Booster station main transformer state evaluation method based on knowledge and data

Abstract

A method for evaluating the state of a main transformer of a booster station based on knowledge and data comprises the following steps: acquiring basic ledger information and historical maintenance records of the main transformer in a wind power booster station on a road, and selecting corresponding test parameters as evaluation indexes; constructing a transformer knowledge and data dual-driven power key equipment state evaluation model, determining subjective and objective weights by adopting an AHP-entropy weight method in the knowledge-driven model, determining the membership degree of each parameter by selecting a cloud model, and obtaining a proper comment through fuzzy operation; the data-driven model generates an initial data source by using a knowledge graph, oversamples a data set by using SMOTE to balance the integrity of a sample, reduces the noise of a synthetic sample and the influence of a boundary sample by using a self-NaN-DE algorithm, selects WOA-SVM classification evaluation, and updates the data source by using a mechanism model; and evaluating the operation state of the booster station main transformer by using the constructed state evaluation model. The method has the advantages of being more accurate in transformer state evaluation, wider in application range, higher in efficiency and the like.

Description

Technical Field

[0001] The present invention relates to the field of state evaluation of main transformers of wind power stations on new energy roads, and in particular to a knowledge- and data-based method for evaluating the state of main transformers of booster stations. Background Art

[0002] The new power system with new energy as the main body has put forward higher requirements for the state perception and operation and maintenance of traditional power equipment. With the large-scale grid connection of new energy wind farms, the number of main transformers has also increased, and the reliable operation of power equipment faces huge challenges. By developing a service health evaluation system for key equipment of new energy power generation, it helps operation and maintenance personnel to more conveniently and quickly grasp the actual working conditions of the main transformer and provide necessary maintenance opinions, bringing considerable economic benefits to the field of new energy power generation. However, the existing main transformer state evaluation methods mostly adopt the mechanism modeling method, the model division is too fine, and the acquisition of subjective factors makes it difficult in practical application, resulting in inaccurate state evaluation. In the context of big data, the normal operation data and abnormal data of the main transformer have been accumulated to a certain extent. The mining and learning of the newly stored data can improve the rationality and accuracy of the evaluation model to a certain extent. Summary of the invention

[0003] The technical problem to be solved by the present invention is to provide a knowledge- and data-based method for evaluating the status of a main transformer in a substation, which has more accurate status evaluation, wider application range and higher efficiency.

[0004] In order to solve the above technical problems, the technical solution adopted by the present invention is: a method for evaluating the state of a main transformer of a booster station based on knowledge and data, comprising the following steps:

[0005] A method for evaluating the state of a main transformer in a booster station based on knowledge and data comprises the following steps:

[0006] Step 1: Obtain the basic ledger information of the onshore wind power main transformer and the historical maintenance records of the same period, and select the corresponding test parameters as evaluation indicators;

[0007] Step 2: Construct a knowledge-driven and data-driven hierarchical condition assessment model for power transformers. The knowledge-driven model uses the AHP-entropy weight algorithm to determine the subjective and objective weights, selects a suitable cloud model to determine the membership of each indicator, and then uses fuzzy calculation to obtain the transformer evaluation. The data-driven model uses an expert system to generate the initial data source, selects the SMOTE algorithm to balance the sample data set, and then uses the self-NaN-DE algorithm to reduce the impact of synthetic sample noise and boundary samples.

[0008] Step 3: Use WOA-SVM classification evaluation, and then use the mechanism model to update the data source. Finally, use the constructed state evaluation model to evaluate the operating state of the main transformer of the booster station.

[0009] In step 2, the knowledge-driven model uses the AHP algorithm to calculate the subjective weights, and uses the entropy weight method to arrange the main indicators according to the discrete degree of information entropy to obtain the objective weights. The subjective weights are combined with the objective weights to obtain the combined weights; the cloud model is selected to determine the normalized membership of different indicators: the membership of the main transformer is based on the bell-shaped membership function, and the digital characteristics of the cloud model are introduced to obtain the membership; the multi-level fuzzy comprehensive evaluation method is used to perform fuzzy operations on the weight and membership matrix, and the transformer evaluation is obtained by mechanism modeling according to the maximum membership criterion.

[0010] The method for obtaining the combined weight comprises the following steps: performing linear addition combination weighting on the subjective weight A and the objective weight B, that is, the combined weight In the formula, k1, k2∈[0,1], k1, k2 are taken according to actual requirements but must satisfy k1+k2=1.

