Training method and device of transformer full life cycle model, terminal equipment and residual life prediction method of transformer
By initializing the full-life cycle model and generating adversarial network in the transformer, and using alternating iterative training methods to generate full-life cycle data, it solves the problem that it is difficult to obtain full-life cycle monitoring data in the industrial field, achieving higher prediction accuracy and lower data acquisition costs.
Patent Information
- Application Number
- CN202411990401.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-31
- Publication Date
- 2025-04-29
AI Technical Summary
It is difficult to obtain transformer life cycle monitoring data in industrial sites, which has affected the accuracy of existing prediction methods based on artificial intelligence and machine learning.
By initializing the transformer full life cycle model and generating adversarial network, an alternating iterative training method is used to generate full life cycle data using non-full life cycle data, reducing the cost and difficulty of data acquisition.
The problem of obtaining full life cycle monitoring data was successfully solved, reducing the cost and difficulty of data acquisition, while enhancing the generalization ability of the model and improving the accuracy of the remaining life prediction of the transformer.
Smart Images
Figure CN119905162A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of power transformer life prediction, and in particular to a training method, device, terminal equipment and a transformer remaining life prediction method for a transformer full life cycle model. Background Art
[0002] Transformers play an important role in the power transmission and distribution system, and their safe and reliable operation is crucial to the stability of the system. Once a transformer fails, it will not only cause a power outage, but may even cause the collapse of the power system, resulting in immeasurable economic losses. Therefore, it is very important to accurately predict the remaining life of the transformer, so that maintenance can be planned in advance, aging can be delayed, reliability can be improved, unexpected power outages can be avoided, and losses can be reduced. Analysis of dissolved gases in oil is one of the most important analytical factors for predicting the remaining life of the transformer. Under normal circumstances, as the operating time of the transformer increases, the insulating oil and organic insulating materials inside it gradually age and decompose under the action of heat and electricity, and then generate a variety of gases, such as H2, CH4, C2H2, C2H6, etc. Therefore, by detecting the changes in the composition and content of the gas dissolved in the transformer oil and conducting in-depth analysis, the remaining life of the transformer can be effectively predicted.
[0003] At present, many transformer life prediction methods using dissolved gas concentration in transformer oil have emerged at home and abroad, including prediction methods based on statistical and probability models, prediction methods based on artificial intelligence and machine learning, etc. However, there is a slow degradation trend from healthy operation to failure of equipment. For the remaining life prediction method based on deep learning, the most critical point is to learn the autocorrelation degradation characteristics of the equipment during its entire life cycle from the historical monitoring data of the equipment operation, so as to obtain accurate prediction results. However, in actual practice, it is often impossible to obtain the monitoring data of the entire life cycle of the equipment operation, which affects the evaluation accuracy of the existing prediction methods based on artificial intelligence and machine learning. Summary of the invention
[0004] The embodiments of the present invention provide a training method, device, terminal equipment and a method for predicting the remaining life of a transformer for a full life cycle model, which can solve the problem of difficulty in obtaining full life cycle monitoring data in industrial sites and reduce the cost and difficulty of data acquisition.
[0005] An embodiment of the present invention provides a training method for a transformer life cycle model, comprising:
[0006] Initialize a transformer life cycle model and a generative adversarial network; the generative adversarial network includes a generator and a discriminator;
[0007] The transformer life cycle model, generator, and discriminator are alternately and iteratively trained until the adversarial loss reaches a Nash equilibrium, and a trained transformer life cycle model is obtained;
[0008] Among them, in each iterative training, the current transformer life cycle model is trained according to the training samples, and the full life cycle data is generated based on the non-full life cycle data through the trained transformer life cycle model, and the generated full life cycle data is used as the true full life cycle data; a random noise is input into the current generator so that the current generator generates pseudo full life cycle data; the parameters of the current discriminator are optimized according to the true full life cycle data and the pseudo full life cycle data; according to the discrimination result of the optimized discriminator on the pseudo full life cycle data, the adversarial loss is calculated, and the parameters of the current generator are optimized; according to the optimized generator, the pseudo full life cycle data is regenerated, and the training samples for the next iterative training are determined according to the pseudo full life cycle data; the initial training samples are determined according to the non-full life cycle data.
[0009] Furthermore, the initial training samples are determined by the following method:
[0010] Obtain non-full life cycle data; non-full life cycle data refers to acetylene gas concentration data of non-full life cycle;
[0011] Normalizing the non-full life cycle data to obtain pre-processed non-full life cycle data;
[0012] Divide the pre-processed non-full life cycle data according to the preset time window length to generate a number of sample data;
[0013] A number of sample data and the acetylene gas concentration data at the next moment corresponding to each sample data are used as initial training samples.
