Converter station overvoltage identification method and system based on attention mechanism feature fusion

Abstract

The invention discloses a converter station overvoltage identification method and system based on attention mechanism feature fusion, which relates to the technical field of power system. After preprocessing, overvoltage data in a power grid converter station is input into a trained identification model to output classification identification of overvoltage in the converter station. The training process of the identification model is as follows: obtaining overvoltage sample data in the converter station and preprocessing; performing Fourier transform and one-dimensional convolution on fault phase voltage sample data in turn to extract frequency domain features, and performing Fourier transform, S transform and two-dimensional convolution on zero-sequence voltage sample data in turn to extract time-frequency features; operating frequency domain features and time-frequency features based on the attention mechanism to obtain final weighted fusion features; compressing the final weighted fusion features into a one-dimensional vector form and inputting them into a classification network to identify overvoltage, and outputting classification identification results of overvoltage; the converter station overvoltage identification method and system improve the accuracy of overvoltage identification.

Description

Technical Field

[0001] The present invention relates to the technical field of power systems, and in particular to a method and system for identifying overvoltage in a converter station based on attention mechanism feature fusion. Background Art

[0002] Overvoltage in converter stations is the main factor causing insulation damage to electrical equipment, which will seriously threaten the safe and stable operation of the power grid. Various reasons such as intrusion waves caused by lightning strikes on transmission lines, filter group circuit breaker opening and closing operations, and grounding faults in converter valves will cause overvoltages that damage the insulation of electrical equipment in converter stations. Lightning overvoltages will rise sharply in a short period of time, and the wave head rise time is extremely short; the overvoltage caused by circuit breaker opening and closing operations lasts for a short time and has high-frequency oscillations; grounding faults will cause the fault phase voltage to drop sharply and the non-fault phase power frequency voltage to increase. Studying the generation mechanism of various types of overvoltages and identifying their types will help converter stations take targeted overvoltage suppression measures and provide a scientific basis for overvoltage insulation coordination.

[0003] The features extracted by the traditional overvoltage identification method based on feature extraction and classifier often cannot meet the needs of various complex overvoltage identification, nor can it realize the automatic identification of overvoltage signals. Professionals are still required to analyze and process them. Manual analysis of a large amount of overvoltage data is time-consuming and laborious. In addition, it is difficult to unify the identification standards of overvoltage by different analysts, which can easily lead to misjudgment. Therefore, it is of great significance to study the self-extraction and identification methods of overvoltage signal features in converter stations. Summary of the invention

[0004] Based on the technical problems existing in the background technology, the present invention proposes a converter station overvoltage identification method and system based on attention mechanism feature fusion, which improves the accuracy of overvoltage identification.

[0005] The present invention proposes a converter station overvoltage identification method based on attention mechanism feature fusion, which pre-processes the overvoltage data in the power grid converter station and inputs it into the trained identification model to output the classification identification of the overvoltage in the converter station;

[0006] The training process of the recognition model is as follows:

[0007] Step 1: Obtain overvoltage sample data in the converter station and preprocess it to construct a training set, wherein the voltage sample data includes fault phase voltage sample data and zero-sequence voltage sample data;

[0008] Step 2: The fault phase voltage sample data is sequentially subjected to Fourier transform and one-dimensional convolution to extract the frequency domain characteristics of the fault phase voltage, and the zero-sequence voltage sample data is sequentially subjected to Fourier transform, S transform and two-dimensional convolution to extract the time-frequency characteristics of the zero-sequence voltage;

[0009] Step 3: Calculate the channel weighted fusion features of the fault phase voltage frequency domain features based on the channel attention mechanism , based on the spatial attention mechanism to calculate the spatial weighted fusion features of zero-sequence voltage time-frequency features , weighted fusion features of channels And spatial weighted fusion features Add the average value to get the final weighted fusion feature ;

[0010] Step 4: Final weighted fusion features After being compressed into a one-dimensional vector form, it is input into a classification network for overvoltage identification, and the classification and identification result of the overvoltage is output. A loss function is constructed based on the output classification and identification result to adjust the model parameters in the identification model.

