Transformer oil chromatographic analysis method based on improved fuzzy clustering algorithm and LSTM algorithm

Abstract

The invention discloses a transformer oil chromatographic analysis method based on an improved fuzzy clustering algorithm and an LSTM (Long Short Term Memory) algorithm. The method comprises the following steps: acquiring sample data; s2, randomly dividing the sample data into a plurality of cluster sets; s3, obtaining an initial clustering center of the clustering cluster set; s4, acquiring a comprehensive distance between the initial clustering center of each clustering cluster set and the sample data about the Euclidean distance and the cosine distance; s5, the membership matrix is updated according to the comprehensive distance, the steps S3 to S4 are repeatedly and iteratively executed until the iteration convergence condition is met, and the final clustering center of the clustering set is obtained; and according to the final clustering center, determining a clustering result of the transformer fault type. S6, dividing the clustering result into a data training set and a data test set according to a preset proportion; s7, establishing an LSTM model and taking the LSTM model as a prediction model for transformer fault analysis; and obtaining an optimal prediction model based on the data training set and the data verification set. The problems that a large amount of manpower and financial resources are consumed in a traditional regular overhaul mode, particularly in remote areas with inconvenient traffic, overhaul work faces many challenges, and in addition, human errors possibly occur in the overhaul process, and equipment is damaged are solved.

Description

Technical Field

[0001] The invention relates to the technical field of power system technology, and in particular to a transformer oil chromatographic analysis method based on an improved fuzzy clustering algorithm and an LSTM algorithm. Background Art

[0002] Power transformers are the core equipment of the power system. The stability of their operation in the substation directly affects the safety of the power grid. If a transformer fails, it will not only cause equipment damage, but also cause load loss and huge economic losses. In order to ensure the safe operation of the main transformer and improve the reliability of power supply, in addition to selecting transformers with advanced technology and excellent quality, it is also necessary to continuously improve its operation, maintenance and overhaul level.

[0003] Internal faults that occur during transformer operation are usually not caused by a single reason, and are often accompanied by hot spots and partial discharge. In addition, the fault situation may continue to develop and change. In order to accurately determine whether the equipment has a fault and its severity, it is necessary to comprehensively consider factors such as the equipment's historical record, equipment characteristics, and external environment. The detection technology of fault gas production and normal gas production is closely related, and the two are inseparable.

[0004] Online monitoring technology plays an increasingly important role in the development of power grids and has attracted widespread attention from scholars at home and abroad in recent years. In order to ensure the safety and stability of transformers, the State Grid Corporation of China has set up a special maintenance company to regularly perform equipment maintenance. However, this regular maintenance not only consumes a lot of manpower and financial resources, but also faces many challenges, especially in remote areas with inconvenient transportation. In addition, human errors may occur during the maintenance process, resulting in equipment damage; and practical experience shows that transformers may still fail even during the maintenance cycle. Summary of the invention

[0005] The present invention provides a transformer oil chromatographic analysis method based on an improved fuzzy clustering algorithm and an LSTM algorithm to overcome the above technical problems.

[0006] In order to achieve the above object, the technical solution of the present invention is:

[0007] A transformer oil chromatographic analysis method based on an improved fuzzy clustering algorithm and an LSTM algorithm specifically comprises the following steps:

[0008] S1: pre-processing the acquired fault gas component data dissolved in transformer oil to obtain sample data;

[0009] S2: Randomly divide the sample data into several cluster sets;

[0010] S3: Obtain the initial membership matrix of the clustering clusters, and improve the fuzzy clustering algorithm by introducing fuzziness to confirm the initial clustering center of the clustering clusters;

[0011] S4: Obtain the Euclidean distance and cosine distance between the initial cluster center of each cluster set and the sample data, and obtain the comprehensive distance based on the Euclidean distance and cosine distance;

[0012] S5: updating the membership matrix according to the comprehensive distance, and repeatedly iteratively executing steps S3 to S4 until an iterative convergence condition is reached, and obtaining the final cluster center of the cluster set;

[0013] And confirm the clustering results of transformer fault types according to the final cluster center;

[0014] S6: Dividing the clustering result into a data training set and a data test set according to a preset ratio;

[0015] S7: Build an LSTM model and use it as a prediction model for transformer fault analysis;

[0016] The prediction model is trained and verified based on the data training set and the data verification set to obtain the optimal prediction model, and then the fault gas components dissolved in the transformer oil are predicted through the optimal prediction model, and the oil chromatography analysis of the transformer is realized according to the prediction results.

