Oil-immersed transformer state identification method based on NRBO-Transformer-BiLSTM

Abstract

An oil-immersed transformer state identification method based on NRBO-Transform-BiLSTM comprises the following steps: acquiring an original data set of an oil-immersed transformer, selecting winding oil temperature, ambient temperature, load current and operating voltage as characteristic parameters, and interpolating low-frequency characteristic data such as ambient temperature to a time step length of high-frequency characteristic data; and classifying the obtained characteristic parameter data as sample data through a K center point clustering algorithm, extracting each category of characteristic center based on the Euclidean distance, and analyzing the state type of the oil-immersed transformer represented by the characteristic center. The method comprises the steps that data are input into a Transform-BiLSTM neural network optimized through an NRBO algorithm, model parameters are adjusted and optimized to improve model performance, an optimized model is used for training the data, the model performance is tested, and the trained model is used for recognizing the operation state of the oil-immersed transformer. According to the invention, the accuracy and real-time performance of state identification of the oil-immersed transformer can be improved.

Description

Technical Field

[0001] The present invention relates to an oil-immersed transformer state recognition method based on NRBO-Transformer-BiLSTM, and belongs to the technical field of oil-immersed transformer state monitoring. Background Art

[0002] In modern power grid systems, oil-immersed transformers are key equipment, and their operating stability and reliability are directly related to the overall safety and service quality of the power system. As the scale of the power grid continues to expand, the demand for electricity in society continues to grow, which puts higher requirements on the status monitoring and fault warning of oil-immersed transformers. Since oil-immersed transformers may be affected by various factors such as load, ambient temperature, and voltage fluctuations during long-term operation, they are prone to problems such as winding temperature rise and insulation aging, which in turn lead to equipment failure. Therefore, timely and accurate acquisition of oil-immersed transformer status information is of great significance to ensure the safe operation of equipment, reduce fault downtime, and improve operation and maintenance efficiency.

[0003] With the development of computer technology, algorithms for optimizing parameter combinations, such as genetic algorithms and particle swarm optimization, have been widely used in oil-immersed transformer data analysis, especially in terms of data missing patching and outlier removal, which effectively improves the integrity of data. In recent years, machine learning and deep learning technologies have gradually been applied to power systems, using big data to mine potential patterns and make more accurate predictions of the state changes of oil-immersed transformers. In the state identification of oil-immersed transformers, especially in the fields of real-time monitoring and fault warning, the research focus has gradually shifted to combining deep neural networks with optimization algorithms to improve the model's processing efficiency for nonlinear complex data. However, existing methods still face many challenges, such as low data processing efficiency, weak model generalization ability, and parameter adjustment relying on experience, which limits their widespread promotion in practical applications. Summary of the invention

[0004] In view of the problems existing in the above-mentioned prior art, the present invention provides a method for oil-immersed transformer state identification based on NRBO-Transformer-BiLSTM, which can improve the accuracy and real-time performance of oil-immersed transformer state identification, realize automatic selection of parameter combinations, enhance the generalization ability of the model, and provide guarantee for the safe and stable operation of the power system.

[0005] In order to achieve the above object, the present invention provides a method for identifying the state of an oil-immersed transformer based on NRBO-Transformer-BiLSTM, comprising the following steps:

[0006] S1. Obtain the original data set of the oil-immersed transformer, including active value, reactive value, load factor, winding oil temperature, ambient temperature, load current and operating voltage;

[0007] S2. Select winding oil temperature Tw, ambient temperature Ta, load current I, and operating voltage U as characteristic parameters; use a time series interpolation method for ambient temperature data to interpolate low-frequency characteristic data to the same time step as high-frequency characteristic data of other characteristic parameters, and align all characteristic parameter data to a unified time axis through a time synchronization alignment method;

[0008] S3, classifying the characteristic parameter data obtained in S2 as sample data through the K-center point clustering algorithm, and obtaining sample clusters and sample centers of each state type;

[0009] S4. Add a category label to each sample according to the clustering result to obtain a sample set for oil-immersed transformer state recognition, and divide the sample set into a training set and a test set;

[0010] S5, normalizing the feature samples so that the feature data are distributed in the interval [0, 1];

