Power transformer state evaluation method based on multi-scale gating attention full convolutional network

Abstract

The invention relates to a power transformer state evaluation method based on a multi-scale gating attention full convolutional network. Compared with the prior art, the problem that a reliable and comprehensive power transformer state evaluation method is lacked in the aspect of utilizing multi-source sensor data is solved. The method comprises the following steps: acquiring and preprocessing data of a multi-source sensor of the power transformer; constructing a state evaluation data set of the power transformer; constructing a state evaluation model of the power transformer; training a power transformer state evaluation model; monitoring the power transformer in real time; and evaluating the state of the power transformer. According to the invention, an end-to-end power transformer state evaluation model is realized based on the data of the multi-source sensor, the model can perform comprehensive real-time evaluation on the operation state of the power transformer by using the time sequence data of the multi-source sensor, and the reliability of transformer state evaluation is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of power transformer monitoring, and in particular to a power transformer state assessment method based on a multi-scale gated attention fully convolutional network. Background Art

[0002] Power transformers are key components of the power system. Their unexpected failures will seriously affect the stable operation of the power grid and cause huge losses to power companies and users. Therefore, scientific and accurate evaluation of the operating status of substation equipment is of great significance for timely discovery of power outage risks and early maintenance and repair of transformer equipment.

[0003] In recent years, scholars have been exploring methods for evaluating the status of substation equipment, and some mainstream methods have emerged. For example, methods based on dissolved gas analysis (DGA) mostly analyze the gas components and concentrations dissolved in transformer oil to evaluate the status of transformer equipment. There are also methods based on partial discharge (PD) detection technology, thermal analysis technology based on infrared thermal imaging or transformer oil temperature, methods based on vibration analysis, and methods based on frequency response analysis. These methods have certain limitations. They are based only on single data and cannot comprehensively evaluate the operating status of the transformer, which makes it difficult to meet current needs.

[0004] In order to solve the limitation of single data evaluation, some scholars have explored the comprehensive evaluation of substation equipment status based on multiple factors. These methods have considered the impact of multiple factors on substation equipment status evaluation when they were designed. However, there are also certain limitations. It is difficult to ensure the reliability of prediction for multivariate complex data using traditional methods, while deep learning has been proven to be able to learn more complex features and achieve higher prediction accuracy.

[0005] In addition, these methods mostly rely on current data when evaluating, and cannot fully evaluate the operating status of the transformer. As we all know, the operating status of the transformer is not only related to the current data, but also to the data over a period of time, that is, a period of time series data can more accurately reflect the operating status of the transformer.

[0006] Based on the limitations of the above work, we have two problems to solve: 1) comprehensively evaluate the operating status of the transformer based on multi-source sensor data; 2) construct the current moment data and historical data of the past period into a time series data, and evaluate the operating status of the transformer based on the time series data.

[0007] Therefore, based on the above two problems, how to develop transformer condition evaluation technology based on multivariate time series data has become a technical problem that needs to be solved urgently. Summary of the invention

[0008] The purpose of the present invention is to solve the problem in the prior art of lack of a reliable and comprehensive power transformer state assessment method using multi-source sensor data, and to provide a power transformer state assessment method based on a multi-scale gated attention fully convolutional network to solve the above problem.

[0009] In order to achieve the above object, the technical solution of the present invention is as follows:

[0010] A method for evaluating the state of a power transformer based on a multi-scale gated attention fully convolutional network comprises the following steps:

[0011] Acquisition and preprocessing of multi-source sensor data of power transformers: Acquisition and preprocessing of multi-source sensor data of power transformers, including dissolved gas in oil, top oil temperature, core grounding current and power supply load data;

[0012] Construction of power transformer condition evaluation data set: Based on the preprocessed multi-source sensor data, a two-dimensional matrix data set with multiple variables and labels with time series characteristics is constructed as the power transformer condition evaluation data set;

[0013] Construction of power transformer condition assessment model: Construct a power transformer condition assessment model based on a multi-scale gated attention module and a fully convolutional network module;

[0014] Training of power transformer condition assessment model: sending the preprocessed power transformer basic data to the power transformer condition assessment model for training;

[0015] Real-time monitoring of power transformers: Real-time monitoring of power transformers to obtain monitoring data and pre-process them;

[0016] Evaluation of the power transformer status: The preprocessed real-time monitoring data is input into the trained power transformer status evaluation model to obtain the evaluation result of the power transformer status.