[0017] Furthermore, the preprocessing described in S1 is specifically as follows:

[0018] The abnormal distribution data of the fault gas component data inside the transformer are obtained by using box plots and scatter plots; the blank items in the abnormal distribution data are filled with linear interpolation to obtain optimized data;

[0019] The z-score method is used to calculate and obtain the z-score value of each data point in the optimized data, and the data whose z-score value does not meet the preset data point threshold are removed as outliers to obtain the final sample data.

[0020] Furthermore, the S3 specifically includes the following steps:

[0021] S31: Obtain the initial membership matrix U of the cluster set = {u i1 ,u i2 ,…,u ik ,…,u ic};

[0022] The sample data is defined as an n-tuple data set X = {x1, x2, ..., x i ,…,x n}, and randomly divide X into c clusters with 2≤c≤n; xi represents the concentration data of transformer fault gas components obtained by the i-th gas detection, and x i ={x i1 ,x i2 ,…,x ij ,…,x is}, i=1,2,…,n,j=1,2,…,s

[0023] j represents the type of transformer fault gas components, including at least hydrogen, methane, ethane, ethylene, acetylene, carbon monoxide and carbon dioxide;

[0024] S32: By introducing fuzziness to improve the fuzzy clustering algorithm, the initial cluster center V of the cluster set is obtained according to the initial membership matrix, and its expression is:

[0025] V={v1,v2,…,v k ,v j …,v c}

[0026]

[0027] Where: m is the fuzzy weighting parameter, which represents the fuzziness of the initial membership matrix U; u ik represents the membership of the i-th sample to the k-th cluster; v j Represents the initial cluster center of the jth cluster set.

[0028] Furthermore, the S4 specifically includes the following steps:

[0029] S41: Get the Euclidean distance and cosine distance between the initial cluster center of each cluster and the sample data, and its expression is:

[0030]

[0031]

[0032] Where: represents the cosine distance of the angle between the i-th sample and the k-th cluster center, and its value range is [0,2]; d ik represents the Euclidean distance between the initial cluster center and the sample data; s represents the number of gas types during gas detection; v k represents the initial cluster center of the kth cluster; v kj Represents the cosine distance between the initial cluster center of the kth cluster and the jth sample;

[0033] S42: Get the comprehensive distance based on Euclidean distance and cosine distance, its expression is

[0034]

[0035]

[0036] Where: ik It represents the Euclidean-cosine distance for comprehensive indicators with similar values ​​and shapes; σ and 1-σ represent the weights of the Euclidean distance and the cosine distance; β represents the multiple of the reduction of the Euclidean distance, that is, the proportional coefficient.

[0037] Furthermore, the S4 specifically includes the following steps:

[0038] S51: Update the membership matrix according to the comprehensive distance, and its expression is:

[0039]

[0040] Where: ir Represents the Euclidean-angle cosine comprehensive distance from the kth cluster center of the data point to another cluster center; ρ ik It represents the Euclidean-cosine combined distance for combined indicators with similar values ​​and shapes;

[0041] S52: Re-determine the new cluster center based on the updated membership matrix, which is expressed as

[0042]

[0043] Where: v k (L+1) represents the new cluster center; (u ik (L)) represents the membership of the i-th sample to the initial k-th cluster set;

[0044] S53: Repeat the iterative execution of steps S3 to S4 until the iterative convergence condition is reached, and the final cluster center of the cluster set is obtained;

[0045] The iterative convergence condition is

[0046] ||V t -V t+1 ||≤ε or ||U t -U t+1 ||≤ε

[0047] Where: V t represents the cluster center of t iterations; V t+1 represents the cluster center of t+1 iteration; U t represents the membership matrix of t iterations; U t+1 represents the membership matrix after t+1 iterations; t represents the number of iterations; ε represents the preset error threshold;

[0048] And confirm the clustering results of transformer fault types based on the final cluster center.