[0011] Furthermore, in step 2, the generation process of the fault phase voltage frequency domain characteristics is as follows:

[0012] ;

[0013] ;

[0014] ;

[0015] Where: N is the time series length of the fault phase voltage sample data, After normalization, The fault phase voltage sample data of a time series, , It represents the frequency domain result after the fault phase voltage sample data is transformed by Fourier transform. is an imaginary number, represents the index of the frequency component after Fourier transformation, , For the The number of one-dimensional convolution kernels of the layer, that is, the number of channels, For the Layer Channels and Layer The one-dimensional convolution kernel weights of channels, Indicates Tier The output vector of channels , as the first The input vector of the layer, For the Tier Channel bias term, symbol represents a one-dimensional convolution, represents the ReLU activation function, represents the pooling function, One-dimensional convolution The output of the layer, is the number of convolution layers of one-dimensional convolution, is the result after pooling operation, As the frequency domain characteristics of the fault phase voltage.

[0016] Furthermore, in step 2, the generation process of the zero-sequence voltage time-frequency characteristics is as follows:

[0017] ;

[0018] ;

[0019] ;

[0020] Wherein: the time series length of the fault phase voltage sample data is equal to the time series length of the zero-sequence voltage sample data, N is the time series length of the fault phase voltage sample data or the time series length of the zero-sequence voltage sample data, After normalization, Time series of zero-sequence voltage sample data, yes The frequency domain result after discrete Fourier transform is: For at the moment The frequency domain result after Fourier transform is: represents the S matrix after S transformation, is the time series corresponding to the S matrix column, is the index of the frequency component corresponding to the row of the S matrix, ; Indicates The number of two-dimensional convolution kernels of the layer, that is, the number of channels, For the Layer and Layer Two-dimensional convolution kernel weights, Indicates Tier The output vector of channels , as the first The input vector of the layer, For the Tier Channel bias term, symbol represents a two-dimensional convolution, represents the ReLU activation function, represents the pooling function, For the two-dimensional convolution The output of the layer, , is the result after pooling operation, As the time-frequency characteristics of zero-sequence voltage.

[0021] Furthermore, in step 3, the channel weighted fusion features The generation process is as follows:

[0022] ;

[0023] ;

[0024] ;

[0025] in, The frequency domain characteristics of the fault phase voltage The initial weight vector obtained by concatenating the global average pooling and maximum pooling operations, represents global average pooling, stands for global average pooling; is the channel attention weight vector, is the weight matrix of the linear transformation, represents the ReLU activation function, It is a Sigmoid function that normalizes the output value to the range of [0,1].

[0026] Furthermore, spatial weighted fusion features The generation process is as follows:

[0027] ;

[0028] ;

[0029] ;

[0030] in, The time-frequency characteristics of the zero-sequence voltage are The weight vector obtained by concatenating the global average pooling and maximum pooling operations, is the weight matrix of the linear transformation, is the spatial attention weight vector; It is the time-frequency characteristic of zero-sequence voltage after weighted fusion.

[0031] The converter station overvoltage identification system based on the attention mechanism feature fusion pre-processes the overvoltage data in the power grid converter station and inputs it into the trained identification model to output the classification identification of the overvoltage in the converter station;

[0032] The training process of the recognition model is as follows:

[0033] Step 1: Obtain overvoltage sample data in the converter station and preprocess it to construct a training set, wherein the voltage sample data includes fault phase voltage sample data and zero-sequence voltage sample data;

[0034] Step 2: The fault phase voltage sample data is sequentially subjected to Fourier transform and one-dimensional convolution to extract the frequency domain characteristics of the fault phase voltage, and the zero-sequence voltage sample data is sequentially subjected to Fourier transform, S transform and two-dimensional convolution to extract the time-frequency characteristics of the zero-sequence voltage;

[0035] Step 3: Calculate the channel weighted fusion features of the fault phase voltage frequency domain features based on the channel attention mechanism , based on the spatial attention mechanism to calculate the spatial weighted fusion features of zero-sequence voltage time-frequency features , weighted fusion features of channels And spatial weighted fusion features Add the average value to get the final weighted fusion feature ;

[0036] Step 4: Final weighted fusion features After being compressed into a one-dimensional vector form, it is input into a classification network for overvoltage identification, and the classification and identification result of the overvoltage is output. A loss function is constructed based on the output classification and identification result to adjust the model parameters in the identification model.

