Extra-high voltage converter transformer early warning method and system based on master-slave mechanism-parameter sharing

Abstract

The invention provides an extra-high voltage converter transformer early warning method based on master-slave mechanism-parameter sharing. The method comprises the steps that firstly, a converter transformer-master-slave mechanism structure is constructed, and historical data of to-be-predicted parameters, current moment data of other parameters in a master node converter transformer and dimension reduction coding information of parameters in other slave node converter transformers are obtained to serve as a supervised learning sample set; constructing a seq2seq time sequence prediction model, and training to obtain a parameter sharing high-precision model; obtaining real or simulated abnormal data of the related variable according to the fault category, and training to obtain a fuzzy discrimination model; and finally, according to a prediction result of the related parameters, calculating a kurtosis coefficient of the membership probability, and removing false alarm information to generate corresponding early warning information. According to the method, the edge node network model is constructed, the master-slave mode is adopted, cross-device key information sharing and inter-device key information sharing are achieved, time sequence prediction of device parameters is carried out in combination with the time sequence prediction model, collaborative information between the parameters is effectively fused, and therefore prediction precision is improved. And meanwhile, a fuzzy set of equipment parameters is constructed by introducing a fuzzy theory, so that alarm triggering conditions are enriched, and the flexibility and robustness of the early warning system are enhanced.

Description

Technical Field

[0001] The present invention relates to the field of parameter monitoring and early warning of ultra-high voltage converter transformer equipment, and in particular to an ultra-high voltage converter transformer early warning method and system based on a master-slave mechanism-parameter sharing. Background Art

[0002] In the UHV DC converter station equipment fault warning framework, the role of model prediction is to fit the time series trend of the power equipment parameters on site, so as to obtain the possible trend of the parameter status of the power equipment in the future time period. The prediction results will be used as the data source of the alarm generation layer to generate fault warning information, so the prediction accuracy of the future value of the power equipment parameters will directly affect the accuracy of the early warning system.

[0003] Given the intricate interconnections between the various production equipment in the converter station, they form a criss-crossing hierarchical dependency network. The functional performance of a single device is closely related to the operating status of other surrounding devices, and there are many complex relationship patterns such as collaborative interaction, functional assistance, and performance complementarity. In this context, only considering a single parameter of the device or even the internal comprehensive multi-parameter prediction is not enough to ensure the accuracy that meets the actual scenario.

[0004] Faced with this challenge, traditional prediction methods that only focus on a single parameter of the device or internal comprehensive multiple parameters can no longer meet the stringent requirements for prediction accuracy in actual scenarios. Summary of the invention

[0005] In order to solve the problems of the above-mentioned background technology, the present invention provides a UHV converter transformer early warning method based on a master-slave mechanism-parameter sharing, which deeply mines the potential correlation and regularity between these data, integrates the monitoring data from multiple device nodes (including the main device nodes) in the region, and provides the prediction accuracy of device parameters.

[0006] By realizing data sharing and collaborative prediction among multiple converter equipment nodes, the overall parameter information of the complex relationships between and within the equipment in the converter station is fully utilized to improve the accuracy and robustness of the prediction. At the same time, the fuzzy theory is introduced, and a fuzzy processing scheme is designed based on the different types of faults that may occur in the equipment. The equipment parameters are fuzzified through the membership function, thereby generating a more detailed equipment fault warning signal with multi-state differentiation capabilities, which enhances the effectiveness of the fault warning system.