[0011] The boundary parameters of the subordinate cloud model calculate the correlation degree according to the different probabilities of different cloud droplets appearing in the data normal cloud, which improves the accuracy of the correlation calculation: Input the test data x of n indicators i And calculate the expectation E in turn x 、Entropy E n and super entropy H e , enter the E required by historical experience summary in the forward cloud generator x 、E n and H e And cloud droplets are continuously generated. Since the dimensions of each transformer data are different, each indicator needs to be normalized and converted into the evaluation standard Zhengtai cloud:

[0012] E x =(c min +c max ) / 2

[0013] E n =(c max -c min ) / 6

[0014] In the formula, C min , C max H is the upper and lower boundaries of the indicator level; e is a constant and can be adjusted according to the fuzziness of specific indicators. The parameters of the evaluation grade cloud model are obtained according to the above formula. The membership of the step-up transformer is based on the bell-shaped membership function, and the digital characteristics of the cloud model are introduced as follows:

[0015]

[0016] In the formula, E nn For E nThe expectation of the cloud droplets is determined by the number of cloud droplets, which determines the accuracy of the qualitative description of the generated cloud image. The membership degree is obtained after calculating 1000 times and averaging.

[0017] Step 2 The specific steps are as follows:

[0018] Step 2.1: SMOTE synthetic samples:

[0019] Step 2.2: Detect noise samples using natural nearest neighbor theory:

[0020] Step 2.3: Improve suspicious samples using differential algorithm.

[0021] In step 2.3, the position or attribute of the suspicious sample SE is optimized using the differential evolution algorithm, which includes four operations: initialization, mutation, crossover, and selection;

[0022] Initialization: DE starts with an initial set of suspicious examples SE, where each suspicious example is called with D-dimensional attributes and a specified class label w i The target vector T_x i :

[0023]

[0024] Where i∈{1, 2, ..., N SE}, and N SE is the sample number represented by SE; g∈[1,2,...,G max ] represents the gth generation, G max represents the upper limit of the generation. Attribute d∈[1,2,...,D];

[0025] Mutation: It is the prime operator of DE and is used for each target vector T_x i,g Generate mutation vector V i,g ;

[0026]

[0027] In the formula, F i is the scaling factor that actively controls the different vectors; T_x rl,SE , T_x r2,SE and T_x r3,SE is randomly selected from SE and is related to T_x i,g have the same category; if there are not enough samples with the same category as T_x in SE i,g The same category label, it is recommended to generate artificial samples T_x rm,SE , m∈[1,2,3];

[0028]

[0029] In the formula, rand(-0.1, 0.1) means selecting a random real number from the interval [-0.1, 0.1];

[0030] Crossover: By linearly combining the target vector T_x i,g and mutation vector V i,g To generate a new test vector U i,g ;

[0031] U i,g =T_x i,g +K×(V i,g -T_x i,g )

[0032] In the formula, K is a random number, K∈[0,1]. By introducing V i,g Simplified public i,g ,Right now:

[0033] U i,g =T_x i,g +K×(T_x r1,SE -T_x i,g )+F i ×(T_x r2,SE -T_x r3,SE )

[0034] By modifying the target vector T_x i,g The attributes or positions of the new trial vector U i,g Therefore, U i,g With T_x i,g Same category label;

[0035] Selection: The selection operator determines U i,g Whether to survive in the next generation g+1:

[0036]

[0037] In the formula, if T_x i,g is misclassified by a given classifier C trained on normal examples NE, then U i,g will survive in the next generation, and vice versa; when all target vectors are correctly classified in the gth generation, the iterative process of DE terminates.

[0038] In step 3,

[0039] Firstly, the WOA optimized SVM algorithm is used to train, predict and classify the sample data to obtain the transformer operating condition information;

[0040] Secondly, the SMOTE algorithm is used to balance the sample set;

[0041] Again, the SMOTE algorithm is used to expand the statistically obtained small amount of abnormal and severe test data, where the F i Affects its convergence speed and optimal performance, and its values ​​are as follows: rand1, rand2 and rand3 are uniform pseudo-random numbers between 0 and 1; SFGSS = 8, SFHC = 20, F l =0.1, F u =0.9, τ1=0.1, τ2=0.03, τ3=0.07,

[0042]

[0043] Finally, the whale optimization algorithm is selected to quickly optimize the penalty factor c and kernel parameter g of SVM.

[0044] The operating status of the transformer includes {excellent (V1), good (V2), caution (V3), abnormal (V4)}; excellent means that the indicators of the transformer fluctuate within the specified range, away from the warning value, and can continue to operate; caution means that the transformer indicators are developing towards the limit value, and can continue to operate, but requires enhanced monitoring and planned maintenance; abnormal means that the transformer indicators change greatly, approaching or slightly exceeding the standard limit, and power outage and maintenance should be arranged appropriately; serious means that one or more indicators of the transformer seriously exceed the standard limit, and power outage and maintenance must be arranged as soon as possible.

