A transformer fault early warning model establishment method, early warning method and device

Abstract

The application provides a transformer fault early warning model establishment method, an early warning method and device, and belongs to the technical field of artificial intelligence. It solves the problem of existing transformer fault detection not being timely. The transformer fault early warning model establishment method, the early warning method and the device comprise the following steps: step S1: collecting historical account data, training a graph neural network model and recording a graph adjacency matrix, step S2: merging the graph adjacency matrices generated in the training process of all transformer devices, and determining the mapping information of the graph adjacency matrix through a decision tree model. The application has the advantages of realizing high-precision early warning of the transformer according to the fault probability output by the graph adjacency matrix and the change of the gas relationship before and after the change in real-time early warning applications.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of artificial intelligence, and relates to a transformer fault early warning model establishment method, an early warning method and a device, in particular to a transformer fault early warning model establishment method and device based on a graph neural network and a decision tree model. BACKGROUND

[0002] Transformer oil is a fractional product of petroleum, and the main components are alkanes, cycloalkane saturated hydrocarbons, aromatic unsaturated hydrocarbons and the like. The insulating oil of the transformer is a mixture of many different molecular weight hydrocarbon molecules, and the molecules contain CH3, CH2 and CH chemical groups and are bonded together by C-C bonds. As a result of electric or thermal faults, some C-H bonds and C-C bonds can be broken, and a small amount of active hydrogen atoms and unstable hydrocarbon free radicals are generated, and these hydrogen atoms or free radicals are recombined through complex chemical reactions to form hydrogen and low hydrocarbon gases such as methane, ethane, ethylene, acetylene and the like. In the initial stage of transformer fault, the generated gas is dissolved in the oil, and analyzing the components and content of the dissolved gas in the oil is one of the most effective measures to monitor the safe operation of oil-filled electrical equipment.

[0003] The traditional transformer fault warning scheme based on dissolved gas in oil is mainly divided into three directions: key value based, machine learning based, and neural network based. The key value based analysis method includes the three-ratio method which has been used in China for more than 30 years. By comparing the relative concentrations of hydrogen, methane, ethane, ethylene, and acetylene with the mutual dependence of temperature, and selecting the ratio of two gas components with similar solubility and diffusion coefficient as the basis for judging the nature of the fault. With the development of power production and the improvement of scientific and technological level, the analysis method and the judgment and interpretation of the analysis results are supplemented and revised. In 2000, the China Electric Power Research Institute proposed the "Guide to Analysis and Judgment of Dissolved Gases in Transformer Oil". Other key value analysis methods include the Rutgers index, Duval triangle, frequency response analysis, spectral analysis (Hilbert Huang transform, fast Fourier transform, wavelet transform, and combined transform), etc. which are also widely used in transformer fault warning. The machine learning based analysis method includes: the mixed least square support vector regression (M-LS-SVR) algorithm proposed in 2016, which uses a combination of linear and nonlinear functions as the prediction function and uses real data to adaptively select the mixing ratio factor; the oil dissolved gas concentration prediction scheme based on quadratic dimension reduction proposed in 2017, which is based on gray model and support vector machine (SVM); a transformer oil dissolved gas concentration prediction scheme based on random forest proposed in 2019, etc. The neural network based analysis method includes: a transformer oil dissolved gas concentration prediction model based on long short-term memory network (LSTM) proposed in 2020; a prediction method named CEEMD-TCN model proposed in 2020, which decomposes the original sequence into sub-sequence components of different scales, pre-processes them, trains the time convolution network, optimizes the network hyperparameters, and superimposes the prediction results of each component to obtain the final prediction result; using BP neural network for prediction; a prediction method based on empirical mode decomposition and long short-term memory neural network; a prediction method based on deep belief network, etc.