[0014] Furthermore, the non-full life cycle data is normalized, including:
[0015] The non-full life cycle data is normalized by the following formula:
[0016]
[0017] Among them, x ′ represents the parameter value after normalization, x represents the parameter value before normalization, and x min represents the minimum parameter value in the non-full life cycle data, x max Indicates the maximum parameter value in non-full life cycle data.
[0018] Furthermore, the parameters of the current discriminator are optimized according to the true full life cycle data and the pseudo full life cycle data, including:
[0019] The true full life cycle data and the pseudo full life cycle data are mixed to obtain the life cycle data to be determined and an actual label used to characterize the data type of the life cycle data to be determined; the actual label includes: a true full life cycle data label and a pseudo full life cycle data label;
[0020] The life cycle data to be judged is input into the current discriminator so that the discriminator generates a predicted label for the life cycle data to be judged, the predicted label is compared with the actual label, and the parameters of the discriminator are optimized according to the comparison result.
[0021] Furthermore, based on the discrimination results of the optimized discriminator on the pseudo full life cycle data, the adversarial loss is calculated, and the parameters of the current generator are optimized, including:
[0022] Calculate the adversarial loss function of the generator based on the pseudo full life cycle data and the discrimination results of the pseudo full life cycle data;
[0023] By minimizing the adversarial loss function of the generator, the parameters of the generator are updated to obtain the optimized generator.
[0024] Further, according to the optimized generator, pseudo full life cycle data is regenerated, and training samples for the next iterative training are determined according to the pseudo full life cycle data, including:
[0025] Inputting a random noise into the optimized generator, so that the optimized generator generates updated pseudo full life cycle data according to the random noise, and using the updated pseudo full life cycle data as a new sample label;
[0026] From the updated pseudo full life cycle data, life cycle data of a preset length is intercepted to obtain corresponding non-full life cycle data, and the corresponding non-full life cycle data is used as corresponding new sample data;
[0027] The new sample labels and the corresponding new sample data are used as training samples for the next iterative training.
[0028] Based on the above method embodiment, the present invention provides a corresponding device embodiment, including: an initialization module and a transformer full life cycle model iterative training module;
[0029] An initialization module is used to initialize a transformer full life cycle model and a generative adversarial network; the generative adversarial network includes a generator and a discriminator;
[0030] The transformer life cycle model iterative training module is used to perform alternate iterative training on the transformer life cycle model, generator, and discriminator until the adversarial loss reaches a Nash equilibrium, thereby obtaining a trained transformer life cycle model.
[0031] Among them, in each iterative training, the current transformer life cycle model is trained according to the training samples, and the full life cycle data is generated based on the non-full life cycle data through the trained transformer life cycle model, and the generated full life cycle data is used as the true full life cycle data; a random noise is input into the current generator so that the current generator generates pseudo full life cycle data; the parameters of the current discriminator are optimized according to the true full life cycle data and the pseudo full life cycle data; according to the discrimination result of the optimized discriminator on the pseudo full life cycle data, the adversarial loss is calculated, and the parameters of the current generator are optimized; according to the optimized generator, the pseudo full life cycle data is regenerated, and the training samples for the next iterative training are determined according to the pseudo full life cycle data; the initial training samples are determined according to the non-full life cycle data.
[0032] Based on the above-mentioned method embodiment, the present invention provides a corresponding terminal device embodiment, including: a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, the steps of the training method of the transformer full life cycle model as described in the present invention are implemented.
[0033] An embodiment of the present invention provides a method for predicting the remaining life of a transformer, comprising:
[0034] Obtain acetylene gas concentration data of the transformer to be evaluated during the non-full life cycle;
[0035] Inputting the non-full life cycle acetylene gas concentration data into the transformer full life cycle model, so that the transformer full life cycle model generates corresponding full life cycle acetylene gas concentration data according to the non-full life cycle acetylene gas concentration data; wherein the transformer full life cycle model is obtained according to the transformer full life cycle model training method provided by one embodiment of the present invention;
[0036] Calculate the remaining life of the transformer based on the acetylene gas concentration data throughout its life cycle.
[0037] Furthermore, the remaining life of the transformer is calculated based on the acetylene gas concentration data of the entire life cycle, including:
[0038] The life time at which the acetylene gas concentration data reaches a preset failure threshold in the acetylene gas concentration data of the whole life cycle is used as the total life prediction value;
[0039] Subtract the current service life from the total life prediction value to obtain the remaining life of the transformer.