[0017] The construction of the power transformer condition evaluation data set includes the following steps:

[0018] The multi-source sensor data of the transformer will be obtained from the data monitoring platform of the substation, namely the dissolved gas in the oil, trace water, core grounding current, top oil temperature, and power supply load data, and the data will be merged according to time to construct a complete data table;

[0019] Preprocess the merged data table, manually analyze multiple variables separately, deal with missing values ​​and outliers for each variable, and then use the min-max normalization method to normalize the data to form a preprocessed data table;

[0020] The preprocessed data table is fused with the transformer operating status data according to the time dimension to construct a power transformer status evaluation dataset for training, in which the last column is the label column and the remaining columns are feature columns.

[0021] The construction of the power transformer condition assessment model comprises the following steps:

[0022] Construct a power transformer condition assessment model, which includes an input layer, a feature extraction layer, a feature fusion layer, and a classification layer;

[0023] Set up the input layer, which is used to read the raw data and reconstruct the data into the data format used by the model;

[0024] Set the feature extraction layer, which consists of a multi-scale gated attention module and a fully convolutional network module.

[0025] Set up a multi-scale gated attention module, which consists of multi-scale convolution, gating mechanism and multi-head self-attention mechanism.

[0026] Multi-scale convolution is set to extract multi-scale time information of different variables. It consists of three one-dimensional convolutions with different convolution kernel sizes. The sizes of the three one-dimensional convolution kernels are set to 1*3, 1*5 and 1*7 respectively. Then, the features extracted by the three one-dimensional convolutions are concatenated to obtain the output of the multi-scale convolution module.

[0027]

[0028] in, Represents the multi-scale features after splicing and fusion, is the transpose of the input data, DWC1d is a one-dimensional depth-wise separable convolution;

[0029] Set the gating mechanism, which consists of a learnable one-dimensional vector, a sigmoid function, and a bitwise multiplication operation, as shown below:

[0030] F gate =σ(Conv1d(F ms ))⊙F ms ,

[0031] in, represents the output feature after gated weighting, σ represents the Sigmoid activation function, and ⊙ is the element-wise multiplication operation;

[0032] Set up a multi-head self-attention mechanism, which consists of feature concatenation and multi-head attention operations.

[0033]

[0034] in, represents the input features of multi-head self-attention, is the linear transformation matrix, are the queries, keys, and values ​​required for self-attention calculation after linear transformation. is a matrix multiplication operation, d is the variable embedding dimension;

[0035] The fully convolutional network module is set up, which consists of three layers of convolution. The first and third layers of convolution are ordinary one-dimensional convolutions, and the second layer is a one-dimensional depth-separable convolution.

[0036] X′ i =Conv1d(DWC1d(Conv1d(X i )))

[0037] f = LN((X′ i ⊙X i )+X i ),

[0038] Among them, X i is the input feature, X′ i It is the output feature after three layers of convolution, and LN represents layer normalization;

[0039] Set the feature fusion layer,

[0040] The feature fusion layer consists of vector concatenation and vector addition operations. The output of the gating mechanism is concatenated with the output of the multi-head self-attention mechanism. The feedforward neural network is set to reconstruct the concatenated features. The reconstructed output features are then added to the output of the full convolutional network module to achieve feature fusion.

[0041] The classification layer is set, and the fused features are input into the MLP network for classification to obtain the status category of the power transformer, thereby realizing the status evaluation of the power transformer.

[0042] The training of the power transformer condition assessment model comprises the following steps:

[0043] The constructed power transformer condition evaluation data set is divided into a training set and a test set, and then the training set and the test set are respectively reconstructed into data formats readable by the model, and the training set data is input into the input layer of the power transformer condition evaluation model;

[0044] The feature extraction layer extracts effective features from the original data. The multi-scale gated attention module and the fully convolutional network module are used to extract the temporal features of the data and the dependencies between variables.

[0045] Multi-scale gated attention module extracts temporal features:

[0046] The multi-scale time information of multiple variables is extracted through multi-scale convolution, and the multi-scale features are obtained through feature concatenation. Then, the multi-scale features are input into the gating mechanism, and the gating mechanism learns a gating vector. By multiplying the gating vector with the multi-scale features, the important time features of each variable are retained and the redundant time scale information is filtered out.