[0049] Furthermore, obtaining the optimal prediction model in S7 specifically includes:

[0050] S71: Input the data training set into the LSTM model for model training to obtain a trained prediction model;

[0051] S72: Verifying the trained prediction model through the data test set;

[0052] That is, judging whether the output of the trained prediction model meets the preset accuracy threshold;

[0053] If the output of the trained prediction model meets the preset accuracy threshold, it is confirmed that the trained prediction model is the optimal prediction model;

[0054] Otherwise, based on the back propagation algorithm, the weight parameters of the trained prediction model are adaptively adjusted, and step S71 is repeated until it is confirmed that the weight parameters of the trained prediction model that meet the preset accuracy threshold are the optimal weight parameters and the prediction model is reconstructed to obtain the optimal prediction model.

[0055] Furthermore, in S72, mean absolute error and root mean square error are used as evaluation functions for judging the output accuracy of the trained prediction model, and their expressions are:

[0056]

[0057]

[0058] Where: N s represents the number of samples to be predicted in the data test set; p(n) and p′(n) represent the actual value and predicted value of the fault type classification respectively; δ MAE With δ RMSE Respectively represent the mean absolute error and root mean square error used for model validation.

[0059] Beneficial effects: The present invention provides a transformer oil chromatographic analysis method based on an improved fuzzy clustering algorithm and an LSTM algorithm. Different from the traditional hard partitioning clustering method, the improved fuzzy clustering algorithm can more accurately describe the intermediary relationship of data in each category by introducing fuzziness, thereby overcoming the defect that the hard partitioning method cannot truly reflect the actual relationship between the object and the cluster set. By continuously updating the cluster center and the membership of each sample to each cluster, and confirming the clustering result of the transformer fault type according to the final cluster center, an optimized objective function is obtained to realize flexible classification of data; the LSTM model is used as a prediction model for transformer fault analysis; and the prediction model is trained and verified based on the data training set and the data verification set to obtain the optimal prediction model. By adopting the long short-term memory network LSTM specially used for the prediction of time series data, the long-term dependency problem is effectively solved through the cell state and the "gate" mechanism, and the gradient disappearance or explosion is avoided. That is, by combining the improved fuzzy clustering algorithm with the LSTM algorithm, the transformer fault classification result can be made more accurate and the prediction model has higher accuracy. Thus, an accurate and efficient prediction and evaluation of the transformer operating status is effectively realized. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0061] Figure 1 The present invention is a flow chart of a transformer oil chromatographic analysis method based on an improved fuzzy clustering algorithm and an LSTM algorithm;

[0062] Figure 2 Schematic diagram of the structure of the LSTM model in this embodiment;

[0063] Figure 3 This is the core block diagram of the transformer oil chromatographic analysis method based on the improved fuzzy clustering algorithm and LSTM algorithm in this embodiment. DETAILED DESCRIPTION

[0064] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are 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.

[0065] A transformer oil chromatography analysis method based on improved fuzzy clustering algorithm and LSTM algorithm, wherein transformer oil chromatography analysis refers to the analysis of dissolved gas components in transformer oil by chromatography technology. This analysis is mainly used to monitor the internal health status of the transformer and assess whether the transformer has faults or potential fault risks. Oil chromatography analysis helps identify problems such as discharge, overheating, and insulation aging inside the transformer by detecting the gas components in the transformer oil. By analyzing the type and concentration of dissolved gases in the oil and judging in accordance with international standards (such as IEC 60599), the operating status and fault type of the transformer can be determined, such as Figure 1 and Figure 3 As shown, the specific steps include:

[0066] S1: pre-processing the acquired fault gas component data dissolved in transformer oil to obtain sample data;

[0067] The pretreatment is specifically:

[0068] The abnormal distribution data of the fault gas component data inside the transformer are obtained by using box plots and scatter plots; the blank items in the abnormal distribution data are filled with linear interpolation to obtain optimized data;

[0069] The z-score method is used to calculate and obtain the z-score value of each data point in the optimized data, and the data whose z-score value does not meet the preset data point threshold are removed as outliers. The proportion of various fault gases including hydrogen, methane, ethane, acetylene, ethylene, etc. is obtained to achieve data standardization, so as to obtain the final sample data;

[0070] S2: Randomly divide the sample data into several cluster sets;

[0071] That is, the sample data is defined as an n-tuple data set X = {x1, x2, ..., x i ,…,x n}, and randomly divide X into c clusters with 2≤c≤n; x i represents the concentration data of the transformer fault gas component obtained by the i-th gas detection, is an s-dimensional vector, and s in this embodiment represents the type of gas during gas detection. In this embodiment, 7 types of gases are considered, so it is a 7-dimensional vector;

