Power grid operation safety detection method based on artificial intelligence
By employing an AI-based modal decomposition and multi-model collaborative prediction method, the problem of low accuracy in predicting dissolved gas in transformers was solved, enabling intelligent assessment of transformer operating status and timely fault identification, thereby improving the safety and reliability of power grid operation.
Patent Information
- Application Number
- CN202511160703.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-19
- Publication Date
- 2025-12-09
AI Technical Summary
Existing technologies for dissolved gas prediction and state analysis suffer from low accuracy, high subjectivity, and high misjudgment rate, making it difficult to detect potential transformer faults in a timely manner and affecting the safe operation of the power grid.
An artificial intelligence-based approach is adopted to predict the historical sequence of dissolved gas in transformers by modal decomposition and reconstruction. An improved complete set empirical mode decomposition adaptive noise method (ICEEMDAN) and a bidirectional long short-term memory network model are used for prediction. Convolutional neural networks and probabilistic neural networks are combined for state recognition to generate a high-quality training sample set for model training.
It improves the accuracy of dissolved gas prediction, enabling timely and accurate identification of potential transformer faults, reducing the false alarm rate, and enhancing the safety and reliability of power grid operation.
Smart Images
Figure CN121094293A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the technical field of power grid safety, and specifically to an artificial intelligence-based method for detecting power grid operation safety. Background Technology
[0002] As a crucial infrastructure of modern society, the safe, stable, and reliable operation of the power system plays a vital role in ensuring national economic and social well-being. The core of power grid operation safety monitoring lies in the ability to identify potential hazards in advance and prevent major accidents. Among the key equipment in the power grid, power transformers are of high value and long lifespan, and their operating status directly affects the reliability and economy of the grid. Latent faults inside transformers (such as localized overheating, partial discharge, and insulation degradation) are major factors leading to power outages. Early faults are difficult to detect visually or through conventional testing methods, but they can cause characteristic changes in the content and proportion of various dissolved gases in the insulating oil. Therefore, accurate analysis and trend prediction of dissolved gases, combined with operational status assessment, has become a key link in ensuring transformer health and improving power grid safety.
[0003] However, existing technologies have significant limitations in dissolved gas prediction and condition analysis. On the one hand, traditional condition analysis relies heavily on empirical judgment or threshold settings, making it difficult to comprehensively reflect the actual operating conditions of transformers and resulting in strong subjectivity and a high rate of misjudgment. On the other hand, dissolved gas sequences exhibit nonlinear and non-stationary characteristics, while existing prediction methods typically employ holistic sequence analysis, which struggles to accurately capture their complex evolution patterns, leading to insufficient prediction accuracy. These problems make it difficult for existing technologies to reflect potential transformer fault risks in a timely and accurate manner, thereby affecting the safe operation of the power grid. Summary of the Invention
[0004] The purpose of this invention is to solve the problems mentioned in the background art, such as low accuracy of dissolved gas prediction, inaccurate state analysis, and difficulty in timely detection of potential faults, and to propose an artificial intelligence-based power grid operation safety detection method.
[0005] A first aspect of this invention provides an artificial intelligence-based method for detecting power grid operation safety, the method comprising:
[0006] Obtain historical monitoring data of various dissolved gases in transformers;
[0007] The historical sequence of the target dissolved gas is subjected to modal decomposition and reconstruction to obtain multiple sub-sequences; the target dissolved gas is any one of a variety of dissolved gases.
[0008] For any subsequence, call the corresponding time series prediction model to make a prediction and obtain the predicted value of the subsequence;
[0009] Adding the plurality of sub-sequence prediction values to obtain a prediction value of the target dissolved gas;
[0010] According to the prediction values of the plurality of dissolved gases, feature extraction is performed to obtain a first feature vector;
[0011] The first feature vector is input into a pre-trained state recognition model to obtain the operating state of the transformer; the state recognition model is an artificial intelligence model obtained by training.