[0045] The equalization processing method includes the following steps: linearly combining the minority class samples in the abnormal state in the original sample set and their neighboring samples into new samples to improve the low recognition accuracy caused by the uneven distribution of sample states:

[0046] Step 1: Randomly select a B from the K nearest neighbor minority classes of the minority class sample A. The difference vector of the sample features of A and B is (B-A).

[0047] Step 2: Randomly select a real number i from the interval (0, 1) as the weight, and multiply the weight i with the difference vector to obtain i(B-A);

[0048] Step 3: Add the result of Step 2 to the feature vector of sample A to obtain the synthetic sample A+i(B-A).

[0049] WOA simulates the behaviors of humpback whales in surrounding, attacking, and searching for prey, and performs parameter optimization to obtain the optimal solution for prey. The basic principles are as follows:

[0050] 1) Encirclement and predation stage: When the prey is discovered by the humpback whale, the mathematical model of the humpback whale encircling the prey is:

[0051] X(t+1)=X * (t)-A|CX *(t)-X(t)|

[0052] A=2ar-a

[0053] C=2r

[0054] Where: t is the current iteration number; X* is the current prey position; X is the whale position vector; A and C are coefficient vectors; is the convergence factor, which decreases linearly from 2 to 0 as the number of iterations increases; r is a random vector between 0 and 1;

[0055] 2) Bubble net attack stage: The bubble net attack behavior of humpback whales is modeled using the contraction and encirclement mechanism and the spiral position update method. A random probability p of 0 to 1 is generated to determine which predation method to use. When p < 0.5, the contraction and encirclement mechanism is selected; when p ≥ 0.5, the spiral motion position update method is selected.

[0056] 3) Random search predation stage: When |A|>1, the whale searches for more suitable prey in a natural search mode, where X rand Indicates the random position of the current whale group individual;

[0057] X(t+1)=X rand -A|CX rand -X|.

[0058] The present invention adopts the above technical solution, in the multi-source heterogeneous state assessment data, uses the initially constructed data set and the gradually improved fault case library, and uses the causal thinking and empirical model to perceive the operation status of the equipment. By using knowledge-assisted data driving, some domain knowledge and expert experience are reflected in the data-driven model, providing it with some features with mechanism knowledge, thereby improving the adaptability and reliability of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0060] Figure 1 is a flow chart of the method described in an embodiment of the present invention;

[0061] Figure 2 is a flow chart of SMOTE-NaN-DE according to an embodiment of the present invention;

[0062] Figure 3 is a flowchart of WOA optimization SVM according to an embodiment of the present invention; DETAILED DESCRIPTION

[0063] The following is a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the drawings 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.

[0064] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0065] Overall, such as Figure 1 As shown, the embodiment of the present invention discloses a transformer state evaluation method based on knowledge and data dual drive, comprising the following steps:

[0066] Step 1: Obtain the basic ledger information of the onshore wind power main transformer and the historical maintenance records of the same period, and select the corresponding test parameters as evaluation indicators;

[0067] Step 2: Construct a knowledge-driven and data-driven hierarchical condition assessment model for power transformers. The knowledge-driven model uses the AHP-entropy weight algorithm to determine the subjective and objective weights, selects a suitable cloud model to determine the membership of each indicator, and then uses fuzzy calculation to obtain the transformer evaluation. The data-driven model uses an expert system to generate the initial data source, selects the SMOTE algorithm to balance the sample data set, and then uses the self-NaN-DE algorithm to reduce the impact of synthetic sample noise and boundary samples.

[0068] In step 2, the knowledge-driven model uses the AHP algorithm to calculate the subjective weights, and uses the entropy weight method to arrange the main indicators according to the discrete degree of information entropy to obtain the objective weights. The subjective weights are combined with the objective weights to obtain the combined weights; the cloud model is selected to determine the normalized membership of different indicators: the membership of the main transformer is based on the bell-shaped membership function, and the digital characteristics of the cloud model are introduced to obtain the membership; the multi-level fuzzy comprehensive evaluation method is used to perform fuzzy operations on the weight and membership matrix, and the transformer evaluation is obtained by mechanism modeling according to the maximum membership criterion.