[0011] S6. Set the parameter range, population size, and number of iterations of the NRBO optimization algorithm, use the NRBO algorithm to obtain the optimal learning rate, L2 regularization coefficient, and number of BiLSTM hidden layer nodes, and find the parameter combination that makes the model perform best on the data set;

[0012] S7. Build a neural network including Transformer and BiLSTM layers, combine the optimal parameters to form the final model structure, and provide a network framework for subsequent training;

[0013] S8. Use the optimized model structure to train the data, output the training set and the test set, test the model performance, and output the accuracy of the model prediction and its various evaluation indicators;

[0014] S9. Use the trained model to identify the operating status of the oil-immersed transformer.

[0015] Furthermore, the original data set in S1 is obtained from the substation site, wherein the sampling interval of active value, reactive value, load factor, winding oil temperature, load current, and operating voltage is 5 minutes, and the sampling interval of ambient temperature is 24 hours.

[0016] Furthermore, the characteristic parameters in S2 refer to data after time series interpolation and synchronization alignment processing.

[0017] Furthermore, the specific process of S3 is as follows:

[0018] S3.1. Classify the sample data using the K center point clustering algorithm, where the K value is pre-set to 3, initialize K cluster center points, and use the Euclidean distance formula to calculate the distance from each sample to all cluster centers. The formula is as follows:

[0019]

[0020] Among them, x i and c k are the feature vectors of two samples respectively, n is the feature dimension, which are winding oil temperature Tw, ambient temperature Ta, load current I, and operating voltage U, a total of 4 dimensions; x i,m Represents sample x i The value in the mth dimension, c k,m Represents sample c k The value in the mth dimension;

[0021] S3.2. According to the minimum distance principle, assign the sample to the nearest cluster center μ k The corresponding category is updated to the mean of all samples in the current category, and the above steps are repeated until the cluster center is stable. The method for updating the cluster center is as follows: for all samples in the kth category, calculate their mean in each dimension to obtain the new cluster center μ′ k , the formula is as follows:

[0022] Among them, G k is the set of samples assigned to the kth category, N k is the number of samples in this category;

[0023] The basis for judging the stability of cluster centers is: all cluster centers μ′ k and the previous round of clustering center point μ k The changes are all lower than the preset threshold ε, and the formula is as follows:

[0024] ||μ' k -μ k ||<∈, for all k=1,2,…,K;

[0025] Among them, ||…|| represents the Euclidean distance;

[0026] If the above stability judgment conditions are met, the iteration is stopped; otherwise, the iteration is continued until the conditions are met.

[0027] Furthermore, the process of S4 is:

[0028] S4.1. According to the clustering result of S3, the category k of the sample is converted into a specific state type label, specifically: k=1 represents a normal load operation state, k=2 represents a high load operation state, and k=3 represents a light load or low load operation state;

[0029] S4.2. Generate an oil-immersed transformer state feature sample set through label data, and randomly divide the sample set into a training set and a test set in a ratio of 7:3.

[0030] Furthermore, the specific process of S5 is: normalizing the feature samples, and mapping the data to the [0,1] interval using the Min-Max normalization method, and the formula is as follows:

[0031]

[0032] Among them, x is the original eigenvalue, x min is the minimum value of the feature, x max is the maximum value of the feature, and x′ is the normalized feature value, ranging from [0,1].

[0033] Furthermore, the specific process of finding the parameter combination that makes the model perform best on the data set in S6 is:

[0034] S6.1. Define the search range: Set the initial search range for each parameter, specifically: learning rate l r ∈[0.0001,0.1], L2 regularization coefficient The number of BiLSTM hidden layer nodes h∈[32,256];

[0035] S6.2. Initialize the population: Randomly generate a set of parameter combinations as the initial population. Each set of parameters includes {l r ,λ L2 ,h};

[0036] S6.3, fitness calculation: find the parameter combination that minimizes the itness value according to the following formula:

[0037]

[0038] Where N is the number of samples used to calculate the classification loss, is the classification loss function, is the predicted value, y i is the true value; ||θ|| 2 represents the L2 regularization term of the model weight; λ L2 is the regularization coefficient;

[0039] S6.4, parameter update: sort the parameter combinations according to the Fitness value and retain the optimal solution; use random perturbations to explore parameters for the next generation population and narrow the parameter search range;

[0040] S6.5, Iterative optimization: Continuously repeat the fitness calculation and parameter update process until the preset number of iterations is reached or the Fitness value converges;

[0041] If the change in Fitness value of consecutive generations is lower than the preset threshold value γ = 0.001, it is considered that the parameter combination has reached the optimal value; if there is no significant change within the specified maximum number of iterations, the parameter combination with the smallest current Fitness value is selected. r ,λ L2 ,h} as the optimal solution.