[0012] Optionally, the plurality of sub-sequences includes a residual sequence and a plurality of recombination sequences;
[0013] The modal decomposition and reconstruction of the historical sequence of the target dissolved gas to obtain the plurality of sub-sequences includes:
[0014] The historical sequence of the target dissolved gas is subjected to modal decomposition by using an improved complete ensemble empirical mode decomposition adaptive noise method to obtain a plurality of modal components and a residual sequence;
[0015] The average frequencies of the modal components are calculated, and the modal components are divided into three groups of high-frequency components, medium-frequency components, and low-frequency components according to the average frequencies of the modal components;
[0016] The modal components in each group are summed to obtain three recombination sequences.
[0017] Optionally, the time series prediction model adopts a bidirectional long short-term memory network model; for each sub-sequence, a corresponding time series prediction model is trained separately.
[0018] Optionally, the feature extraction according to the prediction values of the plurality of dissolved gases to obtain the first feature vector includes:
[0019] According to the prediction values of the plurality of dissolved gases, a plurality of gas ratios are generated;
[0020] The prediction values of the plurality of dissolved gases and the plurality of gas ratios are combined to obtain a second feature vector;
[0021] The second feature vector is converted by using a pre-set weight matrix to obtain the first feature vector; the weight matrix is determined by a training sample set.
[0022] Optionally, the state recognition model is an integrated model including a convolutional neural network base model, a probabilistic neural network base model, and a fully connected layer; the operation process of the state recognition model includes:
[0023] The first feature vector is input into the convolutional neural network base model and the probabilistic neural network base model respectively to obtain a first probability vector and a second probability vector;
[0024] concatenate the first probability vector and the second probability vector to form a third feature vector;
[0025] input the third feature vector into a full connection layer to output a third probability vector;
[0026] determine the operation state of the transformer according to a category with the largest probability value in the third probability vector.
[0027] Optionally, the training process of the state recognition model comprises:
[0028] obtain a historical state data set of the transformer; the historical state data comprises dissolved gas data and corresponding operation state labels;
[0029] perform feature extraction on original dissolved gas data to obtain gas feature data; and establish a first data set according to the gas feature data;
[0030] generate a weight matrix according to the gas feature data;
[0031] perform dimension reduction processing on the gas feature data of the first data set according to the weight matrix to obtain factor score feature data, and establish a second data set according to the factor score feature data;
[0032] generate data of a rare fault state by using a conditional generative adversarial network to obtain a third data set; the rare fault state is determined by the number of samples in the first data set;
[0033] merge the third data set into the second data set to obtain a fourth data set;
[0034] filter the fourth data set to obtain a training sample set;
[0035] perform model training by using the training sample set to obtain a state detection model.
[0036] Optionally, the generating a weight matrix according to the gas feature data comprises:
[0037] perform standardization processing on the gas feature data to obtain a sample matrix;
[0038] generate a correlation matrix according to the sample matrix;
[0039] perform processing on the correlation matrix by using a principal component method to obtain a principal component transformation matrix;
[0040] calculate an initial factor loading matrix according to the principal component transformation matrix;
[0041] perform variance maximization rotation on the initial factor loading matrix to obtain a rotated factor loading matrix;
[0042] According to the rotated factor loading matrix and the correlation matrix, a weight matrix is calculated.
[0043] Optionally, the filtering of the fourth data set to obtain the training sample set comprises:
[0044] The repeated samples in the fourth data set are deleted to obtain a fifth data set.
[0045] Noise recognition is performed on the samples in the fifth data set, and noise data is deleted to obtain a sixth data set.
[0046] The sixth data set is subjected to sample quantity balancing processing to obtain the training sample set.
[0047] Optionally, the noise recognition performed on the samples in the fifth data set and the deletion of the noise data to obtain the sixth data set comprise:
[0048] Each sample in the fifth data set is classified by using a support vector machine, and sample data with classification errors is included in a weak feature data set.