[0007] In order to achieve the above object, the present invention adopts the following technical solution:

[0008] On the one hand, the present invention provides a UHV converter transformer early warning method based on a master-slave mechanism-parameter sharing, which is characterized in that it includes the following steps:

[0009] Construct a master-slave mechanism structure, divide the converter transformer equipment in the substation into master nodes and slave nodes, the slave nodes collect and process the real-time monitoring data, voiceprint data, and beamformed acoustic imaging data of the local equipment, use the autoencoder to extract key information from the real-time monitoring data, voiceprint data, and beamformed acoustic imaging data, and generate coded information to send to the master node; the master node obtains the coded information from multiple slave nodes and the current time data and historical data of the real-time monitoring inside the master node as the input data set of the time series prediction model;

[0010] Construct and train a seq2seq time series prediction model, and use the input data set for training to obtain a parameter-sharing prediction model;

[0011] According to the fault category, real or simulated abnormal data of the real-time monitoring data of the master node is obtained, and a fuzzy discrimination model is trained to fuzzify the real-time monitoring data of the master node and generate a fault warning signal;

[0012] The prediction results of the prediction model are input into the fuzzy discriminant model, the kurtosis coefficient of the membership probability is calculated, and the corresponding warning information is generated after removing the false alarm information.

[0013] Furthermore, the real-time monitoring data of the master node and the slave node include real-time operating data of winding temperature, dissolved gas in oil, three temperature differences, SF6 gas pressure, core and clamp grounding current, oil temperature, and oil pressure. The real-time monitoring data of the slave node also includes infrared image data.

[0014] Furthermore, the seq2seq time series prediction model consists of an encoder-decoder structure, the encoder consists of four layers of GRU recurrent units, and the decoder consists of five layers of GRU recurrent units. The seq2seq time series prediction model is trained by maximizing the joint probability distribution of the output sequence. The RMSProp optimization algorithm is used in the training process, the learning rate is set to 5e-1, and the Epoch and Batch-Size parameters of the model training are adjustable parameters.

[0015] Furthermore, the prediction result output by the time series prediction model is the prediction result of the real-time monitoring data of the master node.

[0016] Furthermore, the fuzzy discriminant model includes an input layer, a membership function layer, a combination layer, a normalization layer and an output layer. Its training adopts the Adam optimization algorithm, takes the cross entropy loss function as the reconstruction error, and performs negative feedback optimization by comparing the network output and the target output until the optimal weight is obtained.

[0017] Furthermore, the fuzzy discrimination model removes false positive information by calculating the kurtosis coefficient of the membership probability. The calculation formula of the kurtosis coefficient U is:

[0018]

[0019] Among them, μ and σ are the mean and variance of the confidence probability respectively; X is the confidence vector; K is the warning category; the kurtosis coefficient threshold is set to 5. When the calculated value is greater than the threshold, the warning is considered to be a false alarm.

[0020] Furthermore, the simulated abnormal data includes instantaneous fault simulation data of the converter transformer, abnormal fluctuation fault simulation data of the converter transformer and trend fault simulation data of the converter transformer. The data values ​​of the real-time monitoring data of the master node in the normal state are simulated by superimposing unit pulse signals, fluctuation time series and rising or falling sequences to simulate instantaneous fault data, abnormal fluctuation fault data and trend fault data.

[0021] Furthermore, the warning information includes equipment state parameters, fault type confidence and prediction time. The warning information corresponds to the equipment state parameters, fault type confidence and prediction time through output information header values ​​0, 1, and 2 respectively.

[0022] On the other hand, the present invention also provides an ultra-high voltage converter transformer early warning system based on the above method, including a master node, a slave node, a timing prediction model management module and a fuzzy discrimination model management module. The master node and the slave node realize early warning of the converter transformer equipment through data sharing and collaborative prediction.

[0023] At the same time, the present invention also provides a computer-readable storage medium, on which a computer program is stored, and the program implements the above method when executed by a processor.