[0037] The advantages of the converter station overvoltage identification method and system based on attention mechanism feature fusion provided by the present invention are: the prediction category is obtained through the identification model. Different from the traditional overvoltage detection classification method, the identification model proposed in this embodiment extracts local features of the fault phase voltage spectrum and the zero-sequence voltage time-frequency matrix respectively, and uses a two-dimensional convolution model to extract features and an attention mechanism feature fusion to enhance the correlation of features. On the basis of avoiding a sharp increase in data dimensions, automatic feature enhancement and weighted fusion processes are performed to achieve correct identification of overvoltage signals, thereby solving the problem of large sample data volume and unclear feature performance, and improving the accuracy of overvoltage identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 It is a schematic diagram of the process of the present invention. DETAILED DESCRIPTION

[0039] Below, the technical solution of the present invention is described in detail through specific embodiments. Many specific details are set forth in the following description to facilitate a full understanding of the present invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific implementation disclosed below.

[0040] like Figure 1As shown, the converter station overvoltage identification method based on attention mechanism feature fusion proposed in the present invention pre-processes the overvoltage data in the power grid converter station and inputs it into the trained recognition model to output the classification identification of the overvoltage in the converter station.

[0041] The training process of the recognition model is as follows:

[0042] Step 1: Obtain overvoltage sample data in the converter station and preprocess it to construct a training set, wherein the voltage sample data includes fault phase voltage sample data and zero-sequence voltage sample data;

[0043] Collect overvoltage sample data in ±1100kV converter station and preprocess: intercept the original sample data from 2 cycles before overvoltage to 4 cycles after overvoltage, and linearly normalize the overvoltage sample data so that the data value is between [0,1]. The normalization calculation formula is:

[0044] ;

[0045] in: Represents the original sample data of overvoltage in the N-point time series, ; represents the normalized data; Indicates the maximum value in the original sample data of overvoltage; Indicates the minimum value in the original overvoltage sample data. The division ratio of the sample data training set, validation set and test set is 8:1:1

[0046] In actual sampling, the monitoring of overvoltage signals generally collects three-phase voltage and zero-sequence voltage values. The traditional overvoltage identification method directly extracts features from the three-phase voltage time series, while the ±1100kV converter station belongs to a neutral point direct grounding system. When a fault occurs in the system, especially a single-phase grounding, the voltage of the non-fault phase to the ground will not rise significantly due to the clamping effect of the neutral point. This leads to the problem of feature redundancy and large amount of calculation in the feature extraction of the non-fault phase voltage.

[0047] In this embodiment, feature extraction is performed for fault phase voltage and zero-sequence voltage. Zero-sequence voltage is the voltage measured when the vector sum of the phase voltages in a three-phase power supply is zero. By analyzing the zero-sequence voltage, the type of overvoltage fault can be determined.

[0048] Step 2: The fault phase voltage sample data is sequentially subjected to Fourier transform and one-dimensional convolution to extract the frequency domain characteristics of the fault phase voltage, and the zero-sequence voltage sample data is sequentially subjected to Fourier transform, S transform and two-dimensional convolution to extract the time-frequency characteristics of the zero-sequence voltage;

[0049] After performing S transformation on the zero-sequence voltage sample data, a two-dimensional complex time-frequency matrix, namely the S matrix, can be obtained. The rows of the S matrix correspond to frequency components, and the columns correspond to time series. The amplitude and phase information can be obtained from the elements of the S matrix to extract the time-frequency characteristics of the zero-sequence voltage.

[0050] Since the overvoltage signal of the converter station is nonlinear and non-stationary, it is difficult to extract effective features from it. Therefore, in order to solve this problem, this embodiment uses a dual-path convolution feature extraction module based on the overvoltage signal to extract frequency domain features of the fault phase voltage and time-frequency features of the zero-sequence voltage in two ways. The convolution layer realizes local perception of the overvoltage data, and the pooling layer realizes downsampling operation, which can reduce the risk of overfitting of the recognition model.