[0049] According to the local neighborhood samples of a target candidate sample, a reference category of the target candidate sample is determined; the target candidate sample is any sample in the weak feature data set.
[0050] If the real label of the target candidate sample is inconsistent with the reference category of the target candidate sample, the target candidate sample is determined as a noise sample.
[0051] All noise samples are deleted from the fifth data set to obtain the sixth data set.
[0052] Optionally, the sample quantity balancing processing performed on the sixth data set to obtain the training sample set comprises:
[0053] According to the label categories, the sixth data set is divided into multiple category groups.
[0054] A feature distribution center of a target category group is calculated; the target category group is any one of the multiple category groups.
[0055] The Euclidean distance between each sample in the target category group and the feature distribution center of the target category group is calculated, and is recorded as the outlier score of the sample.
[0056] The samples in the target category group are sorted in descending order according to the outlier scores, and the first L samples are selected to be included in the training sample set; L is not greater than the minimum value of the sample numbers of the category groups.
[0057] The beneficial effects of the present application are:
[0058] By accurately capturing the dynamic characteristics of gas sequences through modal decomposition, and combining multi-model collaborative prediction to enhance the modeling ability of nonlinear trends and improve prediction accuracy, the system can learn fault state characteristics through pre-trained artificial intelligence models to achieve intelligent assessment of transformer operating status. This enables timely and accurate identification of potential faults, reduces the false judgment rate, improves the scientific nature of transformer maintenance decisions, and provides reliable safety assurance for power grid operation. Attached Figure Description
[0059] Figure 1 A flowchart illustrating an artificial intelligence-based power grid operation safety detection method provided in this embodiment of the invention;
[0060] Figure 2 A flowchart of a model training method provided in an embodiment of the present invention. Detailed Implementation
[0061] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.
[0062] This invention provides an artificial intelligence-based method for detecting power grid operation safety. See also... Figure 1 , Figure 1 A flowchart illustrating an artificial intelligence-based power grid operation safety detection method provided in this embodiment of the invention. The method includes the following steps:
[0063] S101, acquire historical monitoring data of various dissolved gases in the transformer.
[0064] S102, perform modal decomposition and reconstruction on the historical sequence of the target dissolved gas to obtain multiple subsequences.
[0065] S103: For any subsequence, call the corresponding time series prediction model to make a prediction and obtain the predicted value of the subsequence.
[0066] S104. The predicted values of multiple subsequences are added together to obtain the predicted value of the target dissolved gas.
[0067] S105. Based on the predicted values of various dissolved gases, feature extraction is performed to obtain the first feature vector.
[0068] S106, input the first feature vector into the pre-trained state recognition model to obtain the operating state of the transformer.
[0069] The target dissolved gas is any one of multiple dissolved gases; the state recognition model is an artificial intelligence model obtained through training, and the output result includes a normal state and fault states such as partial discharge, low-energy discharge, high-energy discharge, medium-low temperature overheating, and high temperature overheating.
[0070] The method for detecting the safety of power grid operation based on artificial intelligence provided in the embodiment of the application can accurately capture the dynamic change characteristics of the gas sequence through modal decomposition, can improve the modeling capability of the nonlinear trend by combining multi-model collaborative prediction, can improve the prediction accuracy, can learn the fault state characteristics through the pre-trained artificial intelligence model, can realize intelligent evaluation of the operation state of the transformer, can timely and accurately identify potential faults, can reduce the misjudgment rate, can improve the scientific nature of the transformer maintenance decision, and can provide reliable safety protection for the operation of the power grid.
[0071] In one embodiment, step S102, the historical sequence of the target dissolved gas is subjected to modal decomposition and reconstruction to obtain multiple sub-sequences, including:
[0072] Step one, the historical sequence of the target dissolved gas is subjected to modal decomposition by using an improved complete ensemble empirical mode decomposition adaptive noise method (ICEEMDAN) to obtain multiple modal components and a residual sequence.