[0040] Compared with the prior art, the beneficial effects of the embodiment of this solution are:
[0041] The present invention initializes a transformer life cycle model, which is intended to simulate the entire life cycle behavior of the transformer. At the same time, a generative adversarial network is initialized, including a generator and a discriminator; the transformer life cycle model, the generator and the discriminator are alternately iteratively trained until the adversarial loss reaches a Nash equilibrium, and a trained transformer life cycle model is obtained; wherein, during each iterative training, the current transformer life cycle model is trained according to training samples, wherein the initial training samples are determined according to non-full life cycle data, the transformer life cycle model is preliminarily trained, and based on the trained transformer life cycle model, full life cycle data is generated using non-full life cycle data, and the generated full life cycle The data is taken as the true full life cycle data; then, a random noise is input into the current generator so that the current generator generates pseudo full life cycle data, and the parameters of the current discriminator are optimized according to the true full life cycle data and the pseudo full life cycle data, so that the discriminator can more accurately distinguish the real data from the generated data; then, according to the discrimination result of the pseudo full life cycle data by the optimized discriminator, the parameters of the current generator are optimized, the adversarial loss is calculated, and the parameters of the current generator are optimized so that the generator can generate pseudo full life cycle data that is closer and closer to the real data; finally, according to the optimized generator, the pseudo full life cycle data is regenerated, and the training samples for the next iterative training are determined according to the pseudo full life cycle data. When the adversarial loss reaches the Nash equilibrium, that is, the generator and the discriminator reach a stable state, the generated data is close enough to the real full life cycle data. Once the training is completed, the transformer full life cycle model obtained can predict the full life cycle data based on the non-full life cycle data.
[0042] In summary, the present invention successfully solves the problem of difficulty in obtaining full life cycle monitoring data in industrial sites by combining the transformer full life cycle model and the generative adversarial network, reduces the cost and difficulty of data acquisition, and greatly enhances the generalization ability of the model. BRIEF DESCRIPTION OF THE DRAWINGS
[0043] Figure 1 It is a flowchart of a training method for a transformer life cycle model provided by an embodiment of the present invention;
[0044] Figure 2 It is a flowchart of a specific training process of iterative training provided by an embodiment of the present invention;
[0045] Figure 3 It is a structural schematic diagram of a training device for a transformer full life cycle model provided by an embodiment of the present invention;
[0046] Figure 4 It is a flow chart of a method for predicting the remaining life of a transformer provided in one embodiment of the present invention. DETAILED DESCRIPTION
[0047] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0048] like Figure 1 As shown, an embodiment of the present invention provides a training method for a transformer life cycle model, the method comprising at least the following steps:
[0049] Step S11: Initializing a transformer life cycle model and a generative adversarial network; the generative adversarial network includes a generator and a discriminator;
[0050] For step S11, a transformer life cycle model is constructed based on a long short-term memory network (LSTM). LSTM is a neural network architecture that is particularly suitable for processing time series data and can capture long-term dependencies in the data. Next, the transformer life cycle model is initialized, the hierarchical structure of the LSTM network is defined (such as the size of the input layer, hidden layer, and output layer), the initial weight and initial bias of the network are set, and the activation function and loss function are selected.
[0051] At the same time, the Generative Adversarial Network (GAN) is initialized. The Generative Adversarial Network consists of two neural networks, including a generator (G) and a discriminator (D). The goal of the generator is to generate fake data as close to the real data as possible, while the goal of the discriminator is to distinguish whether the input data is real or generated by the generator. In this embodiment, the network of the generator includes an LSTM layer and a fully connected layer, wherein the LSTM layer is mainly used to learn time series features, and the fully connected layer is responsible for mapping the features to data with the same dimension and consistent distribution as the original training data. In addition, the network structure of the discriminator also includes an LSTM layer and a fully connected layer, and its purpose is to discriminate the original sample as 1 as much as possible, and discriminate the sample generated by the generator as 0. Its discrimination process also includes the forward calculation process of the LSTM and the calculation of the fully connected layer, wherein the fully connected layer converts the sequence features extracted by the LSTM into the final discrimination result.
[0052] Step S12: alternately iteratively training the transformer life cycle model, the generator, and the discriminator until the adversarial loss reaches a Nash equilibrium, thereby obtaining a trained transformer life cycle model;
[0053] Among them, in each iterative training, the current transformer life cycle model is trained according to the training samples, and the full life cycle data is generated based on the non-full life cycle data through the trained transformer life cycle model, and the generated full life cycle data is used as the true full life cycle data; a random noise is input into the current generator so that the current generator generates pseudo full life cycle data; the parameters of the current discriminator are optimized according to the true full life cycle data and the pseudo full life cycle data; according to the discrimination result of the optimized discriminator on the pseudo full life cycle data, the adversarial loss is calculated, and the parameters of the current generator are optimized; according to the optimized generator, the pseudo full life cycle data is regenerated, and the training samples for the next iterative training are determined according to the pseudo full life cycle data; the initial training samples are determined according to the non-full life cycle data.