[0004] The key value analysis method relies on prior knowledge and has poor applicability, cannot be fine-tuned on devices in different operating environments, operation and maintenance, communication and external interference, and lacks prediction accuracy and cannot completely capture the information between the account data and the time and space of the data, and cannot achieve early warning. The fault diagnosis scheme based on machine learning has certain applicability, interpretability and predictability, but must pre-establish assumptions such as stationarity and linear dependence between gas pairs, is usually suitable for data with deterministic trends, is difficult to adapt to production data in complex environments, and the calculation difficulty increases geometrically with the increase in the number of gases, resulting in poor prediction accuracy. The fault diagnosis scheme based on neural network does not pre-set data conditions, has strong non-linear fitting ability, and has strong self-learning ability and simple learning rules, but the model training requires a large amount of sample data, the prediction effect of the model is greatly related to the selection of the initial hyperparameters, the learning speed is slower than other popular algorithms, and it is easy to fall into a local minimum point. Although there are improved BP algorithms, genetic algorithms, simulated annealing algorithms and other optimization methods, it is still difficult to obtain the global minimum point in practical application, and the relationship between gas pairs cannot be learned, which weakens the interpretability of the model.

[0005] A graph is a special representation of data used to describe the association between nodes. Graph neural networks have been very successful in processing graph data in recent years based on their permutation invariance, local invariance and composition. In the forward propagation process, the graph neural network allows each node to learn its adjacent node information. The dissolved gas in each oil can also be regarded as a node in the graph, and the relationship between the nodes can be used to map the health status of the device. The most common graph neural network model is the spatio-temporal convolution structure, which inputs multi-dimensional time series data and a graph structure as input to predict the data value or category at the next time. Compared with traditional multi-dimensional time series prediction methods, this scheme has obvious precision improvement, but has two major defects. The current graph neural network scheme needs to predefine the graph structure to achieve time series prediction, but in reality it is impossible to obtain a predefined graph structure of the dissolved gas in the transformer oil, because the dependence relationship between the gases is unknown and needs to be learned and mined through data rather than predefined. In addition, even if the predefined graph structure is obtained, most of the current graph neural networks only focus on information feedforward and ignore the problem that the relationship between the nodes in the graph structure will change during training. For example, in the graph structure generated based on the dissolved gas data in the oil, the one-way relationship between each pair of gases is not fixed, but constantly updated over time, and this change is particularly obvious before and after the fault occurs.

[0006] Graph neural networks can be divided into five categories: graph convolution networks (GCN), graph attention networks, graph autoencoders, graph generative networks, and graph spatial-temporal networks. Graph convolution networks apply convolutional computation to graph data and are the basis of many complex graph neural network models, including models based on autoencoders, generative models, and spatial-temporal networks. GCN methods can be divided into two categories: spectral-based and spatial-based. Spectral-based methods introduce filters from the perspective of graph signal processing to define graph convolution, where graph convolution operations are interpreted as removing noise from graph signals. Spatial-based methods represent graph convolution as aggregating feature information from the neighborhood. When the algorithm of the graph convolution network runs at the node level, the graph pooling module can be interleaved with the graph convolution layer to coarsen the graph into high-level substructures. In this scheme, the idea of GCN is mainly used to generate a graph adjacency matrix to map the one-way relationship of gas.

[0007] In IEC standards, the three-ratio method determines the fault type by comparing the size of the concentration ratio of dissolved gas components in the oil. However, the three-ratio method can only be used after the fault occurs and cannot provide early warning. The ratio is absolutely fixed and cannot be fine-tuned on transformers under different operation, environment, and communication. In this scheme, the graph neural network learns the one-way relationship between the dissolved gases in the oil and generates a graph adjacency matrix to express it, which can map the running state of the equipment at different times. At the same time, a decision tree model is used to determine whether the transformer will fail at the next time based on the graph adjacency matrix, greatly enhancing the explainability of the graph network. SUMMARY

[0008] The purpose of the present application is to solve the above-mentioned problems in the prior art, and to provide a transformer fault early warning model establishment method, early warning method and device based on graph neural network and decision tree model.