[0073] Step two, the average frequency of each modal component is calculated, and the modal components are divided into three groups of high-frequency components, medium-frequency components, and low-frequency components according to the average frequency of each modal component. Specifically, the average frequency is calculated by using Hilbert transform. The division method uses a k-means clustering algorithm.
[0074] Step three, the modal components in each group are summed to obtain three reconstructed sequences. The residual sequence and the three reconstructed sequences are taken as sub-sequences.
[0075] In one implementation, for each sub-sequence, a corresponding time series prediction model is trained separately. The ICEEMDAN decomposition can decompose the target dissolved gas sequence into modal components of different time scales, retaining the nonlinear and non-stationary characteristics. The high-frequency components capture short-term rapid fluctuations, and the medium-low frequency components reflect periodic load changes or long-term trends. After decomposition and reconstruction, the characteristics of different sub-sequences are separated, and the internal mode of each sub-sequence is more single, and the model can be optimized accordingly.
[0076] In one implementation, the time series prediction model can use a bidirectional long short-term memory network model. The hyperparameters of the model (such as the number of hidden layers, the number of hidden units in each layer, the learning rate, etc.) can be optimized by using a meta-heuristic algorithm (such as a particle swarm optimization algorithm, a genetic algorithm, etc.), thereby avoiding the limitations caused by empirical setting.
[0077] In one embodiment, step S105: performing feature extraction according to the predicted values of the plurality of dissolved gases to obtain a first feature vector includes:
[0078] Step one, generating a plurality of gas ratios according to the predicted values of the plurality of dissolved gases.
[0079] Step two, combining the predicted values of the plurality of dissolved gases and the plurality of gas ratios to obtain a second feature vector M is the length of the second feature vector.
[0080] Step three, converting the second feature vector using a preset weight matrix W (M x K) to obtain a first feature vector: Specifically, the length K of the first feature vector is less than the length M of the second feature vector.
[0081] Wherein, the weight matrix W is determined by the training sample set.
[0082] This embodiment maps the second feature vector into a first feature vector with a shorter length by using the weight matrix determined by the training sample, realizes multi-dimensional feature information compression, retains key discriminant features, reduces redundancy, and improves the accuracy and robustness of the state recognition model for fault states.
[0083] In one embodiment, the state recognition model can be a machine learning model (such as a random forest) or a neural network model (such as a convolutional neural network, a multilayer perceptron, etc.).
[0084] In one implementation, the state recognition model is an integrated model including a convolutional neural network base model, a probabilistic neural network base model, and a fully connected layer. The operation process of the state recognition model includes:
[0085] Inputting the first feature vector into the convolutional neural network base model and the probabilistic neural network base model respectively to obtain a first probability vector and a second probability vector.
[0086] Splicing the first probability vector and the second probability vector to form a third feature vector.
[0087] Inputting the third feature vector into the fully connected layer to output a third probability vector.
[0088] Determining the operating state of the transformer according to the class with the maximum probability value in the third probability vector.
[0089] This embodiment improves the discrimination ability for the state of the transformer through multi-model cooperation.
[0090] The embodiment of the application provides a model training method based on data enhancement. Referring to Figure 2 , Figure 2A flowchart of a model training method is provided for an embodiment of the present application. The method comprises the following steps:
[0091] S201, obtaining label data. Obtain a historical state data set of the transformer; the historical state data includes dissolved gas data and corresponding operating state labels. Specifically, a sample data is composed of features and labels. The features include the content data of five types of gases: hydrogen H2, methane CH4, acetylene C2H2, ethylene C2H4, and ethane C2H6.
[0092] S202, extracting gas features. Feature extraction is performed on the original dissolved gas data to obtain gas feature data. According to the gas feature data, a data set is established, denoted as a first data set. Specifically, the gas feature data includes original feature data and derived gas ratio data (21 types). The ratios include: .