[0047] The global temporal features of the original data are extracted through the multi-head self-attention mechanism, and the multi-scale features are fused with the original data features and input into the multi-head self-attention mechanism to learn and capture the global temporal features of the time series;

[0048] The fully convolutional network module extracts dependencies between variables:

[0049] The original data is input into the fully convolutional network to extract the interactive dependencies between multiple variables;

[0050] The feature fusion layer fuses multiple extracted valid features to obtain all-round features:

[0051] The output features of the gating mechanism are fused with the output features of the multi-head self-attention mechanism to obtain the time features, and the time features are fused with the features output by the full convolutional network module to obtain the omnidirectional features.

[0052] The classification layer performs classification prediction:

[0053] The extracted all-round features are used for classification prediction, and the predicted value and the true value are calculated through the cross entropy loss function to calculate the loss value. Then, the model parameters are updated through Adam and back propagation to reduce the loss value continuously until it stabilizes, and the power transformer state evaluation model is obtained. Subsequently, the model performance is evaluated using the test set; if the model performance meets the preset requirements, it is applied to the power transformer state evaluation; otherwise, the model hyperparameters need to be adjusted according to the evaluation results, and the model is retrained using the training set data until a power transformer state evaluation model that meets the performance requirements is obtained;

[0054] When the time series data from multi-source sensors are input into the power transformer state evaluation model, the model will evaluate the current transformer state based on the data characteristics.

[0055] Beneficial Effects

[0056] Compared with the prior art, the power transformer state assessment method based on a multi-scale gated attention fully convolutional network of the present invention realizes an end-to-end power transformer state assessment model based on data from multi-source sensors. The model can use the time series data from multi-source sensors to perform a comprehensive real-time assessment of the operating state of the power transformer, thereby improving the reliability of the transformer state assessment.

[0057] Experiments show that the proposed method achieves state-of-the-art performance in terms of accuracy of substation status recognition. In addition, extensive experiments are conducted on public multivariate time series datasets to verify the generalizability of the proposed method. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 is a method sequence diagram of the present invention;

[0059] Figure 2 The figure is an overall framework diagram of the method described in the present invention. DETAILED DESCRIPTION

[0060] In order to have a further understanding and recognition of the structural features and the effects achieved by the present invention, a preferred embodiment and accompanying drawings are used for detailed description as follows:

[0061] like Figure 1 As shown, the present invention discloses a method for evaluating the state of a power transformer based on a multi-scale gated attention fully convolutional network, and proposes a method for evaluating the state of a transformer based on a multivariate time series classification model. The model is divided into two parts, a multi-scale gated attention module and a fully convolutional network module. Among them, the designed multi-scale gated attention module first performs multi-scale convolution operations channel by channel through deep separable convolution, thereby capturing the multi-scale time information of different variables, and then screens the effective time patterns of different variables through the gating layer. On the other hand, the multi-scale features are fused with the original features, and the global dependencies of the time series are captured through multi-head self-attention. The multi-scale gated attention module can simultaneously capture the effective time information and global dependencies of different variables; secondly, based on the inspiration of the anti-bottleneck structure, a fully convolutional network module is designed, which first maps the original features to a high-dimensional space through a convolution operation, enriches the feature space, and enhances the feature representation capability; then, the receptive field is enlarged through multi-layer convolution operations to capture the cross-channel interactive dependencies between multiple variables.

[0062] The multi-scale gated attention module can capture the temporal patterns of different variables respectively and retain effective temporal information through the gating mechanism, and capture the global temporal dependencies of the sequence through multi-head self-attention. The fully convolutional network module can effectively learn the cross-channel interactive dependencies between multiple variables.

[0063] The specific steps are as follows:

[0064] The first step is to obtain and preprocess the multi-source sensor data of the power transformer: obtain the multi-source sensor data of the power transformer, including dissolved gas in oil, top oil temperature, core grounding current and power supply load data, and preprocess it.

[0065] The second step is the construction of the power transformer condition evaluation dataset: Based on the preprocessed multi-source sensor data, a two-dimensional matrix dataset with multiple variables and labels with time series characteristics is constructed as the power transformer condition evaluation dataset.

[0066] (1) The multi-source sensor data of the transformer, namely, dissolved gas in oil, trace water, core grounding current, top oil temperature, and power supply load data, are obtained from the data monitoring platform of the substation, and the data are merged according to time to construct a complete data table.