[0054] For step S12, since the transformer life cycle model of the present invention adopts a long short-term memory network (LSTM) model, it can capture long-term dependencies in time series data. However, in order to ensure the accuracy of LSTM model prediction, a large sample size and high-quality data are usually required. In practical applications, it is often difficult to directly obtain sufficiently complete and continuous transformer life cycle data.
[0055] Therefore, the present invention performs alternate iterative training on the transformer life cycle model, generator and discriminator until the adversarial loss reaches Nash equilibrium, and obtains a trained transformer life cycle model. By introducing GAN to make up for the deficiency of transformer full cycle life data, the prediction model and data prediction model generated by LSTM can be used to guide the training of the GAN generator. The GAN generator can complete the data required by LSTM. Through the alternating optimization training mechanism, the accuracy of transformer life prediction is improved.
[0056] The specific training process of iterative training is described in detail below:
[0057] like Figure 2 As shown, the specific training process of iterative training includes the following steps:
[0058] Step S121: training the current transformer life cycle model according to the training samples, and generating life cycle data through the trained transformer life cycle model based on the non-life cycle data, and using the generated life cycle data as true life cycle data; the initial training samples are determined according to the non-life cycle data;
[0059] For step S121, first, the training sample is input into the current transformer life cycle model, and the following features are extracted based on the sample data and sample labels in the current training sample:
[0060] feature RUL =R_LSTM(W Ri ,W Rf ,W Ro ,b Ri ,b Rf ,b Ro ;X fake )
[0061] Among them, feature RUL represents the extracted features of the current transformer life cycle model, W Ri , W Rf and W Ro They represent the weights of the input gate, forget gate, and output gate of the LSTM layer in the current transformer life cycle model, respectively. Ri , b Rf and b Ro They represent the bias of the input gate, forget gate, and output gate of the LSTM layer in the current transformer life cycle model, respectively. fake Represents the sample data in the current training sample.
[0062] The features extracted by the LSTM layer RULAs the input of the fully connected layer, the predicted full life cycle data is obtained. The specific formula is as follows:
[0063] y RUL =σ(W R feature RUL +b R )
[0064] Among them, y RUL represents the predicted life cycle data, σ represents the nonlinear activation function, W R represents the comprehensive weight of the LSTM layer in the current transformer life cycle model, b R Represents the comprehensive bias of the LSTM layer in the current transformer life cycle model.
[0065] After obtaining the predicted full life cycle data of each sample data in the current training sample, the loss function is calculated according to the sample label corresponding to each sample data in the current training sample, and the gradient descent method is used to optimize the current transformer full life cycle model to obtain the optimized transformer full life cycle model.
[0066] It should be noted that in the initial training stage, due to the possible lack of complete life cycle data, it is preferred to determine the initial training samples in the following way:
[0067] Obtain non-full life cycle data; non-full life cycle data refers to acetylene gas concentration data of non-full life cycle;
[0068] Normalizing the non-full life cycle data to obtain pre-processed non-full life cycle data;
[0069] Divide the pre-processed non-full life cycle data according to the preset time window length to generate a number of sample data;
[0070] A number of sample data and the acetylene gas concentration data at the next moment corresponding to each sample data are used as initial training samples.
[0071] Specifically, during the initial training, non-full life cycle data of acetylene gas concentration is obtained from data sources such as sensor monitoring or historical records. Although these data are incomplete, they are sufficient for the initial training of the model. Then, in order to improve the training efficiency and prediction accuracy of the model, it is usually necessary to normalize the data. Preferably, the non-full life cycle data is normalized, including:
[0072] The non-full life cycle data is normalized by the following formula:
[0073]
[0074] Among them, x ′ represents the parameter value after normalization, x represents the parameter value before normalization, and x min represents the minimum parameter value in the non-full life cycle data, x max Represents the maximum parameter value in the non-full life cycle data. Scaling the original data to between 0 and 1 through normalization helps the model better learn the distribution of the data.
[0075] Next, the normalized data is divided into several sample data according to the preset time window length. Each sample data contains the acetylene gas concentration data within the continuous time window and the acetylene gas concentration data at the next moment after the time window as the prediction target, and the divided sample data and its corresponding prediction target are used as the initial training samples:
[0076]
[0077] Among them, traindata represents the initial sample data, trainlabel represents the initial sample label, and x l represents the acetylene gas concentration data at the lth moment, l represents the time window length, n represents the number of training samples after division, T represents the time series length of the sample data, and T=n+l.