[0069] The method for obtaining the combined weight comprises the following steps:

[0070] The subjective weight A and the objective weight B are linearly added together to give weights, that is, the combined weight In the formula, k1, k2∈[0,1], k1, k2 are taken according to actual requirements but must satisfy k1+k2=1.

[0071] The boundary parameters of the membership cloud model calculate the correlation degree according to the different probabilities of different cloud droplets appearing in the data normal cloud, thereby improving the accuracy of the correlation calculation. i And calculate the expectation E in turn x 、Entropy E n and super entropy H e Enter the E required by historical experience summary in the forward cloud generator x 、E n and H e And cloud droplets are continuously generated. Since the dimensions of each transformer data are different, each indicator needs to be normalized and converted into the evaluation standard Zhengtai cloud:

[0072] E x =(c min +c max ) / 2

[0073] E n =(c max -c min ) / 6

[0074] In the formula, C min , C max H is the upper and lower boundaries of the indicator level; e is a constant and can be adjusted according to the fuzziness of specific indicators; the parameters of the evaluation level cloud model are obtained according to the above formula. The membership of the step-up transformer is based on the bell-shaped membership function, and the digital characteristics of the cloud model are introduced as follows:

[0075]

[0076] In the formula, E nn For E n The number of cloud droplets determines the accuracy of the qualitative description features of the generated cloud image. The membership degree is obtained after calculating 1000 times and averaging.

[0077] The boundary parameters of the membership cloud model are shown in Table 1. The evaluation parameters of the knowledge-driven model include 12 evaluation indicators, including trace water content C1, oil dielectric loss C2, breakdown voltage C3, furfural content C4, absorption ratio C5, winding dielectric loss C6, leakage current C7, insulation resistance C8, hydrogen content C9, acetylene content C10, total hydrocarbon content C11 and CO relative gas production rate C12.

[0078] Table 1 Boundary parameters of the membership cloud model

[0079]

[0080] In step 2, if Figure 2As shown in Figure 1, the data-driven model inputs data statistics into the expert system to generate sample labels, uses SMOTE to simulate synthetic samples that are consistent with the distribution and change rules of minority class samples as much as possible, and uses the NaN-DE algorithm to reduce the impact of synthetic sample noise and boundary samples, thereby increasing the number of samples while retaining the original information. The specific process is as follows: Figure 2 As shown, the specific steps are as follows:

[0081] Step 2.1: SMOTE synthetic samples:

[0082] Step 2.2: Detect noise samples using natural nearest neighbor theory:

[0083] Step 2.3: Improve suspicious samples using differential algorithm.

[0084] In step 2.2, when each sample has at least one true friend, a harmonious social state is formed in the dataset, which is called a natural neighbor stable structure. Let x be a training set with n samples and D-dimensional features, x i and x j is a subset of x, with:

[0085]

[0086] In the formula, NN r (x i ) is the sample x i The set of r nearest neighbors of r (x j ) is the sample x j Then continue to search for the r nearest neighbors of each sample until each sample has at least one mutual r nearest neighbor and forms the NSS.

[0087] Definition 1: Natural Neighbor Eigenvalue (NaNE) If a natural neighbor stable structure is formed, the natural neighbor eigenvalue λ is equal to the search round r.

[0088]

[0089] Definition 2: If sample x is a natural neighbor (NaN) of sample y, then sample x is one of the λ nearest neighbors of sample y, and sample y is one of the λ nearest neighbors of sample x.

[0090]

[0091] Definition 3: If a sample x i belongs to the suspicious sample set (SE), then one of its natural neighbors has a different class label from it.

[0092]

[0093] In the formula, SE = {x j ,...} is a set of suspicious samples, which contains noise and boundary samples destroyed by noise; NaN(x i )={x j ,...} is the sample x i The set of natural neighbors of i ) returns sample x i The class label w i ; l(x j ) returns sample x j The class label w j .

[0094] Definition 4: If a sample x i belongs to the normal sample set (NE), then it has the same class label as any of its natural neighbors.

[0095]

[0096] In the formula, NE={x j ,...} is the set of normal samples. In Definition 3-4, NaN is used for error detection to find suspicious and normal samples. Normal samples are safe and rarely affected by noise and boundary samples.

[0097] In step 2.3, the differential evolution algorithm is used to optimize the position or attributes of the suspicious samples SE (i.e., noise and boundary samples). Generally speaking, DE includes four operations, i.e., initialization, mutation, crossover and selection, which are explained one by one below.

[0098] Initialization: DE starts with an initial set of suspicious examples SE, where each suspicious example is called with D-dimensional attributes and a specified class label w i The target vector T_x i :

[0099]

[0100] Where i∈{1, 2, ..., N SE}, and N SE is the sample number represented by SE. g∈[1,2,…,G max ] represents the gth generation, G max Represents the upper limit of the generation. Attribute d∈[1,2,…,D].