[0051] The convolution models based on one-dimensional convolution and two-dimensional convolution have 2 convolution layers and 2 pooling layers. The size of the one-dimensional convolution kernel is 1×8, and the size of the two-dimensional convolution kernel is 4×4. The pooling layer selects maximum pooling. The size of the pooling layer in the one-dimensional convolution is 1×2, and the size of the pooling layer in the two-dimensional convolution is 2×2. The activation function selects the ReLU function.

[0052] During the forward propagation process, the ReLU function, unlike the Sigmoid function and the tanh function, needs to calculate the exponent and only needs to set the threshold. Therefore, it can greatly speed up the convergence speed. Compared with the tanh function, its convergence speed can be accelerated by 6 times.

[0053] The generation process of the fault phase voltage frequency domain characteristics is as follows:

[0054] ;

[0055] ;

[0056] ;

[0057] Where: N is the time series length of the fault phase voltage sample data, that is, a fault phase voltage sample data has N points, After normalization, The fault phase voltage sample data of a time series, , It represents the frequency domain result after the fault phase voltage sample data is transformed by Fourier transform. is an imaginary number, Represents the index of the frequency component after Fourier transformation, each Corresponding to a specific frequency component, , For the The number of one-dimensional convolution kernels of the layer, that is, the number of channels, For the Layer Channels and Layer The one-dimensional convolution kernel weights of channels, Indicates Tier The output vector of channels , as the first The input vector of the layer, For the Tier Channel bias term, symbol represents a one-dimensional convolution, represents the ReLU activation function, represents the pooling function, One-dimensional convolution The output of the layer, is the number of convolution layers of one-dimensional convolution, is the result after pooling operation, As the frequency domain characteristics of the barrier phase voltage.

[0058] The generation process of zero-sequence voltage time-frequency characteristics is as follows:

[0059] ;

[0060] ;

[0061] ;

[0062] Wherein: the time series length of the fault phase voltage sample data is equal to the time series length of the zero-sequence voltage sample data, N is the time series length of the fault phase voltage sample data or the time series length of the zero-sequence voltage sample data, After normalization, Time series of zero-sequence voltage sample data, yes The frequency domain result after discrete Fourier transform is: For at the moment The frequency domain result after Fourier transform is the discrete sample of the input zero-sequence voltage data. represents the S matrix after S transformation, The time series corresponding to the S matrix column, The index of the frequency component corresponding to the row of the S matrix, ; Indicates The number of two-dimensional convolution kernels of the layer, that is, the number of channels, For the Layer and Layer Two-dimensional convolution kernel weights, Indicates Tier The output vector of channels , as the first The input vector of the layer, For the Tier Channel bias term, symbol represents a two-dimensional convolution, represents the ReLU activation function, represents the pooling function, For the two-dimensional convolution The output of the layer, , is the result after pooling operation, As the time-frequency characteristics of zero-sequence voltage.

[0063] It should be noted that in the process of generating the fault phase voltage frequency domain characteristics and the zero-sequence voltage time-frequency characteristics, the fault phase voltage sample data and the zero-sequence voltage sample data used need to be transformed by Fourier transform. is the result obtained by Fourier transform of the fault phase voltage sample data, The zero-sequence voltage sample data is obtained by Fourier transform, and the calculation process of the two Fourier transforms is the same. The zero-sequence voltage sample data is transformed by Fourier transform and then S transform.

[0064] Among them, the fault phase voltage sample data is consistent with the zero-sequence voltage sample data in Fourier transform, so when the zero-sequence voltage sample data undergoes Fourier transform, in order to facilitate expression and understanding, it is also used As the index of the frequency component after Fourier transform, it is still used in the subsequent S transform As the index of the frequency component of the S matrix, it is obvious that It does not represent a specific frequency component (ie, frequency value) but is only used as an index to indicate a specific frequency component.

[0065] In addition, for , is the zero-sequence voltage sample data Data after Fourier transformation, but it needs to be input when performing S transformation calculation This is because the S transform is a time-frequency analysis. From the perspective of both the time domain and the frequency domain, is a certain frequency, and As an offset, it is used to access the value at each moment in the signal sequence at that frequency.