[0093] S203, feature dimension reduction. According to the gas feature data, a weight matrix is generated, and the gas feature data of the first data set is processed by dimension reduction according to the weight matrix to obtain factor score feature data. According to the factor score feature data, a data set is established, denoted as a second data set. Specifically, the factor analysis method is used for feature dimension reduction.
[0094] S204, few-sample enhancement. The conditional generative adversarial network is used to generate data of rare fault states to obtain an incremental data set, denoted as a third data set. The third data set is integrated into the second data set to obtain an expanded data set, denoted as a fourth data set. Specifically, when the ratio of the number of samples of a certain fault to the maximum number of samples is less than 0.3, the class is recorded as a rare fault state class.
[0095] S205, sample adjustment. The fourth data set is filtered to obtain a training sample set.
[0096] S206, model training. The training sample set is used for model training to obtain a state detection model.
[0097] The transformer operating state data usually has a sample imbalance problem. Normal state samples are abundant, while some fault state samples are scarce, which makes it difficult for traditional model training to fully learn the characteristics of minority class faults, affecting the fault detection accuracy. This embodiment uses a conditional generative adversarial network to generate rare fault samples, makes up for the lack of minority class samples, and filters the enhanced data to remove abnormal or low-quality samples, ensuring the reliability and representativeness of the training set, thereby ensuring that the model can still maintain high recognition ability when facing rare faults.
[0098] In one embodiment, specifically, step S203 includes:
[0099] Step one, standardize the gas feature data to obtain a sample matrix N is the total number of samples, and M is the number of feature categories.
[0100] Step two, generate a correlation matrix R according to the sample matrix X .
[0101] Step three, process the correlation matrix R using the principal component method to obtain a principal component transformation matrix V . Specifically, perform eigenvalue decomposition on R to obtain M eigenvalues (sorted from large to small ) and their corresponding eigenvectors. Select the eigenvectors corresponding to the first K eigenvalues to form the principal component transformation matrix V . K is the minimum number of factors required to accumulate an acceptable level (such as 90%) of explained variance.
[0102] Step four, calculate the initial factor loading matrix A based on the principal component transformation matrix V . Wherein, is a diagonal matrix composed of the first K eigenvalues. represents the square root of the diagonal elements.
[0103] Step five, perform variance maximization rotation (Varimax) on the initial factor loading matrix to obtain the rotated factor loading matrix A_r . Variance maximization rotation is a commonly used rotation method and will not be described here.
[0104] Step six, calculate the weight matrix W based on the rotated factor loading matrix A_r and the correlation matrix R using the regression method .
[0105] This embodiment generates a weight matrix by factor analysis method, realizes dimension reduction, key feature highlighting, and redundancy suppression of gas features, provides high-quality input features for the state recognition model, and improves the model training efficiency, discrimination ability, and generalization performance.
[0106] In one embodiment, step S205: filtering the fourth data set to obtain a training sample set includes:
[0107] Step one, delete duplicate samples in the fourth data set to obtain a fifth data set.
[0108] Step two, identify noise in the samples in the fifth data set and delete the noise data to obtain a sixth data set.
[0109] Step three, perform sample quantity balancing processing on the sixth data set to obtain a training sample set.
[0110] The embodiment constructs a high-quality and balanced training sample set through deduplication, noise removal and sample balancing processing, which can effectively improve the training effect of the state recognition model.
[0111] In an implementation manner, noise recognition is performed on the samples in the fifth data set, and noise data is removed to obtain a sixth data set.
[0112] Step one, support vector machines are used to classify each sample in the fifth data set, and sample data with classification errors are included in the weak feature data set.
[0113] Step two, any sample in the weak feature data set is taken as a target candidate sample, and the reference class of the target candidate sample is determined according to the local neighborhood samples of the target candidate sample. Specifically, a predetermined number (such as 5) of nearest neighbor samples of the target candidate sample are found by using the Euclidean distance, and the class labels of the plurality of nearest neighbor samples are counted, and the class with the highest proportion is taken as the reference class.