[0024] The technical effects of the present invention are as follows:

[0025] By modeling the equipment in the converter station area and using multiple slave nodes for on-site modeling, a master-slave mechanism that can share data is formed. The slave node uses an autoencoder to extract key equipment information and share it with the master node. The master node combines the timing prediction model and the information provided by other nodes to accurately predict the equipment parameters, thereby giving an early warning before the fault occurs. At the same time, the membership function is used to process the equipment parameter prediction results, and the predicted values ​​are converted into fault alarm levels, and a fuzzy model of the equipment status is constructed. Experiments show that the accuracy of the prediction model is 6.67% and 3.14% higher than that of a single parameter and multiple internal parameters, respectively. Referring to the historical logs of the on-site equipment, the fuzzy model is applied to the equipment early warning, and the alarm information can be known 4 days in advance. At the same time, the average recognition accuracy of the constructed test set reached 91.8%. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 It is the early warning distribution result in the implementation case.

[0027] Figure 2 It is the block diagram of seq2seq model.

[0028] Figure 3 It is the structure diagram of the fuzzy network model.

[0029] Figure 4 It is a flow chart of the UHV converter transformer early warning method based on the master-slave mechanism-parameter sharing of the present invention. DETAILED DESCRIPTION

[0030] The present invention will be further described below in conjunction with the accompanying drawings and examples. It should be noted that the following detailed description is exemplary and is intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used in the present invention have the same meanings as those generally understood by those of ordinary skill in the art to which the present invention belongs.

[0031] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0032] Example 1: Parameter prediction.

[0033] like Figure 4 As shown, a master-slave mechanism structure is constructed to obtain the historical data of the real-time monitoring data of the master node and the current moment data, and the dimension reduction coding information of the internal parameters of the converters of the other slave nodes as a supervised learning sample set;

[0034] Specifically, the main contents of the master-slave mechanism are:

[0035] The converter transformer equipment in the substation is divided into master nodes and slave nodes. As the remaining equipment except for the equipment to be evaluated for early warning, the slave node is mainly responsible for collecting and processing the real-time monitoring data of local equipment, the collected voiceprint data, the beam-formed acoustic imaging data, and the infrared data. The autoencoder is used to extract the key information of the power equipment in the node at the current moment, compress the original high-dimensional equipment parameter data, generate the encoded information and send it to the master node; the master node is the central node, responsible for obtaining the encoded information from multiple slave nodes for comprehensive prediction, integrating the encoded information of all relevant slave nodes, taking into account the historical data of the real-time monitoring data of the master node and the data value at the current moment to construct the input data set of the time series prediction model. Then, these data are input into the time series prediction model to predict the future trend of equipment parameters.

[0036] As shown in Table 1, the real-time monitoring data of the main node are: winding temperature, dissolved gas in oil, three temperature differences, SF6 gas pressure, core and clamp grounding current and other 13 state parameters, i.e., parameters to be predicted, as shown in Table 1 below.

[0037] Table 1

[0038] parameter symbol Winding temperature A1 H2 gas dissolved in oil A2 Dissolved gas C2H2 in oil A3 Dissolved gas CO in oil A4 Dissolved gas CO2 in oil A5 Acetylene gas dissolved in oil A6 Dissolved gas CH4 in oil A7 Oil temperature A8 Oil pressure A9 SF6 concentration A10 SF6 gas pressure A11 Core ground current A12 Clamp grounding current A13

[0039] For the selected power equipment parameters, all sampling results within the time interval of July 2023 to October 2023 are used as input samples to form a time series. The time series is collected by sensors distributed on site with a period of 5 minutes, with a total of 12,113 items.

[0040] Optionally, in order to better apply data to training and prediction, combined with the characteristics of machine learning algorithms and improve the accuracy and training speed of the model, it is necessary to standardize the data and de-standardize the prediction results. We use the Z-Score standardization method to process the existing device parameter data:

[0041]

[0042] μ' and σ' are the mean and variance of the data, respectively.

[0043] like Figure 2 As shown, the dimension reduction encoding of the slave node parameters. We use the unsupervised compression of the autoencoder to compress the device parameters contained in the slave node, such as voiceprint data, infrared image data, and acoustic imaging data, to extract the key information in the device parameter vector. The compressed data and the remaining monitoring data are called "encoded information" and will be used as data sharing content sent from the slave node to the master node.