[0009] The first object of the present application can be achieved by the following technical scheme: a transformer fault early warning model establishment method, characterized in that it comprises the following steps:

[0010] Step S1: collect historical account data, train a graph neural network model, and record a graph adjacency matrix:

[0011] Step S1.1: collect historical account data of N kinds of gases dissolved in transformer oil and arrange them in order;

[0012] Step S1.2: filling in missing values in historical account data;

[0013] Step S1.3: collecting input sample set and output sample set;

[0014] Step S1.4: training each group of input sample and output sample in time sequence, and normalizing the last data of input sample and output sample using standard deviation normalization method;

[0015] Step S1.5: mapping the last N gas values of the normalized output sample to a two-dimensional space through an embedding layer to generate an N×N matrix, setting two mapping layers to obtain two output result matrices: E1 and E2, as formula (3):

[0016]

[0017] ;

[0018] Step S1.6: linearly converting the two output result matrices, and the converted matrices are still N×N matrices: Y1 and Y2, as formula (4):

[0019]

[0020] ;

[0021] Step S1.7: using sigmoid function to activate the two converted matrices, and setting a hyperparameter for controlling the activation rate before activation ;

[0022] Step S1.8: calculating the obtained graph adjacency matrix M;

[0023] Step S1.9: generating a comparison graph adjacency matrix M2 according to the gas change at the next moment;

[0024] Step S1.10: the loss function MAPE is the average absolute percentage error, and the error calculated by M and M2 is back propagated to update the model parameters;

[0025] Step S1.11: training each group of input sample and output sample according to steps S1.1-S1.10, retaining the optimal model parameters of each group of input sample and output sample, calling the optimal model parameters obtained by training the previous group of input sample and output sample for training the next group of input sample and output sample, and finally training the last group of input sample and output sample to obtain the final model, and saving the graph adjacency matrix in the model training process;

[0026] Step S2: Merge the graph adjacency matrix generated in the transformer equipment training process, and judge the mapping information of the graph adjacency matrix through the decision tree model:

[0027] Step S2.1: Merge the graph adjacency matrix generated by training each transformer according to the dissolved gas data in the oil to form a data sample set, and mark the graph adjacency matrix that will fail at the next time as a fault data sample set, and mark the other graph adjacency matrices as normal data sample sets;

[0028] Step S2.2: Convert each graph adjacency matrix into a vector;

[0029] Step S2.3: Divide the training data set and the verification data set according to the preset proportion for all data sample sets;

[0030] Step S2.4: According to the decision tree ID3 algorithm model, select the attribute of the graph adjacency matrix as normal or fault through the information gain ratio criterion;

[0031] Step S2.5: Complete the feature extraction work on the training data set, and verify the model accuracy on the verification data set.

[0032] The present application is directed to the dissolved gas data of different transformer devices, which is regarded as graph data, and the one-way relationship change between different gases is captured through the graph neural network and expressed in the form of graph adjacency matrix. According to the one-way relationship of the gas in the normal and fault states, a decision tree model is constructed, and the judgment of the graph adjacency matrix through the decision tree model can early warn whether the transformer will fail at the next time. The present application proposes to use a polynomial fitting method to repair the low-quality account data, use a graph neural network to capture the one-way relationship between the gases in real time, generate a graph adjacency matrix without prior knowledge, and obtain a decision tree model for judging the graph adjacency matrix in the normal and fault states through an information gain algorithm. In real-time early warning applications, according to the fault probability output by the graph adjacency matrix, the changes before and after the gas relationship are combined to realize high-precision early warning of the transformer.

[0033] In the transformer fault early warning model establishment method described above, in step S1.1, the gases dissolved in the transformer oil are 9, including hydrogen (H2), methane (CH4), ethane (C2H6), ethylene (C2H4), acetylene (C2H2), total hydrocarbon (TOTAL HYDROCARBON), carbon monoxide (CO), carbon dioxide (CO2), and oxygen (O2). In steps S1.5, S1.6 and S2.2, the matrix is a 9x9 matrix.

[0034] In the transformer fault early warning model establishment method described above, in step S1.2, the filling method is a 5-order polynomial fitting method, as shown in formula (1):

[0035]

[0036] .

[0037] In the transformer fault early warning model establishment method described above, in step S1.3, the input sample length is 18, which is the value of 9 kinds of dissolved gases in the past 18 time points, the output sample length is 1, which is the value of 9 kinds of dissolved gases in the next time point, the sliding window length is 1, which means collecting a sample set every 1 time point, and let be the 9 kinds of dissolved gas data values at time t, be the value of the i-th gas at time t, be a group of input samples, be the corresponding group of output samples.