[0072] That is x i ={x i1 ,x i2 ,…,x ij ,…,x is}, i=1,2,…,n,j=1,2,…,s

[0073] j represents the type of transformer fault gas components, including at least hydrogen, methane, ethane, ethylene, acetylene, carbon monoxide and carbon dioxide;

[0074] S3: Obtain the initial membership matrix of the cluster set, and improve the fuzzy clustering algorithm by introducing fuzziness to confirm the initial cluster center of the cluster set, which specifically includes the following steps:

[0075] S31: Obtain the initial membership matrix U of the cluster set = {u i1 ,u i2 ,…,u ik ,…,u ic};

[0076] That is, assuming that X can be randomly divided into c clusters (2≤c≤n), based on the improved fuzzy clustering algorithm, we first define the initial membership matrix U = {u i1 ,u i2 ,…,u ik ,…,u ic};

[0077] S32: By introducing fuzziness to improve the fuzzy clustering algorithm, the initial cluster center V of the cluster set is obtained according to the initial membership matrix, and its expression is:

[0078] V={v1,v2,…,v k ,v j …,v c}

[0079]

[0080] Where: m is the fuzzy weighting parameter, which represents the fuzziness of the initial membership matrix U. Usually, the value of m is 2; u ik represents the membership of the i-th sample to the k-th cluster; v j Represents the initial cluster center of the jth cluster set.

[0081] S4: Obtain the Euclidean distance and cosine distance between the initial cluster center of each cluster set and the sample data, and obtain the comprehensive distance based on the Euclidean distance and cosine distance;

[0082] Since the traditional Euclidean distance has limitations in evaluating the similarity of time series morphological changes, in contrast, the angle cosine distance is not affected by the value size and more accurately reflects the shape and direction of the curve, so it is often used to measure the similarity of time series; therefore, the Euclidean distance and the angle cosine distance are combined to construct a comprehensive distance as a classification standard (the comprehensive distance is used as a criterion for judging which category of the cluster center a data point belongs to, that is, dividing the data points into different clusters). The evaluation methods of the two are consistent, and the smaller the value, the higher the similarity;

[0083] The specific steps include:

[0084] S41: Get the Euclidean distance and cosine distance between the initial cluster center of each cluster and the sample data, and its expression is:

[0085]

[0086]

[0087] Where: It represents the cosine distance of the angle between the i-th sample and the k-th cluster center, and its value range is [0,2]. The smaller the value, the larger the cosine distance. ik represents the Euclidean distance between the initial cluster center and the sample data; s represents the number of gas types during gas detection; v k represents the initial cluster center of the kth cluster; v kj Represents the cosine distance between the initial cluster center of the kth cluster and the jth sample;

[0088] S42: Get the comprehensive distance based on Euclidean distance and cosine distance, its expression is

[0089]

[0090]

[0091] Where: ik represents the Euclidean-angle cosine comprehensive distance for comprehensive indicators with similar values ​​and shapes; σ and 1-σ represent the weights of the Euclidean distance and the angle cosine distance; β represents the multiple of the Euclidean distance reduction, that is, the proportionality coefficient;

[0092] S5: updating the membership matrix according to the comprehensive distance, and iteratively executing steps S3 to S4 repeatedly until the iterative convergence condition is reached, and obtaining the final cluster center of the cluster set, and confirming the clustering result of the transformer fault type according to the final cluster center;

[0093] The specific steps include:

[0094] S51: Update the membership matrix according to the comprehensive distance, and its expression is:

[0095]

[0096] Where: ir Represents the Euclidean-angle cosine comprehensive distance from the kth cluster center of the data point to another cluster center; ρ ikRepresents the Euclidean-angle cosine comprehensive distance for comprehensive indicators with similar values ​​and shapes; in this embodiment, the clustering result is obtained by updating the membership matrix and iterating; after determining the initial cluster center and the membership matrix, the comprehensive distance is calculated to update the membership of each data point to each cluster:

[0097] S52: Re-determine the new cluster center based on the updated membership matrix, which is expressed as

[0098]

[0099] Where: v k (L+1) represents the new cluster center; (u ik (L)) represents the membership of the i-th sample to the initial k-th cluster set;

[0100] S53: Repeat the iterative execution of steps S3 to S4 until the iterative convergence condition is reached, and the final cluster center of the cluster set is obtained;