[0044] At the converter station site, an autoencoder is deployed inside each edge node. The training of the autoencoder model requires the use of the historical parameter values ​​in the node. NM Represents the Mth encoded data value of the Nth slave node at the current moment.

[0045] x=[C 1,1 ,C 1,2 ,L,C 1,M-2 ,C 1,M-1 ,C 1,M ,C 2,1 ,L,C 2,M-2 ,C 2,M-1 ,C 2,M ,L,C N,1 ,L,C N,M-2 ,C N,M-1 ,C N,M ]

[0046] We set the final dimension reduction encoding length to 99, that is, the output dimension of the autoencoder located at the slave node is 99, and the input is an optional parameter that can be changed according to actual conditions. Using the encoder part of the autoencoder, the node sequence composed of slave node monitoring parameters, voiceprints, acoustic imaging, and infrared images is forward propagated to obtain the output of the encoder as the result of data compression. The encoded information represents the device parameter status inside the slave node at the current moment. Since the autoencoder is a nonlinear model, the encoded information obtained does not have actual physical meaning, but according to the characteristics of the model, most of the original information can be restored from the encoded information, indicating that the encoded information already contains the key information of the input data and can characterize the device status in the slave node. The master node can directly put the encoded information into the prediction model for training and prediction.

[0047] The training of the time series prediction model mentioned above requires the construction of input and output information. The selection of input information directly determines the accuracy of the model prediction and the complexity of the model itself. In order to take into account as many current device status parameters as possible and avoid the model being too complex to train, we combined the idea of ​​edge computing to divide the model input information into the following three parts:

[0048] (1) Historical data of the real-time monitoring data of the master node.

[0049] (2) The master node monitors the current time data in real time.

[0050] (3) The remaining encoded information transmitted from the node.

[0051] The first part is composed of the historical data of the parameters to be predicted for the master node. This part contains the historical data sequence with a window size of 25. The historical data sequence is the most important content of time series prediction. Analysis based on historical data is the means to achieve many traditional time series prediction algorithms. Through this method, the model can automatically learn many characteristics of the sequence, such as periodicity, trend, etc.

[0052] The second part is to explore the relationship between multiple parameters. In order to reduce the complexity of the model, only the real-time monitoring data of the master node at the current moment is considered.

[0053] The third part is the encoding information from other slave nodes. The encoding information is a compression of high-dimensional data, which can represent most of the conditions of all devices inside the node and can be directly spliced ​​into the input data for training and prediction of time series models.

[0054] The learning sample set takes three types of time series and moment data and features as input samples X input =[x t-1 ,L,x t-N ,q1,L,q M ,c 11 ,c 12 ,L,c Dl ], set the window size N = 25, q M The values ​​of other parameters of the master node, M is the number of types of real-time monitoring data of the master node, and c is Dl is the value of the lth dimension of the Dth slave node, and the output label is X output =[x t+1 ,L,x t+m ], set the output window size m=5.

[0055] like Figure 4 As shown, a seq2seq time series prediction model is constructed and trained to obtain a high-precision model with parameter sharing.

[0056] The seq2seq model consists of an encoder-decoder structure. The encoder consists of four layers of GRU recurrent units connected to each other. The input of each GRU layer is the feature sequence of the input sample and the hidden relationship at the previous moment, and the output is the hidden relationship at the current moment:

[0057] h t =f′(x t ,h t-1 )

[0058] Among them, f represents the GRU model, h t is the hidden relation output by the current layer; f′ is the GRU unit; x t is the input feature value. The output of the fifth layer is used as one of the inputs of the decoder.

[0059] The decoder structure also selects five layers of GRU recurrent units to determine the output feature probability by maximizing the joint probability distribution of the output sequence. Its input is the output value y at the previous moment t-1 , the hidden relationship s at the last moment t-1 And the output c of the fifth layer GRU of the encoder.