[0066] Step 3: Calculate the channel weighted fusion features of the fault phase voltage frequency domain features based on the channel attention mechanism , based on the spatial attention mechanism to calculate the spatial weighted fusion features of zero-sequence voltage time-frequency features , weighted fusion features of channels And spatial weighted fusion features Add the average value to get the final weighted fusion feature ;

[0067] The extracted time domain features and frequency domain features are weightedly fused based on the attention mechanism, where the attention mechanism includes channel attention mechanism and spatial attention mechanism. The channel attention mechanism learns the importance of different channels through a fully connected network and enhances the weight of important channel features; the spatial attention mechanism can transform feature data in the spatial dimension and automatically capture important regional features. This method further enhances the feature extraction capability of the recognition model, improves the importance and directionality of the effective feature vector, and enables the model to have better overvoltage recognition and classification performance.

[0068] Channel weighted fusion features The generation process is as follows:

[0069] ;

[0070] ;

[0071] ;

[0072] in, The frequency domain characteristics of the fault phase voltage The initial weight vector obtained by concatenating the global average pooling and maximum pooling operations, represents global average pooling, stands for global average pooling; is the channel attention weight vector, is the weight matrix of the linear transformation, represents the ReLU activation function, It is a Sigmoid function that normalizes the output value to the range of [0,1].

[0073] Spatial weighted fusion features The generation process is as follows:

[0074] ;

[0075] ;

[0076] ;

[0077] in, The time-frequency characteristics of the zero-sequence voltage are The weight vector obtained by concatenating the global average pooling and maximum pooling operations, is the weight matrix of the linear transformation, is the channel attention weight vector; It is the time-frequency characteristic of zero-sequence voltage after weighted fusion, that is, the spatial weighted fusion characteristic.

[0078] Final weighted fusion feature .

[0079] This embodiment uses the weighted fusion of attention mechanism features to extract effective time-frequency feature parameters of overvoltage signals, enhance the importance and directionality of effective feature vectors, and thus achieve classification and identification of overvoltage.

[0080] Step 4: Final weighted fusion features After being compressed into a one-dimensional vector form, it is input into the classification network for overvoltage identification, and the classification and identification results of overvoltage are output. A loss function is constructed based on the output classification and identification results to adjust the model parameters in the identification model.

[0081] The classification network is composed of a fully connected layer and a Softmax classifier, and the final weighted fusion feature After being compressed into a one-dimensional vector by the Flatten function, the classification and recognition results of voltage The details are as follows:

[0082] ;

[0083] in, Indicates the total number of categories of overvoltage signals, For the The exponential form of the categorical features, is the exponential sum of all eigenvectors, is the index of the overvoltage signal, Represents the output of the recognition model Prediction category features, Represents all the category features that the recognition model has learned. The Softmax function squeezes the last few features output between 0 and 1. The closer the result is to 1, the greater the possibility of representing the category feature.

[0084] for As a result, the cross entropy loss function is selected To measure the difference between the learning recognition model and the training data, so that the network converges to a better local minimum:

[0085] ;

[0086] in, F is the predicted category feature output by the recognition model, is the set label, that is, the real category feature, For the The real category features.

[0087] The Adam optimization algorithm is selected for gradient descent. The algorithm adds first-order momentum and second-order momentum. The weights and biases of neurons in each layer are continuously adjusted until the recognition accuracy meets the requirements. Compared with the traditional gradient descent method, the Adam algorithm is adaptive, robust, and resistant to noise.

[0088] This embodiment obtains the prediction category through the recognition model. Different from the traditional overvoltage detection classification method, the recognition model proposed in this embodiment extracts local features of the fault phase voltage spectrum and the zero-sequence voltage time-frequency matrix respectively, and adopts two-dimensional convolution model to extract features and attention mechanism feature fusion to enhance the correlation of features. On the basis of avoiding a sharp increase in data dimensions, automatic feature enhancement and weighted fusion process are performed to realize the correct recognition of overvoltage signals, thereby solving the problem of large sample data volume and unclear feature performance and improving the accuracy of overvoltage recognition.