[0038] In the transformer fault early warning model establishment method described above, in step S1.4, the standard deviation normalization method is as formula (2):

[0039] ;

[0040] be the mean value of each gas in each group of input samples, be the standard deviation of each gas, if the value of a certain gas in the input sample is constant, then set be 0.00001 to prevent normalization failure.

[0041] In the transformer fault early warning model establishment method described above, in step S1.7, the activation method is as formula (5):

[0042]

[0043]

[0044] ;

[0045] In step S1.8, the graph adjacency matrix is calculated by formula (6):

[0046] ;

[0047] In step S1.9, the growth rate of gas data is first calculated , as formula (7):

[0048] ;

[0049] be the last data of the normalized input sample, be the normalized output sample data

[0050] M2 is generated by comparing the size between the ratios as formula (8):

[0051] ,

[0052] In step S1.10, the calculation method of the loss function is as formula (9):

[0053] ;

[0054] In step S1.11, each group of input samples and output samples is repeatedly trained for 100 rounds, the optimizer is Adam, the initial learning rate is 0.001, and the learning rate is decayed by 0.5 every 20 rounds.

[0055] In the transformer fault early warning model establishment method described above, in step S2.2, each graph adjacency matrix with a shape of 9x9 is converted into a vector with a length of 81 , represents the one-way relationship between the gas to the gas , according to the characteristics of the graph adjacency matrix, when has a weight, is 0, the weight is 0,

[0056] In step S2.4, the decision tree ID3 algorithm model is as formula (10):

[0057]

[0058]

[0059]

[0060]

[0061] Information gain represents the degree of reduction in uncertainty of the belonging class when knowing the feature , the information gain ratio is the ratio of the information gain and the empirical entropy H(D) of the training set D, the information gain g(D, A) of the feature A to the training data set D is defined as the difference between the empirical entropy H(D) of the set D and the empirical conditional entropy H(D|A) of D under the condition of the feature A, is the probability of the classification k of any sample data in the set D;

[0062] In step S2.5, the evaluation indicators are calculated, including the , , , of the predicted category and the real category:

[0063]

[0064]

[0065]

[0066]

[0067] TP represents that the positive classification prediction of the classifier is correct, TN represents that the negative classification prediction of the classifier is correct, FN represents that the positive classification prediction of the classifier is incorrect, and FP represents that the negative classification prediction of the classifier is incorrect.

[0068] In the transformer fault early warning model establishment method, in step S2.3, the preset ratio of the training data set and the verification data set is 4:1.

[0069] The second object of the application can be achieved by the following technical scheme: a transformer fault early warning method, characterized in that the early warning model is established according to the above-mentioned transformer fault early warning model establishment method, and further comprising the following steps:

[0070] Step S3: receiving the latest dissolved gas in oil value of the transformer equipment returned by the sensor to the server, real-time pushing the code of the transformer equipment, collecting historical gas values at multiple time points in the past, merging the latest data, and arranging the dissolved gas in oil in sequence;

[0071] Step S4: normalizing the input dissolved gas in oil data, recording the mean value and standard deviation of each gas, and the normalized result of the last gas data ;

[0072] Step S5: putting into the trained graph neural network model to obtain a real-time graph adjacency matrix, mapping the one-way relationship between gas nodes;

[0073] Step S6: converting the graph adjacency matrix into a vector and putting it into the trained decision tree model to obtain the probability of whether a fault will occur at the next moment;

[0074] Step S7: combining the predicted value size, judging whether a fault will occur at the next moment through the guide, and the fault probability output by the decision tree model, to judge the running state of the transformer at the next moment.

[0075] In the transformer fault early warning method, in step S3, 17 historical gas values at different times are collected, and the dissolved gases in the oil are 9 kinds, including hydrogen (H2), methane (CH4), ethane (C2H6), ethylene (C2H4), acetylene (C2H2), total hydrocarbon (TOTAL HYDROCARBON), carbon monoxide (CO), carbon dioxide (CO2), and oxygen (O2).