[0060]

[0061] P(y t |y1,...,y t-1 ,c)=g(y t-1 ,s t ,c)

[0062] s t =g(y t-1 ,s t-1 ,c)

[0063] The training method is as follows: set the input length parameter to 25; the feature dimension parameter is the number of slave nodes * 99 + 15; the output data length is 5; the vector dimensions of input and output are both 1; the error function is the mean absolute error (MAE), and its formula is: m is the number of single batch outputs, y i is the actual parameter value, The prediction value is used; the iterative update method of the weight parameter is the RMSProp optimization algorithm; the learning rate is 5e-1; the Epoch and Batch-Size parameters of the model training are set as adjustable parameters. The stop training condition is determined according to the set loss function, and the weight set of the model is saved when the stop training condition is reached.

[0064] At the same time, in order to fully illustrate the beneficial effects of the master-slave mechanism, four control groups are set up, among which the input samples of control group 1 are the historical data of the data to be predicted; the input samples of control group 2 are the historical data of the data to be predicted and the historical data of the real-time monitoring data of the equipment; the input samples of control group 3 are the historical data of the data to be predicted, the current time data of the real-time monitoring data of the equipment, and the dimension reduction information of the online monitoring data sent by other devices; control group 4 is the input data of the master-slave mechanism prediction model of the present invention. Under the premise that the training parameters and test parameters are consistent, the results are shown in Table 2 below:

[0065] Table 2

[0066] Group No. Training time (hours) MAE Error standard deviation 1 3 9.2％ 0.0185 2 5.9 8.7％ 0.0149 3 3.7 8.1％ 0.0121 4 4.1 4.7％ 0.0097

[0067] It can be found that the prediction algorithm proposed in this paper has better performance in the test data than the other three experimental groups. The standard deviation of the error is kept to the minimum in the four groups, reaching less than 10 per thousand, which reduces the error by a large margin compared with the traditional single variable model and reduces the error by 4.5% compared with the multivariate prediction model. The reason for this is that this method has more information sources than the other two methods. And the training time is within an acceptable range to achieve good accuracy.

[0068] like Figure 4 As shown, according to the fault category, real or simulated abnormal data of relevant variables are obtained and a fuzzy abnormality discrimination model is obtained through training.

[0069] According to the system log and inspection report and GB / T17468-2019 transformer guidelines, the fault categories are divided into: abnormal SF6 gas pressure, abnormal winding temperature, abnormal clamp core current, abnormal body oil level, and abnormal oil color spectrum. The corresponding converter transformer equipment parameters are: SF6 gas pressure, winding temperature, core grounding current, clamp grounding current, oil temperature, oil pressure, dissolved gas in oil, etc. 13 equipment parameters, as shown in Table 3 below.

[0070] Table 3

[0071] Warning Category Corresponding symbols Abnormal SF6 gas pressure A10, A11 Abnormal winding temperature A1, A13 Abnormal current in the clamp core A12, A13 Abnormal oil level in the body A8, A9 Abnormal oil chromatography A2~A7

[0072] Furthermore, the simulated abnormal data includes instantaneous fault simulation data of the converter transformer, abnormal fluctuation fault simulation data of the converter transformer and trend fault simulation data of the converter transformer, that is, the data values ​​of the real-time monitoring data of the main node in the normal state are simulated by superimposing unit pulse signals, fluctuation time series and rising or falling sequences to simulate instantaneous faults, abnormal fluctuation faults and trend faults.

[0073] Transient fault simulation: superimpose a unit pulse signal with 80% amplitude of the original signal on the normal timing signal

[0074] Abnormal fluctuation fault simulation: a normal time series signal is superimposed on a time series with 80% amplitude fluctuation of the original signal within the same time.

[0075] Trend fault simulation: a normal signal is superimposed with a rising or falling sequence with an amplitude from 0 to 80% of the original signal and a slope of 1.5 over a period of time.

[0076] like Figure 3 As shown, the fuzzy network model includes: input layer, membership function layer, combination layer, normalization layer and output layer.