[0089] In addition, this embodiment can accurately classify and identify overvoltage in converter stations, is highly robust to noise, and has a certain generalization ability. This is mainly because this embodiment simultaneously considers the multi-dimensional features of overvoltage signals and weights and fuses these features together, making the state feature information of overvoltage signals more effective and complete.

[0090] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A converter station overvoltage identification method based on attention mechanism feature fusion, characterized in that: The overvoltage data in the power grid converter station is pre-processed and input into the trained recognition model to output the classification recognition of the overvoltage in the converter station; The training process of the recognition model is as follows: Step 1: Obtain overvoltage sample data in the converter station and preprocess it to construct a training set, wherein the voltage sample data includes fault phase voltage sample data and zero-sequence voltage sample data; Step 2: The fault phase voltage sample data is sequentially subjected to Fourier transform and one-dimensional convolution to extract the frequency domain characteristics of the fault phase voltage, and the zero-sequence voltage sample data is sequentially subjected to Fourier transform, S transform and two-dimensional convolution to extract the zero-sequence voltage time-frequency characteristics; The generation process of the fault phase voltage frequency domain characteristics is as follows: ； ； ； in: is the time series length of the fault phase voltage sample data, After normalization, The fault phase voltage sample data of a time series, , It represents the frequency domain result after the fault phase voltage sample data is transformed by Fourier transform. is an imaginary number, represents the index of the frequency component after Fourier transformation, , For the The number of one-dimensional convolution kernels of the layer, that is, the number of channels, For the Layer Channels and Layer The one-dimensional convolution kernel weights of channels, Indicates Tier The output vector of channels , as the first The input vector of the layer, For the Tier Channel bias term, symbol represents a one-dimensional convolution, represents the ReLU activation function, represents the pooling function, One-dimensional convolution The output of the layer, is the number of convolution layers of one-dimensional convolution, is the result after pooling operation, As the frequency domain characteristics of the fault phase voltage; The generation process of zero-sequence voltage time-frequency characteristics is as follows: ； ； ； Wherein: the time series length of the fault phase voltage sample data is equal to the time series length of the zero-sequence voltage sample data, N is the time series length of the fault phase voltage sample data or the time series length of the zero-sequence voltage sample data, After normalization, Time series of zero-sequence voltage sample data, yes The frequency domain result after discrete Fourier transform is: For at the moment The frequency domain result after Fourier transform is: represents the S matrix after S transformation, is the time series corresponding to the S matrix column, is the index of the frequency component corresponding to the row of the S matrix, ; Indicates The number of two-dimensional convolution kernels of the layer, that is, the number of channels, For the Layer and Layer Two-dimensional convolution kernel weights, Indicates Tier The output vector of channels , as the first The input vector of the layer, For the Tier Channel bias term, symbol represents a two-dimensional convolution, represents the ReLU activation function, represents the pooling function, For the two-dimensional convolution The output of the layer, , is the result after pooling operation, As the time-frequency characteristics of zero-sequence voltage; Step 3: Calculate the channel weighted fusion features of the fault phase voltage frequency domain features based on the channel attention mechanism , based on the spatial attention mechanism to calculate the spatial weighted fusion features of zero-sequence voltage time-frequency features , weighted fusion features of channels And spatial weighted fusion features Add the average value to get the final weighted fusion feature ; Channel weighted fusion features The generation process is as follows: ； ； ； in, The frequency domain characteristics of the fault phase voltage are The initial weight vector obtained by concatenating the global average pooling and maximum pooling operations, represents global average pooling, stands for global average pooling; is the channel attention weight vector, is the weight matrix of the linear transformation, represents the ReLU activation function, Sigmoid function normalizes the output value to the range of [0,1]; Spatial weighted fusion features The generation process is as follows: ； ； ； in, The time-frequency characteristics of the zero-sequence voltage are The weight vector obtained by concatenating the global average pooling and maximum pooling operations, is the weight matrix of the linear transformation, is the spatial attention weight vector; is the time-frequency characteristic of zero-sequence voltage after weighted fusion; Step 4: Final weighted fusion features After being compressed into a one-dimensional vector form, it is input into a classification network for overvoltage identification, and the classification and identification result of the overvoltage is output. A loss function is constructed based on the output classification and identification result to adjust the model parameters in the identification model.