[0067] (2) Preprocess the merged data table, manually analyze multiple variables separately, deal with missing values ​​and outliers for each variable, and then use the min-max normalization method to normalize the data to form a preprocessed data table.

[0068] (3) The preprocessed data table is fused with the transformer operating status data according to the time dimension to construct a power transformer status evaluation dataset for training, in which the last column is the label column and the remaining columns are feature columns.

[0069] The present invention uses multivariate time series data generated by multiple sensors to evaluate the health status of the transformer. Therefore, the substation equipment status evaluation problem based on these multivariate time series data is defined as a classification problem based on multivariate time series. The multivariate time series classification problem is defined as follows: For a given multivariate time series data set X with n samples = {X1, X2, ..., X n}, and label set Y = {Y1,Y2,…,Y n}, where a single sample s represents the sequence length, c represents the number of variables; Y i ∈{1,2,…,m}m represents the number of categories. The goal of the multivariate time series classification task is to learn a multivariate time series X from the input i To Category Y i The mapping relationship of the classifier.

[0070] The third step is to construct a power transformer state assessment model: construct a power transformer state assessment model based on a multi-scale gated attention module and a fully convolutional network module.

[0071] In order to effectively evaluate the status of substation equipment based on multivariate time series data, a multivariate time series classification network that integrates multi-scale gated attention and fully convolutional network is proposed. This method can efficiently capture and utilize the temporal dependencies of multivariate time series data and the complex correlations and interactive dependencies between multiple variables. Figure 2As shown in the figure, the power transformer condition assessment model consists of two modules: multi-scale gated attention (MSGA) and fully convolutional network. Specifically, first, MSGA extracts time patterns of different time scales through multi-scale convolution; then after passing through the gated unit, it can dynamically adjust the importance of features of different time scales according to context information; finally, the multi-scale features are fused with the original features to capture the global dependencies of the time series through multi-head self-attention. Secondly, the fully convolutional network is used to capture the interactive dependencies between multiple variables. Finally, the features extracted by MSGA and FCN are fused to enrich the feature space and improve the classification performance of the model.

[0072] (1) Construct a power transformer condition assessment model, which includes an input layer, a feature extraction layer, a feature fusion layer, and a classification layer.

[0073] (2) Set the input layer, which is used to read the original data and reconstruct the data into the data format used by the model.

[0074] (3) Setting the feature extraction layer, which consists of a multi-scale gated attention module and a fully convolutional network module.

[0075] For multivariate time series data, different variables may have different time patterns. Single-scale feature extraction may not be able to fully learn the information of multivariate time series, and redundant scale information may lead to a decrease in model performance. Therefore, a multi-scale gating mechanism is used to extract the multi-scale features of the sequence, and the gating mechanism is used to adaptively select important scale information and remove redundant time scales. In addition, in order to effectively capture the global time dependency, multi-head self-attention is also introduced. By fusing the two parts, a multi-scale gated attention module is constructed. This module can capture global time information and multi-scale information respectively, and fuse them to improve the model's multi-scale modeling capabilities for multivariate time series.

[0076] In order to make up for the insufficiency of single-scale modeling in fully exploring the temporal regularity of each variable, we designed a multi-scale convolution module to extract multi-scale information of each variable. We use a series of one-dimensional convolutional layers of different sizes to convolve the input features to capture local and global features at different scales. These multi-scale features are then concatenated and fused. In addition, in order to remove redundant time scale information and retain the effective time pattern of each variable, we designed a gating layer to control the information flow of the fused features and dynamically adjust the importance of each time step and feature channel by applying the gating layer after feature fusion. The introduction of the gating mechanism gives the model greater flexibility and helps capture more complex time series patterns.

[0077] A1) Set up a multi-scale gated attention module, which consists of multi-scale convolution, gating mechanism and multi-head self-attention mechanism.

[0078] Multi-scale convolution is set to extract multi-scale time information of different variables. It consists of three one-dimensional convolutions with different convolution kernel sizes. The sizes of the three one-dimensional convolution kernels are set to 1*3, 1*5 and 1*7 respectively. Then, the features extracted by the three one-dimensional convolutions are concatenated to obtain the output of the multi-scale convolution module.

[0079]

[0080] in, Represents the multi-scale features after splicing and fusion, is the transpose of the input data, and DWC1d is a one-dimensional depth-wise separable convolution.