[0042] Furthermore, the specific process of S7 is as follows:

[0043] S7.1, Transformer layer extracts the correlation between features through multi-head attention mechanism:

[0044]

[0045] Where Q, H, and V are query, key, and value matrices respectively, and d h is the dimension of the key vector;

[0046] S7.2, BiLSTM layer captures time series features through a bidirectional recurrent network:

[0047]

[0048] in, is the hidden state of the forward propagation, is the hidden state of back propagation, x t represents the input feature vector at time step t, W f , W b Represents the weight matrix from input feature to hidden state in forward and backward propagation, U f , U b Denotes the weight matrix from hidden state to hidden state in forward and backward propagation, b f and b b Represent the corresponding bias vectors in forward and backward propagation respectively, σ represents the activation function, and the tanh function is used to increase the nonlinear expression ability of the model;

[0049] S7.3, combine the parameters in S6 {l r ,λ L2,h} is used to train the final NRBO-Transformer-BiLSTM model. The process is as follows: initialize the NRBO-Transformer-BiLSTM model, input the training set into the model, use the optimization algorithm NRBO to train in batches, and use the validation set to evaluate the model performance. Monitor whether the training process converges according to the Fitness value; the training stops when the Fitness value converges or reaches the maximum number of iterations. Finally, save the trained model and use the test set to verify its performance before entering the next stage of model training and application. The final parameter combination is determined It can make the model have the highest classification accuracy and the smallest error on the validation set, ensure the optimal model structure, output the classification results with the fully connected layer, and provide a network framework for subsequent training.

[0050] Furthermore, the specific process of S8 is: to perform performance verification on the test set, output the model's classification accuracy CA, sensitivity SE, specificity SP and FM value evaluation indicators, comprehensively analyze the performance through the confusion matrix, and calculate the classification accuracy and classification accuracy percentage; wherein, the classification accuracy represents the percentage of correctly classified samples among all samples in the total samples, reflecting the overall prediction ability of the model, and the calculation formula is the sum of the diagonal elements of the confusion matrix divided by the total number of samples; the classification accuracy percentage represents whether the model's prediction ability for a specific category is balanced, and is used to determine whether the model is biased towards a certain category by observing the confusion matrix data of each category.

[0051] Furthermore, the specific process of S9 is: by inputting the normalized feature data into the trained NRBO-Transformer-BiLSTM model, using the classification output of the model to determine the operating status category of the transformer, including normal operation, high load operation, and light load operation, and combining the classification results to calculate the evaluation index to realize the classification and identification of the oil-immersed transformer state.

[0052] The present invention obtains the original data set of the oil-immersed transformer, selects the winding oil temperature, ambient temperature, load current, and operating voltage as characteristic parameters in the original data set, uses the time series interpolation method for the ambient temperature data, interpolates the low-frequency characteristic data to the same time step as the high-frequency characteristic data of other characteristic parameters, and aligns all characteristic parameter data to a unified time axis through a time synchronization alignment method, classifies the obtained characteristic parameter data as sample data through a K-center point clustering algorithm, extracts the characteristic center of each category based on the Euclidean distance, and analyzes the state type of the oil-immersed transformer represented by it. On this basis, the data is input into the Transformer-BiLSTM neural network optimized by the NRBO algorithm, and the learning rate, L2 regularization coefficient, and number of hidden layer nodes of the model are tuned to improve the model performance, and the optimized model structure is used to train the data, output the training set and the test set, and test the model performance, output the accuracy of the model prediction and its various evaluation indicators, and the trained model is used for the identification of the operating state of the oil-immersed transformer. Through efficient data processing and optimized neural network structure, the present invention realizes accurate classification of different operating states of oil-immersed transformers, improves the accuracy and real-time performance of oil-immersed transformer state identification, realizes automatic selection of parameter combinations, enhances model generalization capability, provides important basis for operating state analysis and decision-making of oil-immersed transformers, and ensures safe and stable operation of power systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 is a flow chart of the method of the present invention;

