 3, the multi-source heterogeneous data fusion algorithm of SVM includes the following steps: based on the above data cleaning results, the multi-source heterogeneous data set is updated to Y ij (t), the sample data volume is M, the data state is t, and the mean standardization method is used to standardize the data. The result is shown in the following formula: In the formula, the standardized result of the data is marked Format; Data mean labeling Format; constant marker i, j format. Based on the above data, the standardized processing results are obtained, and the feature vector of multi-source heterogeneous data i at the jth attribute is extracted. The calculation results are shown in the following formula: In the formula, the feature vector of multi-source heterogeneous data is expressed as Indicates; the data attribute recording period is expressed by T.

5. The method for evaluating key performance of converter transformers based on multi-source heterogeneous data according to claim 4 is characterized in that: In step 3, the multi-source heterogeneous data fusion algorithm of SVM further includes the following steps: Set the acquired data feature vector as the input vector y i , the kernel function is K(·), the feature vector set is y, and the output vector is u i , in order to obtain the optimal hyperplane of the data, the process is as follows: u=τ i you i K(y i ,y)+θ In the formula, the optimal hyperplane is represented by u; the weighting coefficient is τ i Indicated by; the bias term is denoted by θ; the number of eigenvectors is denoted by m; the constant is marked in the form of i; Based on the obtained optimal super-classification surface, the multi-source heterogeneous data is divided into h groups of data, and the data of different groups come from different sensors. The actual value of the data is set as u, the training output is U, and the generalization parameter is χ, so as to establish the fusion model of multi-source heterogeneous data of SVM, as shown in the following formula: In the formula, the constructed data fusion model is represented by J; The optimal solution for data i and j is labeled l i , l i Format; optimal solution mean label Form; Constraint Tagging form; the number of data sources is represented by g; Based on the determined optimal solution of the data, the decision function of the data is established, the model is solved, and the data fusion is realized. The process is shown in the following formula: In the formula, the constructed decision function is represented by f(x); the data mean set of data source g is labeled form; data optimal solution mark l g form.

6. The method for evaluating key performance of converter transformers based on multi-source heterogeneous data according to claim 1, characterized in that: In step 4, the thermal distribution performance indicators include temperature distribution uniformity, thermal response time, cooling efficiency, heat dissipation area and heat loss rate; the mechanical stress performance evaluation indicators include stress, deformation, fatigue life, seismic resistance, connection strength and material yield strength.

7. The method for evaluating key performance of converter transformers based on multi-source heterogeneous data according to claim 1 is characterized in that: In the step five, Calculate the maximum eigenvalue λ max When the eigenvector W is approximated, the accuracy requirement is not high, and the required parameters can be obtained by using an approximate method. The present invention uses the square root method; First do the square root calculation: Then normalize it: Let the approximate eigenvector W be: In=[in1 in2…in n ] T Calculate the maximum eigenvalue λ of the judgment matrix A max : For the evaluation of the key performance of the converter transformer, the evaluation matrix A of the R layer to the G layer is calculated; the evaluation matrix of the index layer P relative to the criterion layer R is calculated: In addition, the matrix A must be subjected to a consistency check. First, the consistency index is calculated: The larger the C1 value, the further away from consistency, and the smaller the C1 value, the closer to consistency. As shown in Table 7, the R1 is obtained by calculating the positive reciprocal matrix 1000 times; the one-time ratio C R , if C R <0.1, then A meets the consistency condition. If C R ≥0.1, then A does not meet the consistency condition; for those that do not meet the requirements, a ij Revalue and recalculate so that C R <0.1, so that A meets the consistency condition; Calculate the composite weight of the factors of each layer on R and then sort them. In the formula, is the total ranking weight of R, and m is the number of factors of R; The total ordering vector of P relative to G can be found The consistency ratio is: For the evaluation of the key performance of the converter transformer, six evaluation criteria are used: {excellent, good, medium, poor, relatively poor, extremely poor}, and the corresponding score set is {6, 5, 4, 3, 2, 1}; Through actual testing of the commutation transformer, the obtained indicators are used to evaluate the performance of the commutation transformer, and the fuzzy evaluation matrix is ​​obtained based on the comprehensive evaluation of the test results and the scoring statistics of each indicator: In the formula, m is the number of indicators in the indicator layer; n is the number of comment sets; i is the serial number of a single factor, indicating a certain stage of the test design; therefore: By calculating the fuzzy evaluation comprehensive index, the comprehensive evaluation value B is obtained:

Images