[0081] Depthwise separable convolution (DWConv) is an optimized convolution operation that can perform convolution operations channel by channel. To this end, a one-dimensional depthwise separable convolution (DWConv1d) is implemented, and a multi-scale convolution module is designed based on DWConv1d. In DWConv1d, each input time series variable is convolved by a separate convolution kernel to extract the time scale of a single variable, and there is no interaction between different variables. For the input feature map X, the shape of the convolution kernel is (C, K), where C is the number of input variables and K is the size of the convolution kernel.

[0082] Set the gating mechanism, which consists of a learnable one-dimensional vector, a sigmoid function, and a bitwise multiplication operation, as shown below:

[0083] F gate =σ(Conv1d(F ms ))⊙F ms ,

[0084] in, represents the output feature after gated weighting, σ represents the Sigmoid activation function, and ⊙ is the element-wise multiplication operation;

[0085] Set up a multi-head self-attention mechanism, which consists of feature concatenation and multi-head attention operations.

[0086]

[0087] in, represents the input features of multi-head self-attention, is the linear transformation matrix, are the queries, keys, and values ​​required for self-attention calculation after linear transformation. is a matrix multiplication operation, and d is the variable embedding dimension.

[0088] Long-term features reflect the overall trend, while short-term features represent subtle changes in local areas, both of which are crucial. The multi-scale gating module can extract the local time scale information of the time series, but it cannot capture the global dependency. Therefore, we design a multi-head self-attention based on multi-scale features to make up for the lack of global dependency information.

[0089] A2) Set up a fully convolutional network module, which consists of three layers of convolution. The first and third layers of convolution are ordinary one-dimensional convolutions, and the second layer is a one-dimensional depth-separable convolution.

[0090] X′ i =Conv1d(DWC1d(Conv1d(X i )))

[0091] f = LN((X′ i ⊙X i )+X i ),

[0092] Among them, X i is the input feature, X′ i It is the output feature after three layers of convolution, and LN represents layer normalization.

[0093] In order to effectively capture the interactive dependencies between multiple variables, a fully convolutional network is designed. Because the traditional convolutional structure first compresses the features and then expands them when processing features. This operation brings a problem, that is, important information may be lost when compressing features, resulting in no useful information being learned in the feature extraction stage. Therefore, a fully convolutional feature extraction module is designed to overcome the above shortcomings and effectively capture the dependencies between multiple variables.

[0094] (4) Setting the feature fusion layer,

[0095] The feature fusion layer consists of vector concatenation and vector addition operations. The output of the gating mechanism is concatenated with the output of the multi-head self-attention mechanism. The feedforward neural network is set to reconstruct the concatenated features, and then the reconstructed output features are added to the output of the full convolutional network module to achieve feature fusion.

[0096] (5) Set the classification layer and input the fused features into the MLP network for classification to obtain the status category of the power transformer, thereby realizing the status evaluation of the power transformer.

[0097] The fourth step is the training of the power transformer state assessment model: the preprocessed power transformer basic data is sent to the power transformer state assessment model for training.

[0098] (1) The constructed power transformer condition evaluation dataset is divided into a training set and a test set, and then the training set and the test set are reconstructed into data formats readable by the model, and the training set data is input into the input layer of the power transformer condition evaluation model.

[0099] (2) The feature extraction layer extracts effective features from the original data. The multi-scale gated attention module and the fully convolutional network module are used to extract the temporal features of the data and the dependencies between variables.

[0100] B1) Multi-scale gated attention module extracts temporal features:

[0101] The multi-scale time information of multiple variables is extracted through multi-scale convolution, and the multi-scale features are obtained through feature concatenation. Then, the multi-scale features are input into the gating mechanism, and the gating mechanism learns a gating vector. By multiplying the gating vector with the multi-scale features, the important time features of each variable are retained and the redundant time scale information is filtered out.

[0102] The global temporal features of the original data are extracted through the multi-head self-attention mechanism, and the multi-scale features are fused with the original data features and input into the multi-head self-attention mechanism to learn and capture the global temporal features of the time series;

[0103] B2) The fully convolutional network module extracts dependencies between variables:

[0104] The original data is input into the fully convolutional network to extract the interactive dependencies between multiple variables.

[0105] (3) The feature fusion layer fuses multiple extracted effective features to obtain omnidirectional features:

[0106] The output features of the gating mechanism are fused with the output features of the multi-head self-attention mechanism to obtain the temporal features, and the temporal features are fused with the features output by the full convolutional network module to obtain the omnidirectional features.