[0076] In step S6, the length of the vector is 81.

[0077] The third object of the application can be achieved by the following technical solution: a transformer fault early warning device, comprising an input end, an output end, one or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and are configured to be executed by the one or more processors, and the programs comprise a program for executing the above-mentioned transformer fault early warning method.

[0078] Compared with the prior art, the application has the advantages of high-precision early warning of the transformer in real-time warning application, according to the fault probability output by the graph adjacency matrix, and combining the changes before and after the gas relationship. BRIEF DESCRIPTION OF DRAWINGS

[0079] Figure 1 is a flowchart of a transformer fault early warning model establishment method of the application.

[0080] Figure 2 is a flowchart of step 1 of a transformer fault early warning model establishment method of the application.

[0081] Figure 3 is a flowchart of step 2 of a transformer fault early warning model establishment method of the application.

[0082] Figure 4 is a flowchart of a transformer fault early warning method of the application. DETAILED DESCRIPTION

[0083] The following is a specific embodiment of the application and further describes the technical solutions of the application in conjunction with the drawings, but the application is not limited to these embodiments.

[0084] As shown in Figures 1-4 A transformer fault early warning model establishment method, characterized in that it comprises the following steps:

[0085] Step S1: Collect historical account data, train a graph neural network model, and record a graph adjacency matrix:

[0086] Step S1.1: Collect historical account data of N kinds of dissolved gases in transformer oil, and arrange them in order;

[0087] Step S1.2: filling in missing values in historical account data;

[0088] Step S1.3: collecting input sample set and output sample set;

[0089] Step S1.4: training each group of input sample and output sample in time sequence, and normalizing the last data of input sample and output sample using standard deviation normalization method;

[0090] Step S1.5: mapping the last N gas values of the normalized output sample to a two-dimensional space through an embedding layer to generate a matrix of shape N*N, setting two mapping layers to obtain two output result matrices: E1 and E2, as formula (3):

[0091]

[0092] ;

[0093] Step S1.6: linearly converting the two output result matrices, and the converted matrices are still matrices of shape N*N: Y1 and Y2, as formula (4):

[0094]

[0095] ;

[0096] Step S1.7: using sigmoid function to activate the two converted matrices, and setting a hyperparameter for controlling the activation rate before activation ;

[0097] Step S1.8: calculating to obtain graph adjacency matrix M;

[0098] Step S1.9: generating a comparison graph adjacency matrix M2 according to the gas change at the next moment;

[0099] Step S1.10: loss function MAPE is the mean absolute percentage error, and the error calculated by M and M2 is back propagated to update the model parameters;

[0100] Step S1.11: training each group of input sample and output sample according to steps S1.1-S1.10, retaining the optimal model parameters of each group of input sample and output sample, calling the optimal model parameters obtained by training the previous group of input sample and output sample for training the next group of input sample and output sample, and finally training the last group of input sample and output sample to obtain the final model, and saving the graph adjacency matrix in the model training process;

[0101] Step S2: Merge the graph adjacency matrix generated in the transformer equipment training process, and judge the mapping information of the graph adjacency matrix through the decision tree model:

[0102] Step S2.1: Merge the graph adjacency matrix generated by training each transformer according to the dissolved gas data in the oil to form a data sample set, and mark the graph adjacency matrix that will fail at the next time as a fault data sample set, and mark the other graph adjacency matrices as normal data sample sets;

[0103] Step S2.2: Convert each graph adjacency matrix into a vector;

[0104] Step S2.3: Divide the training data set and the verification data set according to the preset proportion for all data sample sets;

[0105] Step S2.4: According to the decision tree ID3 algorithm model, select the attribute of the graph adjacency matrix as normal or fault through the information gain ratio criterion;

[0106] Step S2.5: Complete the feature extraction work on the training data set, and verify the model accuracy on the verification data set.