[0101] The iterative convergence condition is

[0102] ||V t -V t+1 ||≤ε or ||U t -U t+1 ||≤ε

[0103] Where: V t represents the cluster center of t iterations; V t+1 represents the cluster center of t+1 iteration; U t represents the membership matrix of t iterations; U t+1 represents the membership matrix after t+1 iterations; t represents the number of iterations; ε represents the preset error threshold, which is 0.000001; at the same time, the change of V or U is very limited, and after the condition is met, it means that the iteration is over, and the clustering result of the transformer fault type is obtained;

[0104] That is, the clustering results of transformer fault types are confirmed according to the final cluster center;

[0105] S6: Dividing the clustering result into a data training set and a data test set according to a preset ratio;

[0106] In this embodiment, the clustering results classified by the improved fuzzy clustering algorithm are segmented, and 80% of the data of each type of fault is taken as the data training set for the training of the LSTM model; 20% of the data is taken as the data test set for the accuracy analysis of the classification results of the model later; the data obtained by the fuzzy clustering algorithm is segmented, the training set is used to train the model, the test set is used for accuracy test, and the accuracy is calculated according to the test results;

[0107] S7: Build an LSTM model and use it as a prediction model for transformer fault analysis;

[0108] The prediction model is trained and verified based on the data training set and the data verification set to obtain the optimal prediction model, and then the fault gas components dissolved in the transformer oil are predicted through the optimal prediction model, and the oil chromatography analysis of the transformer is realized according to the prediction results;

[0109] The obtaining of the optimal prediction model specifically includes

[0110] S71: Input the data training set into the LSTM model for model training to obtain a trained prediction model;

[0111] like Figure 2 As shown, the LSTM model training and optimization in this embodiment; (the number of LSTM network layers is selected according to the complexity and time length of the preset sensor data collection. For simple short-term data, a single-layer LSTM can effectively extract features, and for complex long-term data, multiple layers of LSTM are required. In the initial training, the number of layers of the LSTM network can be started from 2 layers, the number of units can be started from 64, the input dimension is (number of samples, time step, number of gas types), the output dimension is the number of fault types, the initial learning rate is set to 0.0001, the oil and gas content is used as input, and the classification result is used as output.) According to the type of oil and gas generated by different transformer faults, the number of layers of the LSTM network, the number of units in each layer, the input and output dimensions are determined, and the hyperparameters are set, and the oil and gas content is used as input, and the classification result is used as output. By testing the accuracy of the test set, and calculating the accuracy based on the test results, the number of LSTM network layers and units, and the learning rate are gradually increased and adjusted according to the accuracy to conduct a new model training, and continue to optimize the prediction model. The mechanism by which the LSTM neural network controls the information transmission path is a gating mechanism, which includes input gate, output gate and forget gate. t It controls how much information needs to be retained at the current moment; the output gate o t is the internal state c that controls the current moment t Need to send external state h t How much information to output; forget gate f t Control the internal state c at the previous moment t-1 How much information needs to be forgotten? The expression of the established LSTM model is

[0112] i t =σ(W i ·[h t-1 ,x t ]+b i )

[0113] f t =σ(W f ·[h t-1 ,x t ]+b f )

[0114] i t =σ(W o ·[h t-1 ,x t ]+b o )

[0115] Where: σ represents the sigmoid function; x t Indicates the input at the current moment; h t-1 represents the external state at the previous moment; W represents the weight coefficient matrix; b represents the bias term;

[0116] The operation process of the LSTM model in this embodiment is as follows: First, use the external state h at the previous moment t-1 and the current input x t , calculate three gates and candidate states The second step is to combine the forget gate f t and input gate t For memory unit c t Update the information; the third step is to combine the output gate o t , passing the internal state information to the external state h t ; In the LSTM model constructed in this embodiment, x t h is the transformer gas concentration at the current moment and the fault type code to which it belongs; t-1 The memory of gas concentration and fault type at the previous moment helps the network make more accurate predictions; t Control how much previous state information is forgotten, helping the model decide which historical gas concentration information is irrelevant to the current prediction and should be discarded; i t Controls the influence of current gas concentration information on fault type judgment, and determines whether the current gas concentration and fault type coding input information are important for future judgment. t Controls whether the model passes the "gas concentration" and "fault type" information of the current moment to the hidden state of the next moment, c t Record the relationship between gas concentration and fault type;

[0117] S72: Verifying the trained prediction model through the data test set;

[0118] That is, judging whether the output of the trained prediction model meets the preset accuracy threshold;

[0119] If the output of the trained prediction model meets the preset accuracy threshold, it is confirmed that the trained prediction model is the optimal prediction model;

[0120] Otherwise, based on the back propagation algorithm, the weight parameters of the trained prediction model are adaptively adjusted, and step S71 is repeated until it is confirmed that the weight parameters of the trained prediction model that meet the preset accuracy threshold are the optimal weight parameters and the prediction model is reconstructed to obtain the optimal prediction model.