[0107] (4) The classification layer performs classification prediction:

[0108] The extracted all-round features are used for classification prediction, and the predicted value and the true value are calculated through the cross entropy loss function to calculate the loss value. Then, the model parameters are updated through Adam and back propagation to reduce the loss value continuously until it stabilizes, and the power transformer state evaluation model is obtained. Subsequently, the model performance is evaluated using the test set; if the model performance meets the preset requirements, it is applied to the power transformer state evaluation; otherwise, the model hyperparameters need to be adjusted according to the evaluation results, and the model is retrained using the training set data until a power transformer state evaluation model that meets the performance requirements is obtained;

[0109] When the time series data from multi-source sensors are input into the power transformer state evaluation model, the model will evaluate the current transformer state based on the data characteristics.

[0110] Step 5: Real-time monitoring of power transformers: Real-time monitoring of power transformers is performed to obtain monitoring data and perform preprocessing.

[0111] Step 6: Evaluation of the power transformer status: Input the preprocessed real-time monitoring data into the trained power transformer status evaluation model to obtain the evaluation result of the power transformer status.

[0112] In order to verify the effectiveness of the method described in the present invention, a comparative experiment with multiple methods was conducted on the test set. Experiments with sequence lengths of 24 and 48 were set up, and the experimental accuracy and F1 value indicators were recorded as shown in Table 1.

[0113] Table 1 Comparison of the method of the present invention with other methods

[0114]

[0115]

[0116] Through comparative experiments, it can be seen that the method of the present invention is significantly superior to other methods in both experimental indicators, that is, the method of the present invention is more accurate and reliable than other methods in power transformer status evaluation.

[0117] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions only describe the principles of the present invention. The present invention may be subject to various changes and improvements without departing from the spirit and scope of the present invention. These changes and improvements fall within the scope of the present invention. The scope of protection claimed by the present invention is defined by the attached claims and their equivalents.

Claims

1. A method for power transformer state assessment based on a multi-scale gated attention fully convolutional network, characterized in that: The following steps are involved: 11) Acquisition and preprocessing of multi-source sensor data of power transformers: Acquisition and preprocessing of multi-source sensor data of power transformers, including dissolved gas in oil, top oil temperature, core grounding current and power supply load data; 12) Construction of power transformer condition evaluation data set: Based on the preprocessed multi-source sensor data, a two-dimensional matrix data set with multiple variables and labels with time series characteristics is constructed as the power transformer condition evaluation data set; 13) Construction of power transformer condition assessment model: A power transformer condition assessment model is constructed based on a multi-scale gated attention module and a fully convolutional network module; 14) Training of power transformer state assessment model: sending the preprocessed power transformer basic data to the power transformer state assessment model for training; 15) Real-time monitoring of power transformers: Real-time monitoring of power transformers to obtain monitoring data and pre-process them; 16) Evaluation of power transformer status: The preprocessed real-time monitoring data is input into the trained power transformer status evaluation model to obtain the evaluation result of the power transformer status.

2. The power transformer state assessment method based on a multi-scale gated attention fully convolutional network according to claim 1 is characterized in that: The construction of the power transformer condition evaluation data set includes the following steps: 21) Obtain the multi-source sensor data of the transformer from the data monitoring platform of the substation, namely the dissolved gas in the oil, trace water, core grounding current, top oil temperature, and power supply load data, and merge the data according to time to construct a complete data table; 22) Preprocess the merged data table, manually analyze multiple variables separately, deal with the missing values ​​and outliers of each variable, and then use the min-max normalization method to normalize the data to form a preprocessed data table; 23) The preprocessed data table is integrated with the transformer operating status data according to the time dimension to construct a power transformer status evaluation data set for training, in which the last column is the label column and the remaining columns are feature columns.