[0101] Mutation: It is the prime operator of DE and is used for each target vector T_x i,g Generate mutation vector V i,g Among the frequently mentioned mutation strategies, we chose DE because of its prevalence and robustness to generate mutant vectors as follows.

[0102]

[0103] In the formula, F i is the scaling factor that actively controls the different vectors; T_x rl,SE , T_x r2,SE and T_x r3,SE is randomly selected from SE and is related to T_x i,g If there are not enough samples with the same category as T_x in SE i,g The same category label, it is recommended to generate artificial samples T_x rm,SE , m∈[1,2,3].

[0104]

[0105] In the formula, rand(-0.1, 0.1) means selecting a random real number from the interval [-0.1, 0.1].

[0106] Crossover: By linearly combining the target vector T_x i,g and mutation vector V i,g To generate a new test vector U i,g Specifically, the arithmetic crossover strategy is as follows:

[0107] U i,g =T_x i,g +K×(V i,g -T_x i,g )

[0108] In the formula, K is a random number, K∈[0,1]. By introducing V i,g Simplified public i,g ,Right now:

[0109]

[0110] By modifying the target vector T_x i,g The attributes or positions of the new trial vector U i,g Therefore, U i,g With T_x i,g same class label. However, we should check that U i,g Whether the attribute value exceeds the range [min(T_X i，g ), max(T_X i，g )]. The inspection conditions are described as follows:

[0111]

[0112] Selection: The selection operator determines U i,gWhether to survive in the next generation g+1:

[0113]

[0114] In the formula, if T_x i,g is misclassified by a given classifier C trained on normal examples NE, then U i,g will survive in the next generation, and vice versa. In other words, when T_x i,g When correctly classified by classifier C, T_x i,g has been optimized and does not change in the next generation. Note that the iterative process of DE terminates when all target vectors are correctly classified in the gth generation.

[0115] Step 3: Use WOA-SVM classification evaluation. The specific process is as follows: Figure 3 As shown. Then the mechanism model is used to update the data source. Finally, the constructed state assessment model is used to assess the operating state of the main transformer of the booster station. Specifically:

[0116] First, the WOA optimized SVM algorithm is combined to train, predict and classify the sample data to obtain the transformer operating condition information. The operating status of the transformer includes {excellent (V1), good (V2), caution (V3), abnormal (V4)}; excellent means that the indicators of the transformer fluctuate within the specified range, far away from the warning value, and can continue to operate; caution means that the indicators of the transformer are developing towards the limit value, and can continue to operate, but it is necessary to strengthen monitoring and overhaul according to the plan; abnormal means that the indicators of the transformer change greatly, close to or slightly exceed the standard limit, and power outage maintenance should be arranged appropriately; severe means that one or more indicators of the transformer seriously exceed the standard limit, and power outage maintenance must be arranged as soon as possible.

[0117] Secondly, the method of using the SMOTE algorithm to balance the sample set includes the following steps: linearly combining the minority class samples in the abnormal state and their neighboring samples in the original sample set into new samples to improve the low recognition accuracy caused by the uneven distribution of sample states.

[0118] Step 1: Randomly select a B from the K nearest neighbor minority classes of the minority class sample A. The difference vector of the sample features of A and B is (B-A).

[0119] Step 2: Randomly select a real number i from the interval (0, 1) as the weight. Multiply the weight i with the difference vector to get i(B-A).

[0120] Step 3: Add the result of Step 2 to the feature vector of sample A to obtain the synthetic sample A+i(B-A).

[0121] Again, the SMOTE algorithm is used to expand the statistically obtained small amount of abnormal and severe test data, where the F i Affects its convergence speed and optimal performance, and its values ​​are as follows: rand1, rand2 and rand3 are uniform pseudo-random numbers between 0 and 1; SFGSS = 8, SFHC = 20, F l =0.1, F u =0.9, τ1=0.1, τ2=0.03, τ3=0.07.

[0122]

[0123] Finally, the whale optimization algorithm is selected to quickly optimize the penalty factor c and kernel parameter g of SVM. The optimization process is as follows: Figure 3 As shown. The number of whales is set to 30, the maximum number of iterations is 100, and the average classification accuracy of the support vector machine under cross-validation is used as the fitness function to determine the fitness value of each individual in the whale group, and the minimum value is used as the optimal position of the current individual. Then, the whale individual is determined to perform spiral or encircling movement based on the randomly generated number p, and finally the C and g corresponding to the best fitness in the search space are cycled 100 times. Finally, the sample data is trained, predicted and classified in combination with the SVM algorithm to obtain the transformer operating condition information. To improve the accuracy of state recognition.