[0121] In a specific embodiment, the mean absolute error and the root mean square error are used as evaluation functions for judging the output accuracy of the trained prediction model in S72, and the expression is:

[0122]

[0123]

[0124] Where: N s represents the number of samples to be predicted in the data test set; p(n) and p′(n) represent the actual value and predicted value of the fault type classification respectively; δ MAE With δ RMSE Respectively represent the mean absolute error and root mean square error used for model validation.

[0125] This embodiment is different from the traditional hard partition clustering method. The fuzzy clustering algorithm can more accurately describe the intermediary relationship of data in each category by introducing fuzziness to improve the fuzzy clustering algorithm, thereby overcoming the defect that the hard partition method cannot truly reflect the actual relationship between the object and the cluster. This embodiment continuously updates the cluster center and the membership of each sample to each cluster, and finally optimizes the objective function to achieve flexible classification of data. On the other hand, LSTM (Long Short-Term Memory Network) is an improved recursive neural network specifically used for the prediction of time series data. It effectively solves the long-term dependency problem through cell state and "gate" mechanism, avoids the situation of gradient disappearance or explosion, and can keep important information in long-term memory. This embodiment can avoid the limitations of a single method. The combination of the improved fuzzy clustering algorithm and the LSTM algorithm can make the classification results more accurate and the prediction model more accurate, so it can analyze the transformer oil chromatogram more accurately and evaluate the transformer operating status accurately and efficiently.

[0126] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A transformer oil chromatographic analysis method based on improved fuzzy clustering algorithm and LSTM algorithm, characterized in that: The specific steps include: S1: pre-processing the acquired fault gas component data dissolved in transformer oil to obtain sample data; S2: Randomly divide the sample data into several cluster sets; S3: Obtain the initial membership matrix of the clustering clusters, and improve the fuzzy clustering algorithm by introducing fuzziness to confirm the initial clustering center of the clustering clusters; S4: Obtain the Euclidean distance and cosine distance between the initial cluster center of each cluster set and the sample data, and obtain the comprehensive distance based on the Euclidean distance and cosine distance; S5: updating the membership matrix according to the comprehensive distance, and repeatedly iteratively executing steps S3 to S4 until an iterative convergence condition is reached, and obtaining the final cluster center of the cluster set; And confirm the clustering results of transformer fault types according to the final cluster center; S6: Dividing the clustering result into a data training set and a data test set according to a preset ratio; S7: Build an LSTM model and use it as a prediction model for transformer fault analysis; The prediction model is trained and verified based on the data training set and the data verification set to obtain the optimal prediction model, and then the fault gas components dissolved in the transformer oil are predicted through the optimal prediction model, and the oil chromatography analysis of the transformer is realized according to the prediction results.

2. A transformer oil chromatographic analysis method based on an improved fuzzy clustering algorithm and LSTM algorithm according to claim 1, characterized in that: The preprocessing described in S1 is specifically as follows: The abnormal distribution data of the fault gas component data inside the transformer are obtained by using box plots and scatter plots; the blank items in the abnormal distribution data are filled with linear interpolation to obtain optimized data; The z-score method is used to calculate and obtain the z-score value of each data point in the optimized data, and the data whose z-score value does not meet the preset data point threshold are removed as outliers to obtain the final sample data.