[0078] Next, extract the features from the initial sample data traindata:
[0079] feature=LSTM(W i ,W f ,W o ,b i ,b f ,b o ;traindata)
[0080] Among them, W i , W f and W o Represent the weights of the input gate, forget gate, and output gate, respectively, and b i , b f and b o Represent the bias of the input gate, forget gate and output gate respectively.
[0081] The predicted result of acetylene gas concentration of the transformer at a certain moment in the future is calculated layer by layer. According to the difference between the predicted result and the true label, the loss function (such as mean square error MSE) is calculated. Then, the gradient of each weight is calculated layer by layer through the gradient of the loss function and the chain rule, and the weight is updated using an optimizer (such as Adam) to minimize the loss. After repeated iterative training, a rough transformer life cycle model is finally obtained:
[0082] net=P_LSTM(W i ,W f ,W o ,W Dense ,W out ; b i ,b f ,b o ,b Dense ,b out ;traindata)
[0083] Where net represents a rough transformer life cycle model, W Dense represents the weight of the fully connected layer, W out represents the weight of the output layer, b Dense represents the bias of the fully connected layer, b out Represents the bias of the output layer.
[0084] Input the non-full life cycle data into the rough transformer life cycle model to obtain the predicted full life cycle data:
[0085] X ′ =net(W i ,W f ,W o ,W Dense ,W out ; b i ,b f ,b o ,b Dense ,b out ; X)
[0086] Where X represents non-full life cycle data, X ′ Finally, the life cycle data X generated by the rough transformer life cycle model is ′ As true full life cycle data.
[0087] Step S122: inputting a random noise into the current generator so that the current generator generates pseudo full life cycle data;
[0088] For step S122, the network parameters of the generator are fixed, and then a random noise vector is input into the current generator to generate some pseudo life cycle data. These pseudo life cycle data are not real in the strict sense, but they are generated according to the data distribution learned by the generator, so they can simulate the characteristics of real data to some extent.
[0089] Step S123: Optimizing the parameters of the current discriminator according to the true full life cycle data and the pseudo full life cycle data;
[0090] In a preferred embodiment, the parameters of the current discriminator are optimized according to the true full life cycle data and the pseudo full life cycle data, including:
[0091] The true full life cycle data and the pseudo full life cycle data are mixed to obtain the life cycle data to be determined and an actual label used to characterize the data type of the life cycle data to be determined; the actual label includes: a true full life cycle data label and a pseudo full life cycle data label;
[0092] The life cycle data to be judged is input into the current discriminator so that the discriminator generates a predicted label for the life cycle data to be judged, the predicted label is compared with the actual label, and the parameters of the discriminator are optimized according to the comparison result.
[0093] For step S123, the true full life cycle data and the pseudo full life cycle data are mixed to form a new data set, namely the life cycle data to be judged. When mixing the data, an actual label is assigned to each data item to identify its data type. There are two actual labels: a true full life cycle data label, indicating that the data item is real, and a pseudo full life cycle data label, indicating that the data item is generated by the generator. Next, the life cycle data to be judged is input into the current discriminator model. The network structure of the discriminator includes an LSTM layer and a fully connected layer. Its task is to distinguish whether the input data is real or generated by the generator. After receiving the data, the discriminator calculates the predicted label of each data item through its internal network structure and parameters. The predicted label indicates the probability or confidence that the discriminator believes that the data item is real data. Then, the predicted label generated by the discriminator is compared with the actual label, and the loss function is calculated based on the comparison result:
[0094]
[0095] Among them, L g represents the loss function of the discriminator, and d represents the discriminant result of the pseudo full life cycle data. This function measures the difference between the predicted label and the actual label, and is used to guide the optimization of the discriminator parameters to obtain the optimized discriminator.
[0096] Step S124: Calculate the adversarial loss based on the discrimination result of the optimized discriminator on the pseudo full life cycle data, and optimize the parameters of the current generator;
[0097] In a preferred embodiment, according to the discrimination result of the optimized discriminator on the pseudo full life cycle data, the adversarial loss is calculated, and the parameters of the current generator are optimized, including:
[0098] Calculate the adversarial loss function of the generator based on the pseudo full life cycle data and the discrimination results of the pseudo full life cycle data;
[0099] By minimizing the adversarial loss function of the generator, the parameters of the generator are updated to obtain the optimized generator.
[0100] For step S124, the network parameters of the discriminator and the transformer life cycle model are fixed, and the adversarial loss function of the generator is calculated based on the pseudo life cycle data and the discrimination results of the pseudo life cycle data:
[0101]
[0102] Among them, L R Represents the loss function of the generator, i.e., the adversarial loss function. The adversarial loss function measures the realism of the data generated by the generator. It is usually calculated based on the probability value given by the discriminator for the generated data. If the discriminator thinks that the generated data is very realistic (i.e., it gives a higher probability value), the adversarial loss of the generator will be smaller; conversely, if the discriminator can easily distinguish that the generated data is forged, the adversarial loss of the generator will be larger.