[0124] WOA simulates humpback whales’ behaviors of surrounding, attacking, and searching for prey, and performs parameter optimization to obtain the optimal solution for prey. Its basic principles are as follows:

[0125] 1) Encirclement and predation stage: When the prey is discovered by the humpback whale, the mathematical model of the humpback whale encircling the prey is:

[0126] X(t+1)=X * (t)-A|CX * (t)-X(t)|

[0127] A=2ar-a

[0128] C=2r

[0129] Where: t is the current iteration number; X* is the current prey position; X is the whale position vector; A and C are coefficient vectors; is the convergence factor, which decreases linearly from 2 to 0 as the number of iterations increases; r is a random vector between 0 and 1.

[0130] 2) Bubble net attack stage: The bubble net attack behavior of humpback whales is modeled using the contraction and encirclement mechanism and the spiral position update method. A random probability p of 0 to 1 is generated to determine which predation method to use. When p<0.5, the contraction and encirclement mechanism is selected; when p≥0.5, the spiral motion position update is selected.

[0131] 3) Random search predation stage: When |A|>1, the whale searches for more suitable prey in a natural search mode, where X rand Indicates the random position of the current whale group individual.

[0132] X(t+1)=X rand -A|CX rand -X|. Specific embodiment:

[0134] The historical fault data and test management data of the main transformers of each new energy onshore wind power station were selected. Among them, 60 excellent samples, 30 attention samples, 20 abnormal samples and 10 severe samples, a total of 120 groups, were used as the initial data source. Finally, the first 70% of the data was used for training, and the last 30% of the data was sent to the trained model for status evaluation.

[0135] Table 2 Comparison of multiple model state evaluation accuracy

[0136]

[0137] Table 2 shows the classification evaluation accuracy of the test set of different training models before and after sampling. It can be seen that the data evaluation in this paper uses NaN-DE to improve the SMOTE oversampled data set, and uses WOA to optimize the SVM classification evaluation. Its recognition accuracy is higher than other models.

[0138] The multi-level fuzzy comprehensive evaluation of the knowledge-driven model includes three secondary indicators: oil test, electrical test and gas in oil, and 12 tertiary indicators: water content, oil dielectric loss, breakdown voltage, furfural content, absorption ratio, winding dielectric loss, leakage current, insulation resistance, hydrogen content, acetylene content, total hydrocarbon content and CO relative gas production rate.

[0139] The present invention selects a power transformer of type DZ9-QY-31500 / 110GY at a certain station as a verification object. The hydrogen content of 120μL / L, acetylene content of 0.83μL / L, total hydrocarbon content of 63μL / L and CO gas production rate of 74.5% / month in the gas module of the target transformer oil are obtained. The winding absorption ratio of 1.42%, winding dielectric loss of 0.33%, leakage current of 27mA and insulation resistance of 1339MΩ under the electrical test module. The factors of micro-water content of 11.79mg / L, oil dielectric loss of 2.65%, breakdown voltage of 46kV and furfural content of 0.38μL / L under the oil test module. A comprehensive evaluation of the equipment operation status is carried out through the status evaluation module.

[0140] The transformer status assessment is now carried out using the mechanism modeling method. The field measured data is brought into the cloud model to obtain the calculation results of the fuzzy relationship matrix:

[0141] Step 1: Calculation of gas in oil

[0142]

[0143]

[0144] Step 2: Electrical test calculations

[0145] C 绝缘介质 =(0.22,0.16,0.25,0.36)

[0146]

[0147] Step 3: Oil test calculation

[0148] R 油化试验 =(0.15, 0.32, 0.33, 0.2)

[0149]

[0150] Step 4: Knowledge Assessment Calculation

[0151] The evaluation results of gas in oil, electrical test and oil chemical test are used as the membership of the evaluation factors at this level to construct the evaluation matrix, then

[0152] R 变压器 =(0.29, 0.27, 0.44)

[0153] Combining weight distribution and corresponding fuzzy operators to perform fuzzy operations, we get

[0154]

[0155] In the multi-source heterogeneous state assessment data, with the application and exploration of artificial intelligence technology in power equipment fault diagnosis, state assessment and state prediction; and the application of this system to other new energy booster stations, the main transformer test data and the number of state assessments will continue to increase, and a knowledge and data joint assessment model can be formed in the later stage. That is, according to the initially constructed data set and the gradually improved fault case library, the causal thinking and empirical model are combined to perceive the operating status of the equipment. Using knowledge to assist data-driven, some domain knowledge and expert experience are reflected in the data-driven model, providing it with some features with mechanism knowledge, thereby improving the adaptability and reliability of the model.