2. The converter station overvoltage identification system based on attention mechanism feature fusion is characterized by: The overvoltage data in the power grid converter station is pre-processed and input into the trained recognition model to output the classification recognition of the overvoltage in the converter station; The training process of the recognition model is as follows: Step 1: Obtain overvoltage sample data in the converter station and preprocess it to construct a training set, wherein the voltage sample data includes fault phase voltage sample data and zero-sequence voltage sample data; Step 2: The fault phase voltage sample data is sequentially subjected to Fourier transform and one-dimensional convolution to extract the frequency domain characteristics of the fault phase voltage, and the zero-sequence voltage sample data is sequentially subjected to Fourier transform, S transform and two-dimensional convolution to extract the zero-sequence voltage time-frequency characteristics; The generation process of the fault phase voltage frequency domain characteristics is as follows: ； ； ； in: is the time series length of the fault phase voltage sample data, After normalization, The fault phase voltage sample data of a time series, , It represents the frequency domain result after the fault phase voltage sample data is transformed by Fourier transform. is an imaginary number, represents the index of the frequency component after Fourier transformation, , For the The number of one-dimensional convolution kernels of the layer, that is, the number of channels, For the Layer Channels and Layer The one-dimensional convolution kernel weights of channels, Indicates Tier The output vector of channels , as the first The input vector of the layer, For the Tier Channel bias term, symbol represents a one-dimensional convolution, represents the ReLU activation function, represents the pooling function, One-dimensional convolution The output of the layer, is the number of convolution layers of one-dimensional convolution, is the result after pooling operation, As the frequency domain characteristics of the fault phase voltage; The generation process of zero-sequence voltage time-frequency characteristics is as follows: ； ； ； Wherein: the time series length of the fault phase voltage sample data is equal to the time series length of the zero-sequence voltage sample data, N is the time series length of the fault phase voltage sample data or the time series length of the zero-sequence voltage sample data, After normalization, Time series of zero-sequence voltage sample data, yes The frequency domain result after discrete Fourier transform is: For at the moment The frequency domain result after Fourier transform is: represents the S matrix after S transformation, is the time series corresponding to the S matrix column, is the index of the frequency component corresponding to the row of the S matrix, ; Indicates The number of two-dimensional convolution kernels of the layer, that is, the number of channels, For the Layer and Layer Two-dimensional convolution kernel weights, Indicates Tier The output vector of channels , as the first The input vector of the layer, For the Tier Channel bias term, symbol represents a two-dimensional convolution, represents the ReLU activation function, represents the pooling function, For the two-dimensional convolution The output of the layer, , is the result after pooling operation, As the time-frequency characteristics of zero-sequence voltage; Step 3: Calculate the channel weighted fusion features of the fault phase voltage frequency domain features based on the channel attention mechanism , based on the spatial attention mechanism to calculate the spatial weighted fusion features of zero-sequence voltage time-frequency features , weighted fusion features of channels And spatial weighted fusion features Add the average value to get the final weighted fusion feature ; Channel weighted fusion features The generation process is as follows: ； ； ； in, The frequency domain characteristics of the fault phase voltage are The initial weight vector obtained by concatenating the global average pooling and maximum pooling operations, represents global average pooling, stands for global average pooling; is the channel attention weight vector, is the weight matrix of the linear transformation, represents the ReLU activation function, Sigmoid function normalizes the output value to the range of [0,1]; Spatial weighted fusion features The generation process is as follows: ； ； ； in, The time-frequency characteristics of the zero-sequence voltage are The weight vector obtained by concatenating the global average pooling and maximum pooling operations, is the weight matrix of the linear transformation, is the spatial attention weight vector; is the time-frequency characteristic of zero-sequence voltage after weighted fusion; Step 4: Final weighted fusion features After being compressed into a one-dimensional vector form, it is input into a classification network for overvoltage identification, and the classification and identification result of the overvoltage is output. A loss function is constructed based on the output classification and identification result to adjust the model parameters in the identification model.

Images