[0103] Next, by minimizing the adversarial loss function of the generator, the parameters of the generator are updated so that new data closer to the real data is generated, and the optimized generator is obtained.
[0104] Step S125: regenerate pseudo full life cycle data according to the optimized generator, and determine the training samples for the next iterative training according to the pseudo full life cycle data;
[0105] In a preferred embodiment, regenerating pseudo full life cycle data according to the optimized generator, and determining training samples for the next iterative training according to the pseudo full life cycle data include:
[0106] Inputting a random noise into the optimized generator, so that the optimized generator generates updated pseudo full life cycle data according to the random noise, and using the updated pseudo full life cycle data as a new sample label;
[0107] From the updated pseudo full life cycle data, life cycle data of a preset length is intercepted to obtain corresponding non-full life cycle data, and the corresponding non-full life cycle data is used as corresponding new sample data;
[0108] The new sample labels and the corresponding new sample data are used as training samples for the next iterative training.
[0109] For step S125, a random noise vector is input into the optimized generator, and the generator uses its internal optimized neural network structure to convert the random noise into a new pseudo full life cycle data, and uses several new pseudo full life cycle data as new sample labels. Then, for each new pseudo full life cycle data, a subsequence of a preset length is intercepted from the new pseudo full life cycle data, which represents non-full life cycle data, and the non-full life cycle data is used as new sample data. Finally, the new pseudo full life cycle data (i.e., new sample labels) and the corresponding non-full life cycle data (i.e., new sample data) are combined together to form new training samples. These new training samples will be used for the next iterative training to further adjust the parameters of the transformer full life cycle model, generator, and discriminator, thereby improving the performance and accuracy of the model.
[0110] The iterative training process repeats steps S121 to S125. During the iterative training process, the generator and the discriminator jointly promote the improvement of the performance of the transformer life cycle model through mutual competition and collaboration. Each iteration is a correction and optimization of the previous result, so that the transformer life cycle model can more accurately capture and express the inherent laws of the data.
[0111] When the iterative training reaches a certain stage, the adversarial loss gradually stabilizes and reaches a Nash equilibrium state, which means that a dynamic balance is formed between the generator and the discriminator, and the competition and cooperation between them have reached a relatively stable state. In this state, the data generated by the generator has become so realistic that the discriminator can no longer effectively distinguish the generated data from the real data. At the same time, the transformer life cycle model has also been well trained. This model continuously learns and adjusts during the iteration process, gradually mastering the distribution laws and characteristics of the data, so that it can more accurately predict and simulate the full life cycle behavior of the transformer.
[0112] like Figure 3 As shown, based on the above method embodiment, a corresponding device embodiment is provided;
[0113] An embodiment of the present invention provides a training device for a transformer life cycle model, including: an initialization module and a transformer life cycle model iterative training module;
[0114] An initialization module is used to initialize a transformer full life cycle model and a generative adversarial network; the generative adversarial network includes a generator and a discriminator;
[0115] The transformer life cycle model iterative training module is used to perform alternate iterative training on the transformer life cycle model, generator, and discriminator until the adversarial loss reaches a Nash equilibrium, thereby obtaining a trained transformer life cycle model.
[0116] Among them, in each iterative training, the current transformer life cycle model is trained according to the training samples, and the full life cycle data is generated based on the non-full life cycle data through the trained transformer life cycle model, and the generated full life cycle data is used as the true full life cycle data; a random noise is input into the current generator so that the current generator generates pseudo full life cycle data; the parameters of the current discriminator are optimized according to the true full life cycle data and the pseudo full life cycle data; according to the discrimination result of the optimized discriminator on the pseudo full life cycle data, the adversarial loss is calculated, and the parameters of the current generator are optimized; according to the optimized generator, the pseudo full life cycle data is regenerated, and the training samples for the next iterative training are determined according to the pseudo full life cycle data; the initial training samples are determined according to the non-full life cycle data.
[0117] It can be understood that the above-mentioned device item embodiment corresponds to the method item embodiment of the present invention, which can implement the training method of the transformer full life cycle model provided by any one of the above-mentioned method item embodiments of the present invention.
[0118] It should be noted that the device embodiments described above are merely illustrative, and some or all of the modules may be selected according to actual needs to achieve the purpose of the present embodiment. In addition, in the accompanying drawings of the device embodiments provided by the present invention, the connection relationship between the modules indicates that there is a communication connection between them, which may be specifically implemented as one or more communication buses or signal lines. Those of ordinary skill in the art may understand and implement the present invention without creative work.