[0054] Figure 2 This is a graph showing the change in Fitness value with the number of iterations;

[0055] Figure 3 This is the architecture diagram of the neural network model based on Transformer-BiLSTM;

[0056] Figure 4 This is a performance evaluation result diagram of the optimized model. DETAILED DESCRIPTION

[0057] The present invention will be further described below in conjunction with the accompanying drawings.

[0058] like Figure 1 As shown, a method for identifying the state of an oil-immersed transformer based on NRBO-Transformer-BiLSTM comprises the following steps:

[0059] S1. Obtain the original data set of the oil-immersed transformer, including active value, reactive value, load factor, winding oil temperature, ambient temperature, load current and operating voltage;

[0060] S2. Select winding oil temperature Tw, ambient temperature Ta, load current I, and operating voltage U as characteristic parameters; use a time series interpolation method for ambient temperature data to interpolate low-frequency characteristic data to the same time step as high-frequency characteristic data of other characteristic parameters, and align all characteristic parameter data to a unified time axis through a time synchronization alignment method;

[0061] S3, classifying the characteristic parameter data obtained in S2 as sample data through the K-center point clustering algorithm, and obtaining sample clusters and sample centers of each state type;

[0062] S4. Add a category label to each sample according to the clustering result to obtain a sample set for oil-immersed transformer state recognition, and divide the sample set into a training set and a test set;

[0063] S5, normalizing the feature samples so that the feature data are distributed in the interval [0, 1];

[0064] S6. Set the parameter range, population size, and number of iterations of the NRBO optimization algorithm, use the NRBO algorithm to obtain the optimal learning rate, L2 regularization coefficient, and number of BiLSTM hidden layer nodes, and find the parameter combination that makes the model perform best on the data set;

[0065] S7. Build a neural network including Transformer and BiLSTM layers, combine the optimal parameters to form the final model structure, and provide a network framework for subsequent training;

[0066] S8. Use the optimized model structure to train the data, output the training set and the test set, test the model performance, and output the accuracy of the model prediction and its various evaluation indicators;

[0067] S9. Use the trained model to identify the operating status of the oil-immersed transformer.

[0068] This embodiment collects the original data set of a 525kV oil-immersed transformer in a power supply company substation in a certain city. The basic parameters of the transformer are shown in Table 1:

[0069] Table 1 Basic parameters of transformer

[0070]

[0071] The specific process of identifying its status is as follows:

[0072] (1) Obtain the original data set of the oil-immersed transformer at the substation site, including active value, reactive value, load factor, winding oil temperature, ambient temperature, load current, and operating voltage. The sampling interval of active value, reactive value, load factor, winding oil temperature, load current, and operating voltage is 5 minutes, and the sampling interval of ambient temperature is 24 hours;

[0073] (2) Select winding oil temperature, ambient temperature, load current, and operating voltage as characteristic parameters. Characteristic parameters refer to data that has been processed by time series interpolation and synchronization alignment. In this embodiment, the characteristic parameters are winding oil temperature Tw, ambient temperature Ta, load current I, and operating voltage U. In case the sampling frequency of ambient temperature data is lower than that of other features, the time series interpolation method is used to convert the low-frequency characteristic data ambient temperature T a Interpolation to high frequency characteristic data winding oil temperature T w , load current I and operating voltage U, and align all feature data to a unified time axis through a time synchronization alignment method to obtain a feature data set with a consistent time dimension;

[0074] (3) The sample data is classified by the K-center point clustering algorithm to obtain the sample clusters and sample centers of each state type. The K value is pre-set to 3, K cluster center points are initialized, and the Euclidean distance formula is used to calculate the distance between each sample x i To all cluster centers c k The distance is as follows:

[0075]

[0076] x i and c k are the feature vectors of two samples respectively, n is the feature dimension (winding oil temperature, ambient temperature, load current, operating voltage, a total of 4 dimensions), x i,m Represents sample x i The value in the mth dimension;

[0077] According to the minimum distance principle, the samples are assigned to the nearest cluster center μ k , the corresponding category, update the cluster center to the mean of all samples in the current category, repeat the above steps until the cluster center is stable

[0078] The specific method of updating the cluster center point is: for all samples of the kth category, calculate its mean in each dimension and get the new cluster center point μ′ k , the formula is as follows:

[0079] Among them, G k is the set of samples assigned to the kth category, N kis the number of samples in this category, x i,m Represents sample x i The value in the mth dimension.