3. The power transformer state assessment method based on a multi-scale gated attention fully convolutional network according to claim 1 is characterized in that: The construction of the power transformer condition assessment model comprises the following steps: 31) Construct a power transformer condition assessment model, which includes an input layer, a feature extraction layer, a feature fusion layer, and a classification layer; 32) Set the input layer, which is used to read the original data and reconstruct the data into the data format used by the model; 33) Set the feature extraction layer, which consists of a multi-scale gated attention module and a fully convolutional network module. 331) Set up a multi-scale gated attention module, which consists of multi-scale convolution, gating mechanism and multi-head self-attention mechanism, Multi-scale convolution is set to extract multi-scale time information of different variables. It consists of three one-dimensional convolutions with different convolution kernel sizes. The sizes of the three one-dimensional convolution kernels are set to 1*3, 1*5 and 1*7 respectively. Then, the features extracted by the three one-dimensional convolutions are concatenated to obtain the output of the multi-scale convolution module. in, Represents the multi-scale features after splicing and fusion, is the transpose of the input data, DWC1d is a one-dimensional depth-wise separable convolution; Set the gating mechanism, which consists of a learnable one-dimensional vector, a sigmoid function, and a bitwise multiplication operation, as shown below: F gate =σ(Conv1d(F ms ))⊙F ms , in, represents the output feature after gated weighting, σ represents the Sigmoid activation function, and ⊙ is the element-wise multiplication operation; Set up a multi-head self-attention mechanism, which consists of feature concatenation and multi-head attention operations. in, represents the input features of multi-head self-attention, is the linear transformation matrix, are the queries, keys, and values ​​required for self-attention calculation after linear transformation. is a matrix multiplication operation, d is the variable embedding dimension; 332) Set up a fully convolutional network module, which consists of three layers of convolution. The first and third layers of convolution are ordinary one-dimensional convolutions, and the second layer is a one-dimensional depth-separable convolution. X′ i =Conv1d(DWC1d(Conv1d(X i ))) f=LN((X′ i ⊙X i )+X i ), Among them, X i is the input feature, X′ i It is the output feature after three layers of convolution, and LN represents layer normalization; 34) Set the feature fusion layer, The feature fusion layer consists of vector concatenation and vector addition operations. The output of the gating mechanism is concatenated with the output of the multi-head self-attention mechanism. The feedforward neural network is set to reconstruct the concatenated features. The reconstructed output features are then added to the output of the full convolutional network module to achieve feature fusion. 35) Set the classification layer, input the fused features into the MLP network for classification, and obtain the status category of the power transformer, thereby realizing the status evaluation of the power transformer.

4. The power transformer state assessment method based on a multi-scale gated attention fully convolutional network according to claim 1 is characterized in that: The training of the power transformer condition assessment model comprises the following steps: 41) dividing the constructed power transformer condition evaluation data set into a training set and a test set, then reconstructing the training set and the test set into data formats readable by the model, and inputting the training set data into the input layer of the power transformer condition evaluation model; 42) The feature extraction layer extracts effective features from the original data. The multi-scale gated attention module and the fully convolutional network module are used to extract the temporal features of the data and the dependencies between variables. 421) Multi-scale gated attention module extracts temporal features: The multi-scale time information of multiple variables is extracted through multi-scale convolution, and the multi-scale features are obtained through feature concatenation. Then, the multi-scale features are input into the gating mechanism, and the gating mechanism learns a gating vector. By multiplying the gating vector with the multi-scale features, the important time features of each variable are retained and the redundant time scale information is filtered out. The global temporal features of the original data are extracted through the multi-head self-attention mechanism, and the multi-scale features are fused with the original data features and input into the multi-head self-attention mechanism to learn and capture the global temporal features of the time series; 422) The fully convolutional network module extracts dependencies between variables: The original data is input into the fully convolutional network to extract the interactive dependencies between multiple variables; 43) The feature fusion layer fuses multiple extracted effective features to obtain all-round features: The output features of the gating mechanism are fused with the output features of the multi-head self-attention mechanism to obtain the time features, and the time features are fused with the features output by the full convolutional network module to obtain the omnidirectional features. 44) The classification layer performs classification prediction: The extracted all-round features are used for classification prediction, and the loss value is calculated by the cross entropy loss function between the predicted value and the true value. Then, the model parameters are updated through Adam and back propagation to reduce the loss value continuously until it stabilizes, and the power transformer state evaluation model is obtained. Subsequently, the model performance is evaluated using the test set; if the model performance meets the preset requirements, it is applied to the power transformer state evaluation; otherwise, the model hyperparameters are adjusted according to the evaluation results, and the model is retrained using the training set data until a power transformer state evaluation model that meets the performance requirements is obtained; When the time series data from multi-source sensors are input into the power transformer state evaluation model, the model evaluates the current transformer state based on the data characteristics.

Images