[0119] Based on the above-mentioned embodiment of the training method for the transformer full life cycle model, another embodiment of the present invention provides a terminal device, which includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, the training method for the transformer full life cycle model of any embodiment of the present invention is implemented.
[0120] Exemplarily, in this embodiment, the computer program may be divided into one or more modules, which are stored in the memory and executed by the processor to implement the present invention. The one or more module elements may be a series of computer program instruction segments capable of implementing specific functions, which are used to describe the execution process of the computer program in the terminal device.
[0121] The terminal device may be a computing device such as a desktop computer, a notebook, a PDA, a cloud server, etc. The terminal device may include, but is not limited to, a processor and a memory.
[0122] The processor may be a central processing unit (CPU), other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor, etc. The processor is the control center of the terminal device, and uses various interfaces and lines to connect various parts of the entire terminal device.
[0123] like Figure 4 As shown, an embodiment of the present invention provides a method for predicting the remaining life of a transformer, the method comprising at least the following steps:
[0124] Step S21: Acquire acetylene gas concentration data of a transformer to be evaluated during a non-full life cycle;
[0125] For step S21 , acetylene gas concentration data are collected from the transformer to be evaluated. These data may not cover the entire service life cycle of the transformer, that is, they may be recent, local or intermittent measurement data.
[0126] Step S22: inputting the non-full life cycle acetylene gas concentration data into the transformer full life cycle model, so that the transformer full life cycle model generates corresponding full life cycle acetylene gas concentration data according to the non-full life cycle acetylene gas concentration data; wherein the transformer full life cycle model is obtained according to the transformer full life cycle model training method provided by one embodiment of the present invention;
[0127] For step S22, a trained transformer full life cycle model is obtained by the training method of the transformer full life cycle model provided by an embodiment of the present invention, and the non-full life cycle acetylene gas concentration data collected in step S21 is input into the trained transformer full life cycle model. The model will use these data and the law of change of transformer acetylene gas concentration over time that it has learned to generate corresponding full life cycle acetylene gas concentration data.
[0128] Step S23: Calculate the remaining life of the transformer according to the acetylene gas concentration data of the entire life cycle.
[0129] In a preferred embodiment, the remaining life of the transformer is calculated based on the acetylene gas concentration data of the entire life cycle, including:
[0130] The life time at which the acetylene gas concentration data reaches a preset failure threshold in the acetylene gas concentration data of the whole life cycle is used as the total life prediction value;
[0131] Subtract the current service life from the total life prediction value to obtain the remaining life of the transformer.
[0132] For step S23, the life time of the acetylene gas concentration data reaching the preset failure threshold in the acetylene gas concentration data of the whole life cycle is used as the total life prediction value. The failure threshold is a preset critical value of the acetylene gas concentration. When the acetylene gas concentration in the transformer reaches or exceeds this value, it is generally believed that the transformer has approached or reached the end of its service life and there is a high risk of failure. Next, the time from the start of use of the transformer to the current moment is obtained, and the service life at the current moment is subtracted from the total life prediction value to obtain the remaining life of the transformer. The specific expression is as follows:
[0133] rul=predict-RUL t
[0134] Among them, rul represents the remaining life of the transformer, predict represents the total life prediction value, RUL t Remaining life rul of the transformer indicates the length of time the transformer can continue to operate safely under the current operating conditions.
[0135] The above is a preferred embodiment of the present invention. It should be pointed out that a person skilled in the art can make several improvements and modifications without departing from the principle of the present invention. These improvements and modifications are also considered to be within the scope of protection of the present invention.
Claims
1. A training method for a transformer life cycle model, characterized in that: include: Initialize a transformer life cycle model and generate adversarial networks; The generative adversarial network includes a generator and a discriminator; The transformer life cycle model, the generator and the discriminator are alternately and iteratively trained until the adversarial loss reaches a Nash equilibrium, thereby obtaining a trained transformer life cycle model; In each iterative training, the current transformer life cycle model is trained according to the training samples, and the full life cycle data is generated through the trained transformer life cycle model based on the non-full life cycle data, and the generated full life cycle data is used as the true full life cycle data; A random noise is input into the current generator to make the current generator generate pseudo full life cycle data; the parameters of the current discriminator are optimized according to the true full life cycle data and the pseudo full life cycle data; the adversarial loss is calculated according to the discrimination result of the optimized discriminator on the pseudo full life cycle data, and the parameters of the current generator are optimized; the pseudo full life cycle data is regenerated according to the optimized generator, and the training samples for the next iterative training are determined according to the pseudo full life cycle data; the initial training samples are determined according to the non-full life cycle data.