[0107] Further, in step S1.1, the dissolved gas in the transformer oil is 9 kinds, including hydrogen (H2), methane (CH4), ethane (C2H6), ethylene (C2H4), acetylene (C2H2), total hydrocarbon (TOTAL HYDROCARBON), carbon monoxide (CO), carbon dioxide (CO2), and oxygen (O2). The matrix in steps S1.5, S1.6 and S2.2 is a 9x9 matrix.

[0108] Further, in step S1.2, the filling method is a 5th order polynomial fitting method, as shown in formula (1):

[0109]

[0110] .

[0111] Further, in step S1.3, the input sample length is 18, which is the 9 kinds of dissolved gas values in the past 18 times, the output sample length is 1, which is the 9 kinds of dissolved gas values in the next time, the sliding window length is 1, which is to collect a sample set every 1 time, and let be the 9 kinds of dissolved gas data values at time t, be the value of the i-th gas at time t, be a group of input samples, be a corresponding group of output samples.

[0112] Further, in step S1.4, the standard deviation normalization method is as formula (2):

[0113] ;

[0114] is the mean value of each gas in each group of input samples, is the standard deviation of each gas, and if the value of a certain gas in the input sample is constant, it is set to 0.00001 to prevent non-normalization.

[0115] Further, in step S1.7, the activation method is as formula (5):

[0116]

[0117]

[0118] ;

[0119] In step S1.8, the graph adjacency matrix is calculated by formula (6):

[0120] ;

[0121] In step S1.9, the growth rate of gas data is calculated first , as formula (7):

[0122] ;

[0123] is the last data of the normalized input sample, is the normalized output sample data

[0124] M2 is generated by comparing the size between the rates, as formula (8):

[0125] .

[0126] In step S1.10, the calculation method of the loss function is as formula (9):

[0127] ;

[0128] In step S1.11, each group of input samples and output samples is repeatedly trained for 100 rounds, the optimizer is Adam, the initial learning rate is 0.001, and the learning rate is decayed by 0.5 every 20 rounds.

[0129] Further, in step S2.2, each graph adjacency matrix with a shape of 9x9 is converted into a vector with a length of 81 , representing the gas one-way relationship between the gas According to the characteristics of the graph adjacency matrix, when has a weight, then 0, the weight is 0,

[0130] In step S2.4, the decision tree ID3 algorithm model is as formula (10):

[0131]

[0132]

[0133]

[0134]

[0135] Information gain represents the degree of information that the feature reduces the uncertainty of the belonging category, and the information gain ratio is the ratio of the information gain and the empirical entropy H(D) of the training set D, and the information gain g(D, A) of the feature A to the training data set D is defined as the difference between the empirical entropy H(D) of the set D and the empirical conditional entropy H(D|A) of D under the condition of the feature A, is the probability of the classification k of any sample data in the set D;

[0136] In step S2.5, the evaluation index is calculated, including the , , , :

[0137]

[0138]

[0139]

[0140]

[0141] TP represents the number of correct positive classification predictions by the classifier, TN represents the number of correct negative classification predictions by the classifier, FN represents the number of incorrect positive classification predictions by the classifier, and FP represents the number of incorrect negative classification predictions by the classifier.

[0142] Further, in step S2.3, the preset ratio of the training data set and the validation data set is 4:1.

[0143] The application discloses a transformer fault early warning method, and relates to the technical field of transformer fault early warning.

[0144] Step S3: receiving the latest oil dissolved gas value of the transformer device returned by the sensor to the server, pushing the code of the transformer device in real time, collecting historical gas values at multiple time points in the past, merging the latest data, and arranging the oil dissolved gas in sequence;

[0145] Step S4: normalizing the input oil dissolved gas data, recording the mean value and standard deviation of each gas, and the normalized result of the last gas data ;

[0146] Step S5: putting into the trained graph neural network model to obtain a real-time graph adjacency matrix, and mapping the one-way relationship between the gas nodes;

[0147] Step S6: converting the graph adjacency matrix into a vector and putting it into the trained decision tree model to obtain the probability of whether a fault will occur at the next moment;

[0148] Step S7: combining the predicted value size, judging whether a fault will occur at the next moment through the guide, and judging the running state of the transformer at the next moment according to the fault probability output by the decision tree model.