[0114] Step three, if the true label of the target candidate sample is inconsistent with the reference class, the target candidate sample is determined as a noise sample.
[0115] Step four, all noise samples are removed from the fifth data set to obtain a sixth data set.
[0116] The implementation manner can efficiently and accurately identify and remove noise samples by combining simple classification results and neighborhood statistics.
[0117] In an implementation manner, the sixth data set is subjected to sample number balancing processing to obtain a training sample set.
[0118] Step one, the sixth data set is divided into a plurality of class groups according to the label class.
[0119] Step two, the feature distribution center of the target class group is calculated. The target class group is any one of the plurality of class groups. The feature distribution center is the mean vector of the features of all samples in the class group, that is, the mean in vector form.
[0120] Step three, the Euclidean distance between each sample in the target class group and the feature distribution center thereof is calculated, which is recorded as the outlier score of the sample.
[0121] Step four, the samples in the target class group are sorted in descending order according to the outlier score, and the first L samples are selected to be included in the training sample set. L is not greater than the minimum value of the number of samples in each class group, and can be set to the minimum value of the number of samples in each class group.
[0122] The implementation mode performs sample screening based on an outlier score, retains typical samples with features close to a category center, rejects edge samples, and makes the feature distribution of each category more concentrated. The number of samples of each category is balanced to ensure that the training set is uniformly distributed among categories, avoid model bias to the majority class, and improve the identification ability of the minority class fault.
[0123] The above is only a preferred embodiment of the present application, and does not limit the present application in any form. Although the present application has been disclosed as above with a preferred embodiment, it is not intended to limit the present application. Any person skilled in the art can make slight changes or modifications to the above disclosed technical content to obtain equivalent embodiments with equivalent changes, without departing from the technical solution of the present application. Any simple modification, equivalent change and modification of the above embodiments made according to the technical essence of the present application, without departing from the technical solution of the present application, are within the scope of the technical solution of the present application.
Claims
1. An artificial intelligence-based power grid operation safety detection method, characterized in that, The method comprises: obtaining historical monitoring data of multiple dissolved gases of a transformer; performing modal decomposition and reconstruction on a historical sequence of a target dissolved gas to obtain multiple sub-sequences; the target dissolved gas is any one of the multiple dissolved gases; for any one sub-sequence, calling a corresponding time series prediction model to perform prediction to obtain a sub-sequence prediction value; adding the multiple sub-sequence prediction values to obtain a prediction value of the target dissolved gas; performing feature extraction according to the prediction values of the multiple dissolved gases to obtain a first feature vector; inputting the first feature vector into a pre-trained state recognition model to obtain the running state of the transformer; the state recognition model is an artificial intelligence model obtained by training.
2. The power grid operation safety detection method based on artificial intelligence according to claim 1, characterized in that, The multiple sub-sequences include a residual sequence and multiple reconstructed sequences; the modal decomposition and reconstruction on the historical sequence of the target dissolved gas to obtain the multiple sub-sequences comprise: performing modal decomposition on the historical sequence of the target dissolved gas by using an improved complete ensemble empirical mode decomposition adaptive noise method to obtain multiple modal components and a residual sequence; calculating the average frequencies of the modal components, and dividing the modal components into three groups of high-frequency components, medium-frequency components and low-frequency components according to the average frequencies of the modal components; summing the modal components in each group to obtain three reconstructed sequences.
3. The method of claim 1, wherein the method further comprises: The time series prediction model adopts a bidirectional long short-term memory network model; for each sub-sequence, a corresponding time series prediction model is trained separately.