[0077] Input layer: includes 13 neurons corresponding to the 13 device parameters of this method, and each neuron corresponds to one device parameter.

[0078] Membership function layer: m*13 weight elements where i ranges from [1,13]; j ranges from [1,m]; m is set to 5 or 10; each input layer neuron is connected to m weight elements and outputs the intermediate result:

[0079]

[0080] Simplified to

[0081]

[0082] Composition layer: Each neuron node is composed of a combination of neurons in the membership function layer, that is, the output of the combination layer is:

[0083]

[0084] in represent The neuron weight value, f is the tanh activation function, and m is consistent with the m of the membership function layer.

[0085] Normalization layer: Normalizes the values ​​of the constituent layers.

[0086]

[0087] Output layer: Each neuron represents a type of fault, indicating the degree of conformity of the fuzzy neural network input to each fault category.

[0088]

[0089] Among them, ω ij represents the weight of the neurons connecting the normalization layer and the output layer, K represents the warning category, j represents the number of neurons in the normalization layer, which is consistent with the number of neurons in the input layer, and f represents the sigmoid activation function.

[0090]

[0091] The training method is as follows: the Adam optimization algorithm is used to optimize the parameters of the fuzzy network model, and the cross entropy loss function is used as the reconstruction error; by comparing the network output and the target output, negative feedback is used for optimization until the most weighted one is obtained.

[0092] As mentioned above, the kurtosis coefficient U is:

[0093]

[0094] Among them, μ and σ are the mean and variance of the confidence probability respectively; X is the confidence vector; K is the warning category. Set the kurtosis coefficient threshold to 5. When the calculated value is greater than the threshold, the warning is considered to be a false alarm.

[0095] Furthermore, the warning information includes: equipment status parameters, fault type confidence, and prediction time corresponding to the output information header values ​​0, 1, and 2 respectively. The fuzzy network is constructed using pytorch, and the actual production data of the converter station from October 2023 to December 2023 are selected. The parameters corresponding to the fuzzy neural network model are sampled, and a total of 2993 samples are obtained. The training set and the test set are divided into 7:3, and epoch=200 is selected.

[0096] Table 3

[0097] Warning Category Total quantity Correct quantity Accuracy Abnormal SF6 gas pressure 450 412 91.5％ Abnormal winding temperature 550 465 84.7％ Abnormal current in the clamp core 550 518 94.1％ Abnormal oil level in the body 450 382 85.5％ Abnormal oil chromatography 450 387 86.4％

[0098] As shown in Table 3, the correct warning rates of various anomalies are all above 80%. Figure 1 As shown, the probability of missed alarms and false alarms also meets the actual needs. When implementing false alarms, as an auxiliary mechanism, the false alarm filtering of winding temperature is better, and a total of 13 pieces of information are filtered out. With the help of this warning sample data, a certain warning information is tested as follows:

[0099] Table 4

[0100] Status parameters (0) Confidence(1) Warning time (2) Abnormal oil temperature Abnormal oil level in the body (92.4%) 2023-11-25

[0101] As shown in Table 4, the converter transformer issued an early warning on November 25, and the converter transformer equipment had an abnormality on November 29, which means that the early warning information was issued four days in advance.

[0102] A UHV converter transformer early warning system based on the above method comprises a master node, a slave node, a time series prediction model management module and a fuzzy discrimination model management module. The master node and the slave node realize early warning of converter transformer equipment through data sharing and collaborative prediction.

[0103] At the same time, the present invention also provides a computer-readable storage medium, on which a computer program is stored, and the program implements the above method when executed by a processor.