[0149] Further, in step S3, 17 time points of historical gas values are collected, and the oil dissolved gas is 9 kinds, including hydrogen (H2), methane (CH4), ethane (C2H6), ethylene (C2H4), acetylene (C2H2), total hydrocarbon (TOTAL HYDROCARBON), carbon monoxide (CO), carbon dioxide (CO2) and oxygen (O2).

[0150] In step S6, the length of the vector is 81.

[0151] A transformer fault early warning device comprises an input end, an output end, one or more processors, a memory and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, and the programs comprise a program for executing the transformer fault early warning method.

[0152] The specific embodiments described herein are merely illustrative of the spirit of the application. Those skilled in the art of the application can make various modifications or supplements to the described specific embodiments or replace them with similar ways, without departing from the spirit of the application or exceeding the scope defined by the appended claims.

[0153] Although a large number of terms are used herein, the possibility of using other terms is not excluded. The use of these terms is merely intended to facilitate the description and explanation of the essence of the present application; it is against the spirit of the present application to interpret them as any kind of additional limitation.

Claims

1. A method for establishing a transformer fault early warning model, characterized in that, Includes the following steps: Step S1: Collect historical ledger data, train the graph neural network model, and record the graph adjacency matrix: Step S1.1: Collect historical data on the N gases dissolved in transformer oil and arrange them in order; Step S1.2: Fill in the missing values ​​in the historical ledger data; Step S1.3: Collect the input and output sample sets; in this step S1.3, the input sample length is 18, representing the dissolved gas values ​​of 9 types of oil over the past 18 time points; the output sample length is 1, representing the dissolved gas values ​​of 9 types of oil at the next time point; the sliding window length is 1, meaning that a sample set is collected every 1 time point. Here are the dissolved gas data values ​​for nine types of oil at time t. Let be the value of the i-th gas at time t. For a set of input samples, For the corresponding set of output samples; Step S1.4: Train each group of input and output samples in chronological order, and normalize the last data point of the input sample and the output sample using the standard deviation normalization method; Step S1.5: Map the N normalized gas values ​​of the last output sample to a two-dimensional space through an embedding layer to generate an N×N matrix. Set two mapping layers to obtain two output result matrices: E1 and E2, as shown in formula (3): ； Step S1.6: Perform a linear transformation on the two output matrices. The transformed matrices are still N×N matrices: Y1 and Y2, as shown in formula (4): ； Step S1.7: Activate the two transformed matrices using the sigmoid function, and set a hyperparameter to control the activation rate before activation. ; Step S1.8: Calculate and obtain the graph adjacency matrix M; Step S1.9: Generate the adjacency matrix M2 of the comparison map based on the gas changes at the next time step; Step S1.10: The loss function MAPE is the mean absolute percentage error. The errors calculated by M and M2 are backpropagated to update the model parameters. Step S1.11: Train each set of input and output samples according to steps S1.1 to S1.

10. The optimal model parameters are retained for each set of input and output samples. The next set of input and output samples is trained using the optimal model parameters obtained from the previous set of input and output samples. The final model is obtained after the last set of input and output samples is trained, and the graph adjacency matrix during the model training process is saved. Step S2: Merge the graph adjacency matrices generated during the training of all transformer devices, and determine the mapping information of the graph adjacency matrices using a decision tree model: Step S2.1: Merge the graph adjacency matrices generated by training based on dissolved gas data in oil for each transformer to form a data sample set. Mark the graph adjacency matrices that will fail in the next moment as the fault data sample set, and mark the other graph adjacency matrices as the normal data sample set. Step S2.2: Convert each graph adjacency matrix into a vector; Step S2.3: Divide all data sample sets into training datasets and validation datasets according to a preset ratio; Step S2.4: Based on the decision tree ID3 algorithm model, select the attribute of the adjacency matrix of the partition graph as normal or faulty using the information gain ratio criterion; Step S2.5: Perform feature extraction on the training dataset and verify the model accuracy on the validation dataset.