3. A transformer oil chromatographic analysis method based on an improved fuzzy clustering algorithm and LSTM algorithm according to claim 2, characterized in that: The S3 specifically includes the following steps: S31: Obtain the initial membership matrix U of the cluster set = {u i1 ,u i2 ,…,u ik ,…,u ic }; The sample data is defined as an n-tuple data set X = {x1, x2, ..., x i ,…,x n }, and randomly divide X into c clusters with 2≤c≤n; x i represents the concentration data of transformer fault gas components obtained by the i-th gas detection, and x i ={x i1 ,x i2 ,…,x ij ,…,x is }, i = 1, 2, ..., n, j = 1, 2, ..., sj represents the type of transformer fault gas components, including at least hydrogen, methane, ethane, ethylene, acetylene, carbon monoxide and carbon dioxide; S32: By introducing fuzziness to improve the fuzzy clustering algorithm, the initial cluster center V of the cluster set is obtained according to the initial membership matrix, and its expression is: V={v1,v2,…,v k ,v j …,v c } Where: m is the fuzzy weighting parameter, which represents the fuzziness of the initial membership matrix U; u ik represents the membership of the i-th sample to the k-th cluster; v j Represents the initial cluster center of the jth cluster set.

4. A transformer oil chromatographic analysis method based on an improved fuzzy clustering algorithm and LSTM algorithm according to claim 3, characterized in that: The S4 specifically comprises the following steps: S41: Get the Euclidean distance and cosine distance between the initial cluster center of each cluster and the sample data, and its expression is: Where: represents the cosine distance of the angle between the i-th sample and the k-th cluster center, and its value range is [0,2]; d ik represents the Euclidean distance between the initial cluster center and the sample data; s represents the number of gas types during gas detection; v k represents the initial cluster center of the kth cluster; v kj Represents the cosine distance between the initial cluster center of the kth cluster and the jth sample; S42: Get the comprehensive distance based on Euclidean distance and cosine distance, its expression is Where: ik It represents the Euclidean-cosine distance for comprehensive indicators with similar values ​​and shapes; σ and 1-σ represent the weights of the Euclidean distance and the cosine distance; β represents the multiple of the reduction of the Euclidean distance, that is, the proportional coefficient.

5. A transformer oil chromatographic analysis method based on improved fuzzy clustering algorithm and LSTM algorithm according to claim 4, characterized in that: The S4 specifically comprises the following steps: S51: Update the membership matrix according to the comprehensive distance, and its expression is: Where: ir Represents the Euclidean-angle cosine comprehensive distance from the kth cluster center of the data point to another cluster center; ρ ik It represents the Euclidean-cosine combined distance for combined indicators with similar values ​​and shapes; S52: Re-determine the new cluster center based on the updated membership matrix, which is expressed as Where: v k (L+1) represents the new cluster center; (u ik (L)) represents the membership of the i-th sample to the initial k-th cluster set; S53: Repeat the iterative execution of steps S3 to S4 until the iterative convergence condition is reached, and the final cluster center of the cluster set is obtained; The iterative convergence condition is ||V t -V t+1 || ≤ ε or ||U t -U t+1 || ≤ ε Where: V t represents the cluster center of t iterations; V t+1 represents the cluster center of t+1 iteration; U t represents the membership matrix of t iterations; U t+1 represents the membership matrix after t+1 iterations; t represents the number of iterations; ε represents the preset error threshold; And confirm the clustering results of transformer fault types based on the final cluster center.

6. A transformer oil chromatographic analysis method based on improved fuzzy clustering algorithm and LSTM algorithm according to claim 4, characterized in that: The obtaining of the optimal prediction model in S7 specifically includes: S71: Input the data training set into the LSTM model for model training to obtain a trained prediction model; S72: Verifying the trained prediction model through the data test set; That is, determine whether the output of the trained prediction model meets the preset accuracy threshold; If the output of the trained prediction model meets the preset accuracy threshold, then the trained prediction model is confirmed to be the optimal prediction model; Otherwise, based on the back propagation algorithm, the weight parameters of the trained prediction model are adaptively adjusted, and step S71 is repeated until it is confirmed that the weight parameters of the trained prediction model that meet the preset accuracy threshold are the optimal weight parameters and the prediction model is reconstructed to obtain the optimal prediction model.

7. A transformer oil chromatographic analysis method based on improved fuzzy clustering algorithm and LSTM algorithm according to claim 6, characterized in that: In S72, mean absolute error and root mean square error are used as evaluation functions to judge the output accuracy of the trained prediction model, and their expressions are: Where: N s represents the number of samples to be predicted in the data test set; p(n) and p′(n) represent the actual value and predicted value of the fault type classification respectively; δ MAE With δ RMSE Respectively represent the mean absolute error and root mean square error used for model validation.

Images