[0080] Repeat the above steps until the cluster center is stable. The basis for judging the stability of the cluster center is: all cluster center points μ′ k and the previous round of clustering center point μ k The changes are all lower than the preset threshold ∈=10 -4 ; Table 2 shows the cluster center samples of each operating condition after determining the number of clusters K = 3:

[0081] Table 2 Cluster center samples

[0082]

[0083] (4) According to the clustering result of step (3), the category k to which the sample belongs is converted into a specific state type label. k = 1 represents "normal load operation state", k = 2 represents "high load operation state", and k = 3 represents "light load or low load operation state". Generate the state feature sample set of the oil-immersed transformer through the label data, and randomly divide the sample set into a training set and a test set in a ratio of 7:3 to ensure the effect of model training and testing; Table 3 shows the generated state feature sample set:

[0084] Table 3 State feature sample set

[0085]

[0086] (5) Normalize the feature samples and use the Min-Max normalization method to map the data to the [0,1] interval. The formula is as follows:

[0087]

[0088] x is the original eigenvalue, x min is the minimum value of the feature, x max is the maximum value of the feature, x′ is the normalized feature value, and the range is [0,1];

[0089] (6) By defining the optimization objective function, the classification error of the model on the validation set is used as the fitness function, and the NRBO algorithm is used to optimize the learning rate l r , L2 regularization coefficient The number of nodes h in the BiLSTM hidden layer is optimized. The NRBO (Neural Random-Based Optimization) algorithm is an optimization algorithm based on random search, which is suitable for finding the optimal solution of multiple parameter combinations. The optimization process is as follows:

[0090] (a) Define the search scope:

[0091] Set the initial search range for each parameter: learning rate l r ∈[0.0001,0.1], L2 regularization coefficient The number of BiLSTM hidden layer nodes h∈[32,256];

[0092] (b) Initialize the population:

[0093] A set of parameter combinations is randomly generated as the initial population, each set of parameters includes {l r ,λ L2 ,h};

[0094] (c) Fitness calculation:

[0095] like Figure 2 As shown, find the parameter combination that minimizes the Fitness value according to the following formula:

[0096]

[0097] N is the number of samples used to calculate the classification loss, is the classification loss function (e.g. cross entropy loss), is the predicted value, y i is the true value; ||θ|| 2 represents the L2 regularization term of the model weight; λ L2 is the regularization coefficient;

[0098] (d) Parameter update:

[0099] Sort the parameter combinations by Fitness value and keep the best solution. Use random perturbations to explore parameters for the next generation population and narrow the parameter search range.

[0100] (e) Iterative Optimization:

[0101] The fitness calculation and parameter update process is repeated until the preset number of iterations is reached or the Fitness value converges.

[0102] If the change in Fitness value of consecutive generations is lower than the preset threshold value γ = 0.001, the parameter combination is considered to be optimal. If no significant change occurs within the specified maximum number of iterations of 100, the parameter combination with the smallest current Fitness value is selected. r ,λ L2 ,h} as the optimal solution;

[0103] The optimal solution of the parameter combination finally found is the learning rate lr = 0.001, the L2 regularization coefficient The number of BiLSTM hidden layer nodes h=128, and the minimum Fitness value is 0.025;

[0104] (7) Figure 3 As shown, build a neural network containing Transformer and BiLSTM layers:

[0105] The Transformer layer extracts the correlation between features through a multi-head attention mechanism:

[0106]

[0107] Where Q, H, and V are query, key, and value matrices respectively, and d h is the dimension of the key vector;

[0108] The BiLSTM layer captures time series features through a bidirectional recurrent network:

[0109]

[0110] is the hidden state of the forward propagation, is the hidden state of back propagation, x t represents the input feature vector at time step t, W f , W b Represents the weight matrix from input features to hidden states in forward and backward propagation, U f , U b represents the weight matrix from hidden state to hidden state in forward and backward propagation, b f and b b Represents the corresponding bias vector in forward and backward propagation, σ represents the activation function, and the tanh function is used to increase the nonlinear expression ability of the model;

[0111] The parameter combination in step (6) {l r ,λ L2 ,h} is used to train the final NRBO-Transformer-BiLSTM model. The training process is as follows: first initialize the NRBO-Transformer-BiLSTM model, input the training set into the model, use the optimization algorithm NRBO to train in batches, and use the validation set to evaluate the model performance, and monitor whether the training process converges according to the Fitness value. The training will stop when the Fitness value converges or reaches the maximum number of iterations (100 times). Finally, save the trained model and use the test set to verify its performance to enter the next step of model training and application;

[0112] (8) The optimized neural network model is used to train the training set and to verify the performance of the test set. The model’s classification accuracy CA, sensitivity SE, specificity SP, and FM value (harmonic mean of precision and recall) and other evaluation indicators are output, such as Figure 4 As shown in the figure, the classification accuracy (CA) reflects the correctness of the overall prediction of the model, indicating the proportion of correctly classified samples in all samples. The CA index in the figure is close to the vertex 1, indicating that the overall classification performance of the model is very high and can effectively distinguish different operating states of the transformer;

[0113] Sensitivity (SE) indicates the ability of the model to correctly identify the positive class (a certain state), that is, the proportion of samples that are correctly classified as a certain category among samples that are actually of a certain category. A high value of sensitivity indicates that the model is stable in identifying samples of various categories, especially in detecting key states without omissions.

[0114] Specificity (SP) indicates the ability of the model to correctly identify negative classes, that is, the proportion of samples that are correctly classified as non-classes among samples that do not belong to a certain class. A high SP value indicates that the model has a strong ability to exclude wrong classes, which helps to reduce false positives.

[0115] The FM value is the harmonic mean of precision and recall, which is used to comprehensively measure the performance of the model in classification. A high FM value indicates that the model has achieved a good balance between correct classification and reducing misclassification.

[0116] Youden index J is used to evaluate the overall discriminative ability of the model. Its calculation formula is, J = SE + SP-1. The J value close to the vertex indicates that the model performs well in balancing sensitivity and specificity;

[0117] The area under the ROC curve (AUC) represents the comprehensive performance of the model under all possible classification thresholds. The larger the value, the higher the discrimination ability of the model in the classification task. The AUC value in the figure is close to the vertex 1, indicating that the model has extremely high accuracy and stability in classifying different operating states.

[0118] The performance is comprehensively analyzed through the confusion matrix to ensure the performance of the model in identifying the state of the oil-immersed transformer. The confusion matrix analysis results and classification performance indicators are shown in Table 4:

[0119] Table 4 Confusion matrix analysis results and classification performance indicators

[0120]

[0121] The classification accuracy indicates the percentage of correctly classified samples in each category. As can be seen from Table 4, the classification accuracy of all categories in the training set is between 98.8% and 100.0%, indicating that the model fits the training data very well; the classification accuracy in the test set remains at 99.4% to 100.0%, showing that the model has excellent classification performance for new data;

[0122] The classification accuracy percentage reflects whether the model's prediction ability for a specific category is balanced. The classification accuracy percentages in both the training set and the test set are close to 100%, indicating that the model performs evenly across all categories, with no one category being significantly lower than the others.

[0123] Overall, the results of the training set and test set show that this method performs well in classification accuracy, classification precision and misclassification rate, and the number of misclassified samples is extremely low, especially in the test set, which shows the strong generalization ability of the model on unseen data. This performance advantage is due to the combination of Transformer and BiLSTM network structures optimized by the NRBO algorithm, as well as efficient data preprocessing methods.