2. The training method for transformer life cycle model according to claim 1, characterized in that: The initial training samples are determined in the following way: Acquire non-full life cycle data; the non-full life cycle data is acetylene gas concentration data of the non-full life cycle; Normalizing the non-full life cycle data to obtain pre-processed non-full life cycle data; Divide the preprocessed non-full life cycle data according to a preset time window length to generate a number of sample data; The plurality of sample data and the acetylene gas concentration data at the next moment corresponding to each sample data are used as initial training samples.
3. The training method for transformer life cycle model according to claim 2, characterized in that: The non-full life cycle data is normalized, including: The non-full life cycle data is normalized by the following formula: Among them, x ′ represents the parameter value after normalization, x represents the parameter value before normalization, and x min represents the minimum parameter value in the non-full life cycle data, x max Indicates the maximum parameter value in non-full life cycle data.
4. The training method for transformer life cycle model according to claim 2, characterized in that: Optimizing the parameters of the current discriminator according to the true full life cycle data and the pseudo full life cycle data, including: Mixing the true full life cycle data and the pseudo full life cycle data to obtain the life cycle data to be identified and an actual label for characterizing the data type of the life cycle data to be identified; the actual label includes: a true full life cycle data label and a pseudo full life cycle data label; The life cycle data to be judged is input into the current discriminator so that the discriminator generates a predicted label for the life cycle data to be judged, the predicted label is compared with the actual label, and the parameters of the discriminator are optimized according to the comparison result.
5. The training method for transformer life cycle model according to claim 4, characterized in that: According to the discrimination results of the optimized discriminator on the pseudo full life cycle data, the adversarial loss is calculated, and the parameters of the current generator are optimized, including: Calculating the adversarial loss function of the generator according to the pseudo full life cycle data and the discrimination result of the pseudo full life cycle data; By minimizing the adversarial loss function of the generator, the parameters of the generator are updated to obtain the optimized generator.
6. The training method for transformer life cycle model according to claim 5, characterized in that: According to the optimized generator, pseudo full life cycle data is regenerated, and training samples for the next iterative training are determined according to the pseudo full life cycle data, including: Inputting a random noise into the optimized generator, so that the optimized generator generates updated pseudo full life cycle data according to the random noise, and using the updated pseudo full life cycle data as a new sample label; From the updated pseudo full life cycle data, life cycle data of a preset length is intercepted to obtain corresponding non-full life cycle data, and the corresponding non-full life cycle data is used as corresponding new sample data; The new sample labels and the corresponding new sample data are used as training samples for the next iterative training.
7. A training device for a transformer life cycle model, characterized in that: include: Initialization module and transformer full life cycle model iterative training module; The initialization module is used to initialize a transformer life cycle model and generate an adversarial network; The generative adversarial network includes a generator and a discriminator; The transformer life cycle model iterative training module is used to perform alternate iterative training on the transformer life cycle model, the generator and the discriminator until the adversarial loss reaches a Nash equilibrium, thereby obtaining a trained transformer life cycle model; In each iterative training, the current transformer life cycle model is trained according to the training samples, and the full life cycle data is generated through the trained transformer life cycle model based on the non-full life cycle data, and the generated full life cycle data is used as the true full life cycle data; A random noise is input into the current generator to make the current generator generate pseudo full life cycle data; the parameters of the current discriminator are optimized according to the true full life cycle data and the pseudo full life cycle data; the adversarial loss is calculated according to the discrimination result of the optimized discriminator on the pseudo full life cycle data, and the parameters of the current generator are optimized; the pseudo full life cycle data is regenerated according to the optimized generator, and the training samples for the next iterative training are determined according to the pseudo full life cycle data; the initial training samples are determined according to the non-full life cycle data.
8. A terminal device, characterized in that: The method comprises a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, the training method for the transformer life cycle model according to any one of claims 1 to 6 is implemented.
9. A method for predicting the remaining life of a transformer, characterized in that: include: Obtain acetylene gas concentration data for the transformer to be evaluated during the non-full life cycle; Inputting the non-full life cycle acetylene gas concentration data into the transformer full life cycle model, so that the transformer full life cycle model generates corresponding full life cycle acetylene gas concentration data according to the non-full life cycle acetylene gas concentration data; wherein the transformer full life cycle model is obtained according to the training method of the transformer full life cycle model according to any one of claims 1 to 6; The remaining life of the transformer is calculated based on the acetylene gas concentration data of the entire life cycle.
10. The method for predicting the remaining life of a transformer according to claim 9, characterized in that: Based on the acetylene gas concentration data of the whole life cycle, the remaining life of the transformer is calculated, including: The life time at which the acetylene gas concentration data reaches a preset failure threshold value in the acetylene gas concentration data of the whole life cycle is used as the total life prediction value; The total life prediction value is subtracted from the current service life to obtain the remaining life of the transformer.