4. The power grid operation safety detection method based on artificial intelligence according to claim 1, characterized in that, the feature extraction according to the prediction values of the multiple dissolved gases to obtain the first feature vector comprises: generating multiple gas ratios according to the prediction values of the multiple dissolved gases; combining the prediction values of the multiple dissolved gases and the multiple gas ratios to obtain a second feature vector; performing conversion on the second feature vector by using a pre-set weight matrix to obtain the first feature vector; the weight matrix is determined by a training sample set.
5. The method of claim 1, wherein the method is based on artificial intelligence. The state recognition model is an integrated model comprising a convolutional neural network base model, a probabilistic neural network base model and a fully connected layer; the operation process of the state recognition model comprises: inputting the first feature vector into the convolutional neural network base model and the probabilistic neural network base model respectively to obtain a first probability vector and a second probability vector; splicing the first probability vector and the second probability vector to form a third feature vector; inputting the third feature vector into the fully connected layer to output a third probability vector; determining the running state of the transformer according to the class with the maximum probability value in the third probability vector.
6. The power grid operation safety detection method based on artificial intelligence according to claim 5, characterized in that, The training process of the state recognition model comprises: obtaining a historical state data set of the transformer; the historical state data comprises dissolved gas data and corresponding running state labels; performing feature extraction on the original dissolved gas data to obtain gas feature data; a first data set is established according to the gas feature data; generating a weight matrix according to the gas feature data; performing dimensionality reduction processing on the gas feature data of the first data set according to the weight matrix to obtain factor score feature data, and establishing a second data set according to the factor score feature data; Adopting a conditional generative adversarial network to generate data of a rare fault state to obtain a third data set; the rare fault state is determined by the number of samples in the first data set; Integrating the third data set into the second data set to obtain a fourth data set; Filtering the fourth data set to obtain a training sample set; Adopting the training sample set to perform model training to obtain a state detection model.
7. The power grid operation safety detection method based on artificial intelligence according to claim 6, characterized in that, The generating a weight matrix according to the gas feature data comprises: Performing standardization processing on the gas feature data to obtain a sample matrix; Generating a correlation matrix according to the sample matrix; Adopting a principal component method to process the correlation matrix to obtain a principal component transformation matrix; Calculating an initial factor loading matrix according to the principal component transformation matrix; Performing variance maximization rotation on the initial factor loading matrix to obtain a rotated factor loading matrix; Calculating a weight matrix according to the rotated factor loading matrix and the correlation matrix.
8. The power grid operation safety detection method based on artificial intelligence according to claim 6, characterized in that, The filtering the fourth data set to obtain a training sample set comprises: Deleting duplicate samples in the fourth data set to obtain a fifth data set; Performing noise identification on samples in the fifth data set and deleting noise data to obtain a sixth data set; Performing sample number balancing processing on the sixth data set to obtain a training sample set.
9. The power grid operation safety detection method based on artificial intelligence according to claim 8, characterized in that, The performing noise identification on samples in the fifth data set and deleting noise data to obtain a sixth data set comprises: Adopting a support vector machine to classify each sample of the fifth data set and including sample data with classification errors into a weak feature data set; According to local neighborhood samples of a target candidate sample, performing category statistics to determine a reference category thereof; the target candidate sample is any sample in the weak feature data set; If a true label of the target candidate sample is inconsistent with the reference category thereof, the target candidate sample is determined as a noise sample; Deleting all noise samples from the fifth data set to obtain a sixth data set.
10. The power grid operation safety detection method based on artificial intelligence according to claim 8, characterized in that, The performing sample number balancing processing on the sixth data set to obtain a training sample set comprises: Dividing the sixth data set into multiple category groups according to label categories; Calculating a feature distribution center of a target category group; the target category group is any one of the multiple category groups; Calculating the Euclidean distance between each sample in the target category group and the feature distribution center thereof, recorded as an outlier score of the sample; Performing descending order sorting on samples in the target category group according to the outlier scores, and selecting the first L samples to be included in the training sample set; L is not greater than the minimum value of the number of samples in each category group.