Claims

1. A UHV converter transformer early warning method based on master-slave mechanism-parameter sharing, characterized in that: The following steps are involved: Construct a master-slave mechanism structure, divide the converter transformer equipment in the substation into master nodes and slave nodes, the slave nodes collect and process the real-time monitoring data, voiceprint data, and beamformed acoustic imaging data of the local equipment, use the autoencoder to extract key information from the real-time monitoring data, voiceprint data, and beamformed acoustic imaging data, and generate coded information to send to the master node; the master node obtains the coded information from multiple slave nodes and the current time data and historical data of the real-time monitoring inside the master node as the input data set of the time series prediction model; Construct and train a seq2seq time series prediction model, and use the input data set for training to obtain a parameter-sharing prediction model; According to the fault category, real or simulated abnormal data of the real-time monitoring data of the master node is obtained, and a fuzzy discrimination model is trained to fuzzify the real-time monitoring data of the master node and generate a fault warning signal; The prediction results of the prediction model are input into the fuzzy discriminant model, the kurtosis coefficient of the membership probability is calculated, and the corresponding warning information is generated after removing the false alarm information.

2. The UHV converter transformer early warning method based on master-slave mechanism-parameter sharing according to claim 1 is characterized in that: The real-time monitoring data of the master node and the slave node include real-time operating data of winding temperature, dissolved gas in oil, three temperature differences, SF6 gas pressure, core and clamp grounding current, oil temperature, and oil pressure. The real-time monitoring data of the slave node also includes infrared image data.

3. The UHV converter transformer early warning method based on master-slave mechanism-parameter sharing according to claim 1 is characterized in that: The seq2seq time series prediction model consists of an encoder-decoder structure, the encoder consists of four layers of GRU recurrent units, and the decoder consists of five layers of GRU recurrent units. The seq2seq time series prediction model is trained by maximizing the joint probability distribution of the output sequence. The RMSProp optimization algorithm is used in the training process, the learning rate is set to 5e-1, and the Epoch and Batch-Size parameters of the model training are adjustable parameters.

4. The UHV converter transformer early warning method based on master-slave mechanism-parameter sharing according to claim 1 is characterized in that: The prediction result output by the time series prediction model is the prediction result of the real-time monitoring data of the master node.

5. The UHV converter transformer early warning method based on master-slave mechanism-parameter sharing according to claim 1 is characterized in that: The fuzzy discriminant model includes an input layer, a membership function layer, a combination layer, a normalization layer and an output layer. The training adopts the Adam optimization algorithm, takes the cross entropy loss function as the reconstruction error, and performs negative feedback optimization by comparing the network output and the target output until the optimal weight is obtained.

6. The UHV converter transformer early warning method based on master-slave mechanism-parameter sharing according to claim 5 is characterized in that: The fuzzy discrimination model removes false positive information by calculating the kurtosis coefficient of the membership probability. The calculation formula of the kurtosis coefficient U is: Among them, μ and σ are the mean and variance of the confidence probability respectively; X is the confidence vector; K is the warning category; the kurtosis coefficient threshold is set to 5. When the calculated value is greater than the threshold, the warning is considered to be a false alarm.

7. The method according to claim 1, characterized in that The simulated abnormal data include instantaneous fault simulation data of converter transformer, abnormal fluctuation fault simulation data of converter transformer and trend fault simulation data of converter transformer. The data values ​​of the real-time monitoring data of the master node in the normal state are simulated by superimposing unit pulse signals, fluctuation time series and rising or falling sequences to simulate instantaneous faults, abnormal fluctuation faults and trend faults.

8. The method according to claim 1, characterized in that: The warning information includes equipment status parameters, fault type confidence and prediction time. The warning information corresponds to the equipment status parameters, fault type confidence and prediction time through the output information header values ​​0, 1, and 2 respectively.

9. An ultra-high voltage converter transformer early warning system based on the method described in any one of claims 1 to 8, comprising a master node, a slave node, a timing prediction model management module and a fuzzy discrimination model management module, wherein the master node and the slave node realize early warning of converter transformer equipment through data sharing and collaborative prediction.

10. A computer-readable storage medium having a computer program stored thereon, wherein the program, when executed by a processor, implements the method according to any one of claims 1 to 8.

Images