Claims

1. A knowledge and data-based method for evaluating the status of a main transformer in a booster station, characterized by: The steps include: Step 1: Obtain basic inventory information and historical maintenance records of onshore wind power main transformers, and select corresponding test parameters as evaluation indicators; Step 2: Construct a knowledge-driven and data-driven hierarchical condition assessment model for power transformers. The knowledge-driven model uses the AHP-entropy weight algorithm to determine the subjective and objective weights, selects an appropriate cloud model to determine the membership of each indicator, and then performs fuzzy operations to obtain transformer evaluations. The data-driven model uses an expert system to generate the initial data source, selects the SMOTE algorithm to equalize the sample data set, and then uses the self-NaN-DE algorithm to reduce the impact of synthetic sample noise and boundary samples. Step 3: Use WOA-SVM classification evaluation, then use the mechanism model to update the data source. Finally, use the constructed state assessment model to evaluate the operating status of the main transformer of the booster station.

2. The method for evaluating the status of a main transformer in a booster station based on knowledge and data according to claim 1, characterized in that: In step 2, the knowledge-driven model uses the AHP algorithm to calculate the subjective weight, and uses the entropy weight method to arrange the main indicators according to the discrete degree of information entropy to obtain the objective weight. The subjective weight is combined with the objective weight to obtain the combined weight; the cloud model is selected to determine the normalized membership of different indicators: the membership of the main transformer is based on the bell-shaped membership function, and the digital characteristics of the cloud model are introduced to obtain the membership; the multi-level fuzzy comprehensive evaluation method is used to perform fuzzy operations on the weight and membership matrix, and the transformer evaluation is obtained by mechanism modeling according to the maximum membership criterion.

3. The method for evaluating the status of a main transformer in a booster station based on knowledge and data according to claim 2, characterized in that: The method for obtaining the combined weight comprises the following steps: performing linear addition combination weighting on the subjective weight A and the objective weight B, i.e., the combined weight In the formula, k1, k2∈[0,1], k1, k2 are taken according to actual requirements but must satisfy k1+k2=1.

4. The method for evaluating the status of a main transformer in a booster station based on knowledge and data according to claim 2, characterized in that: The boundary parameters of the membership cloud model are calculated based on the different probabilities of different cloud droplets appearing in the data normal cloud to improve the accuracy of the correlation calculation: Input the test data x of n indicators i And calculate the expectation E in turn x , Entropy E n and super entropy H e , enter the E required by historical experience summary in the forward cloud generator x 、E n and H e And continuously generate cloud droplets. Since the dimensions of transformer data are different, each indicator needs to be normalized and converted into the evaluation standard Shota cloud: E x =(c min +c max ) / 2 E n =(c max -c min ) / 6 Where C min 、C max The upper and lower boundaries of the indicator level; H e is a constant and can be adjusted according to the fuzziness of specific indicators. The parameters of the evaluation level cloud model are obtained according to the above formula. The membership of the step-up transformer is based on the bell-shaped membership function, and the digital characteristics of the cloud model are introduced as follows: Where, E nn For E n The number of cloud droplets determines the accuracy of the qualitative description features of the generated cloud image. The membership degree is obtained after calculating 1000 times and taking the average.

5. The method for evaluating the status of a main transformer in a booster station based on knowledge and data according to claim 1, characterized in that: Step 2: Step 2.1: SMOTE synthesis sample: Step 2.2: Detect noise samples using the natural nearest neighbor theory: Step 2.3: Improve suspicious samples using differential algorithm.