Claims

1. A method for identifying the state of an oil-immersed transformer based on NRBO-Transformer-BiLSTM, characterized in that: The following steps are involved: S1. Obtain the original data set of the oil-immersed transformer, including active value, reactive value, load factor, winding oil temperature, ambient temperature, load current and operating voltage; S2. Select winding oil temperature Tw, ambient temperature Ta, load current I, and operating voltage U as characteristic parameters; use a time series interpolation method for ambient temperature data to interpolate low-frequency characteristic data to the same time step as high-frequency characteristic data of other characteristic parameters, and align all characteristic parameter data to a unified time axis through a time synchronization alignment method; S3, classifying the characteristic parameter data obtained in S2 as sample data through the K-center point clustering algorithm, and obtaining sample clusters and sample centers of each state type; S4. Add a category label to each sample according to the clustering result to obtain a sample set for oil-immersed transformer state recognition, and divide the sample set into a training set and a test set; S5, normalizing the feature samples so that the feature data are distributed in the interval [0, 1]; S6. Set the parameter range, population size, and number of iterations of the NRBO optimization algorithm, use the NRBO algorithm to obtain the optimal learning rate, L2 regularization coefficient, and number of BiLSTM hidden layer nodes, and find the parameter combination that makes the model perform best on the data set; S7. Build a neural network including Transformer and BiLSTM layers, combine the optimal parameters to form the final model structure, and provide a network framework for subsequent training; S8. Use the optimized model structure to train the data, output the training set and the test set, test the model performance, and output the accuracy of the model prediction and its various evaluation indicators; S9. Use the trained model to identify the operating status of the oil-immersed transformer.

2. The oil-immersed transformer state identification method based on NRBO-Transformer-BiLSTM according to claim 1 is characterized in that: The original data set in S1 is obtained from the substation site, wherein the sampling interval of active value, reactive value, load factor, winding oil temperature, load current, and operating voltage is 5 minutes, and the sampling interval of ambient temperature is 24 hours.

3. The oil-immersed transformer state identification method based on NRBO-Transformer-BiLSTM according to claim 1 or 2 is characterized in that: The characteristic parameters in S2 refer to the data after time series interpolation and synchronization alignment processing.

4. The oil-immersed transformer state identification method based on NRBO-Transformer-BiLSTM according to claim 3 is characterized in that: The specific process of S3 is as follows: S3.

1. Classify the sample data using the K center point clustering algorithm, where the K value is pre-set to 3, initialize K cluster center points, and use the Euclidean distance formula to calculate the distance from each sample to all cluster centers. The formula is as follows: Among them, x i and c k are the feature vectors of two samples respectively, n is the feature dimension, which are winding oil temperature Tw, ambient temperature Ta, load current I, and operating voltage U, a total of 4 dimensions; x i,m Represents sample x i The value in the mth dimension, c k,m Represents sample c k The value in the mth dimension; S3.

2. According to the minimum distance principle, assign the sample to the nearest cluster center μ k The corresponding category is updated to the mean of all samples in the current category, and the above steps are repeated until the cluster center is stable. The method for updating the cluster center is as follows: for all samples in the kth category, calculate their mean in each dimension to obtain the new cluster center μ′ k , the formula is as follows: Among them, G k is the set of samples assigned to the kth category, N k is the number of samples in this category; The basis for judging the stability of cluster centers is: all cluster centers μ′ k and the previous round of clustering center point μ k The changes are all lower than the preset threshold ∈, and the formula is as follows: ||μ’ k -μ k ||<ε, for all k = 1, 2, …, K; Among them, ||…|| represents the Euclidean distance; If the above stability judgment conditions are met, the iteration is stopped; otherwise, the iteration is continued until the conditions are met.

5. The oil-immersed transformer state identification method based on NRBO-Transformer-BiLSTM according to claim 4 is characterized in that: The process of S4 is as follows: S4.

1. According to the clustering result of S3, the category k of the sample is converted into a specific state type label, specifically: k=1 represents a normal load operation state, k=2 represents a high load operation state, and k=3 represents a light load or low load operation state; S4.

2. Generate an oil-immersed transformer state feature sample set through label data, and randomly divide the sample set into a training set and a test set in a ratio of 7:

3.