2. The method for establishing a transformer fault early warning model according to claim 1, characterized in that, In step S1.1, the transformer oil contains nine dissolved gases, including hydrogen, methane, ethane, ethylene, acetylene, total hydrocarbons, carbon monoxide, carbon dioxide, and oxygen. The matrices in steps S1.5, S1.6, and S2.2 are 9×9 matrices.

3. The method for establishing a transformer fault early warning model according to claim 1, characterized in that, In step S1.2, the filling method is a 5th-order polynomial fitting method, as shown in formula (1): 。 4. The method for establishing a transformer fault early warning model according to claim 1, characterized in that, In step S1.4, the standard deviation normalization method is as shown in formula (2): ； This represents the mean value for each gas in each input sample group. For each gas, set the standard deviation. If the value of a certain gas in the input sample remains constant, then... The value is set to 0.00001 to prevent non-normalization.

5. The method for establishing a transformer fault early warning model according to claim 1, characterized in that, In step S1.7, the activation method is as shown in formula (5): ； In step S1.8, the graph adjacency matrix is ​​calculated using formula (6): ； In step S1.9, the growth rate of the gas data is first calculated. As shown in formula (7): ； This is the last data point of the input sample after normalization. For the normalized output sample data, M2 is generated by comparing the magnitudes of the ratios, as shown in formula (8): ； In step S1.10, the loss function is calculated as shown in formula (9): ； In step S1.11, each set of input and output samples is trained repeatedly for 100 rounds. The optimizer is Adam, the initial learning rate is 0.001, and the learning rate decays by 0.5 every 20 rounds.

6. The method for establishing a transformer fault early warning model according to claim 1, characterized in that, In step S2.2, each 9×9 graph adjacency matrix is ​​converted into a vector of length 81. , Representative gas to gas The unidirectional relationship between them, according to the properties of the graph adjacency matrix, when When weighted, Then it is 0. The weight is 0. In step S2.4, the decision tree ID3 algorithm model is as shown in formula (10): Information gain represents the knowledge of features The information gain ratio is defined as the degree to which the uncertainty of the classification is reduced by the amount of information provided. For information gain The information gain g(D, A) of feature A on training dataset D, compared to the empirical entropy H(D) of training set D, is defined as the difference between the empirical entropy H(D) of set D and the empirical conditional entropy H(D|A) of D given feature A. The probability of classifying any sample data in set D into category k; In step S2.5, the evaluation index is calculated, including the predicted category and the true category. , , , : TP represents the number of positive classifications correctly predicted by the classifier, TN represents the number of negative classifications correctly predicted by the classifier, FN represents the number of positive classifications incorrectly predicted by the classifier, and FP represents the number of negative classifications incorrectly predicted by the classifier.

7. A method for early warning of transformer faults, characterized in that, The method for establishing a transformer fault early warning model according to any one of claims 1-6 further includes the following steps: Step S3: Receive the latest dissolved gas values ​​in the oil of the transformer equipment transmitted back to the server from the sensor, push the code of the transformer equipment in real time, collect historical gas values ​​from multiple past moments, merge them with the latest data, and arrange the dissolved gas values ​​in the oil in order. Step S4: Normalize the input dissolved gas data in the oil, and record the mean and standard deviation of each gas, as well as the normalized result of the last gas data. ; Step S5: By feeding the pre-trained graph neural network model, a real-time graph adjacency matrix is ​​obtained, which maps the one-way relationships between gas nodes. Step S6: Convert the graph adjacency matrix into a vector and put it into the trained decision tree model to obtain the probability of whether a failure will occur in the next time step; Step S7: Combining the predicted value, determine whether a fault will occur at the next moment by using the guideline and the fault probability output by the decision tree model, and determine the operating status of the transformer at the next moment.

8. A transformer fault early warning method according to claim 7, characterized in that, In step S3, historical gas values ​​at 17 time points are collected. There are 9 dissolved gases in the oil, including hydrogen, methane, ethane, ethylene, acetylene, total hydrocarbons, carbon monoxide, carbon dioxide, and oxygen. In step S6, the vector length is 81.

9. A transformer fault early warning device, comprising an input terminal, an output terminal, one or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the programs comprising methods for performing a transformer fault early warning method according to any one of claims 7-8.

Images