6. The method for evaluating the status of a main transformer in a booster station based on knowledge and data according to claim 5, characterized in that: In step 2.3, the position or attribute of the suspicious sample SE is optimized using the differential evolution algorithm, which includes four operations: initialization, mutation, crossover, and selection; Initialization: DE starts with an initial set of suspicious examples SE, where each suspicious example is called with D-dimensional attributes and a specified class label w i The target vector T_x i : Where, i∈{1, 2, ..., N SE }, and N SE is the sample number represented by SE; g∈[1,2,…,G max ] represents the gth generation, G max Represents the upper limit of the generation. Attribute d∈[1,2,…,D]; Mutation: It is the prime operator of DE and is used for each target vector T_x i,g Generate mutation vector V i,g ; Where, F i is the scaling factor that actively controls the different vectors; T_x rl,SE , T_x r2,SE and T_x r3,SE is randomly selected from SE and is the same as T_x i,g Have the same category; if there are not enough samples with T_x in SE i,g The same category label, it is recommended to generate artificial samples T_x rm,SE , m∈[1,2,3]; In the formula, rand(-0.1, 0.1) means selecting a random real number from the interval [-0.1, 0.1]; Cross: By linearly combining the target vector T_x i,g and mutation vector V i,g To generate a new test vector U i,g ; U i,g =T_x i,g +K×(V i,g -T_x i,g ) Where K is a random number, K∈[0,1]. By introducing V i,g Simplified public i,g ,Right now: By modifying the target vector T_x i,g The attributes or positions of the new trial vector U are generated i,g Therefore, U i,g With T_x i,g Same category label; Selection: The selection operator determines U i,g Whether to survive in the next generation g+1: In the formula, if T_x i,g is misclassified by a given classifier C trained on normal examples NE, then U i,g will survive the next generation and vice versa; The iterative process of DE terminates when all target vectors are correctly classified in the gth generation.

7. The method for evaluating the status of a main transformer in a booster station based on knowledge and data according to claim 1, characterized in that: In step 3, Firstly, the WOA optimized SVM algorithm is used to train, predict and classify the sample data to obtain the transformer operating condition information; Secondly, the SMOTE algorithm is used to balance the sample set; Again, the SMOTE algorithm is used to expand the statistically obtained small amount of abnormal and severe test data, where the F i Affects its convergence speed and optimal performance, and its values ​​are as follows: rand1, rand2 and rand3 are uniform pseudo-random numbers between 0 and 1; SFGSS = 8, SFHC = 20, F l =0.1, F u =0.9, τ1=0.1, τ2=0.03, τ3=0.07, Finally, the whale optimization algorithm is selected to quickly optimize the penalty factor c and kernel parameter g of SVM.

8. The knowledge and data-based method for evaluating the status of a booster station main transformer according to claim 7, characterized in that: The operating status of the transformer includes {Excellent (V1), Good (V2), Caution (V3), Abnormal (V4)}; Excellent means that the indicators of the transformer fluctuate within the specified range, away from the warning value, and can continue to operate; Caution means that the transformer indicators are developing towards the limit value, and can continue to operate, but requires strengthened monitoring and planned maintenance; Abnormal means that the transformer indicators change greatly, approaching or slightly exceeding the standard limit value, and power outage maintenance should be arranged appropriately; Severe means that one or more indicators of the transformer seriously exceed the standard limit value, and power outage maintenance must be arranged as soon as possible.

9. The method for evaluating the status of a main transformer in a booster station based on knowledge and data according to claim 7, characterized in that: The equalization method includes the following steps: linearly combining the minority class samples with abnormal status in the original sample set and their neighboring samples to form new samples, so as to improve the low recognition accuracy caused by the uneven distribution of sample status: Step 1: Randomly select a B from the K nearest neighbor minority class of the minority class sample A. The difference vector of the sample features of A and B is (B-A); Step 2: Randomly select a real number i from the interval (0, 1) as the weight, and multiply the weight i with the difference vector to obtain i(B-A); Step 3: Add the result of Step 2 to the feature vector of sample A to obtain the synthetic sample A+i(B-A).

10. The knowledge and data-based method for evaluating the status of a booster station main transformer according to claim 7, characterized in that: WOA simulates humpback whales' behaviors of surrounding, attacking, and searching for prey, and optimizes parameters to obtain the optimal solution for prey. The basic principles are as follows: 1) Encirclement and predation stage: When the prey is discovered by the humpback whale, the mathematical model of the humpback whale encircling the prey is: X(t+1)=X * (t)-A|CX * (t)-X(t)| A=2ar-a C＝2r Where: t is the current iteration number; X* is the current prey position; X is the whale position vector; A and C are coefficient vectors; is the convergence factor, which decreases linearly from 2 to 0 as the number of iterations increases; r is a random vector between 0 and 1; 2) Bubble net attack phase: The bubble net attack behavior of humpback whales was modeled using two methods: a contraction and encirclement mechanism and a spiral position update. The predation method was determined by generating a random probability p between 0 and 1. When p < 0.5, the contraction and encirclement mechanism was selected; when p ≥ 0.5, the spiral motion position update was selected. 3) Random search and predation stage: When |A|>1, the whale searches for more suitable prey in a natural search mode, where X ran d represents the random position of the current whale group individual; X(t+1)=X rand -A|CX rand -X|。

Images