6. The oil-immersed transformer state identification method based on NRBO-Transformer-BiLSTM according to claim 5 is characterized in that: The specific process of S5 is: normalize the feature samples and map the data to the [0,1] interval using the Min-Max normalization method. The formula is as follows: Among them, x is the original eigenvalue, x min is the minimum value of the feature, x max is the maximum value of the feature, and x′ is the normalized feature value, ranging from [0,1].

7. The oil-immersed transformer state identification method based on NRBO-Transformer-BiLSTM according to claim 6 is characterized in that: The specific process of finding the parameter combination that makes the model perform best on the data set in S6 is: S6.

1. Define the search range: Set the initial search range for each parameter, specifically: learning rate l r ∈[0.0001,0.1], L2 regularization coefficient λ L2 ∈[0.0001,0.01], number of BiLSTM hidden layer nodes h∈[32,256]; S6.

2. Initialize the population: Randomly generate a set of parameter combinations as the initial population. Each set of parameters includes {l r ,λ L2 ,h}; S6.3, fitness calculation: find the parameter combination that minimizes the itness value according to the following formula: Where N is the number of samples used to calculate the classification loss, is the classification loss function, is the predicted value, y i is the true value; ||θ|| 2 represents the L2 regularization term of the model weight; λ L2 is the regularization coefficient; S6.4, parameter update: sort the parameter combinations according to the Fitness value and retain the optimal solution; use random perturbations to explore parameters for the next generation population and narrow the parameter search range; S6.5, Iterative optimization: Continuously repeat the fitness calculation and parameter update process until the preset number of iterations is reached or the Fitness value converges; If the change in Fitness value of consecutive generations is lower than the preset threshold value γ = 0.001, it is considered that the parameter combination has reached the optimal value; if there is no significant change within the specified maximum number of iterations, the parameter combination with the smallest current Fitness value is selected. r ,λ L2 ,h} as the optimal solution.

8. The oil-immersed transformer state identification method based on NRBO-Transformer-BiLSTM according to claim 7 is characterized in that: The specific process of S7 is as follows: S7.1, Transformer layer extracts the correlation between features through multi-head attention mechanism: Where Q, H, and V are query, key, and value matrices respectively, and d h is the dimension of the key vector; S7.2, BiLSTM layer captures time series features through a bidirectional recurrent network: in, is the hidden state of the forward propagation, is the hidden state of back propagation, x t represents the input feature vector at time step t, W f , W b Represents the weight matrix from input feature to hidden state in forward and backward propagation, U f , U b Denotes the weight matrix from hidden state to hidden state in forward and backward propagation, b f and b b Respectively represent the corresponding bias vectors in forward and backward propagation, σ represents the activation function; S7.3, combine the parameters in S6 {l r ,λ L2 ,h} is used to train the final NRBO-Transformer-BiLSTM model. The process is: initialize the NRBO-Transformer-BiLSTM model, input the training set into the model, use the optimization algorithm NRBO to train in batches, and use the validation set to evaluate the model performance. Monitor whether the training process converges according to the Fitness value; the training stops when the Fitness value converges or reaches the maximum number of iterations. Finally, save the trained model and use the test set to verify its performance to enter the next step of model training and application.

9. The oil-immersed transformer state identification method based on NRBO-Transformer-BiLSTM according to claim 8 is characterized in that: The specific process of S8 is: to perform performance verification on the test set, output the model's classification accuracy CA, sensitivity SE, specificity SP and FM value evaluation indicators, comprehensively analyze the performance through the confusion matrix, and calculate the classification accuracy and classification accuracy percentage; wherein, the classification accuracy represents the percentage of correctly classified samples among all samples in the total samples, and the calculation formula is the sum of the diagonal elements of the confusion matrix divided by the total number of samples; the classification accuracy percentage represents whether the model's prediction ability for a specific category is balanced, and is used to determine whether the model is biased towards a certain category by observing the confusion matrix data of each category.

10. The oil-immersed transformer state identification method based on NRBO-Transformer-BiLSTM according to claim 9 is characterized in that: The specific process of S9 is: by inputting the normalized feature data into the trained NRBO-Transformer-BiLSTM model, using the classification output of the model to determine the operating status category of the transformer, including normal operation, high load operation, and light load operation, and combining the classification results to calculate the evaluation index to realize the classification and identification of the oil-immersed transformer status.

Images