Feature extraction method, model training method, medium and equipment of synchronous phasor data

Abstract

The application discloses a feature extraction method and model training method for synchronous phasor data, a medium and equipment. The method uses a dynamic convolution kernel to adaptively capture multi-dimensional features of synchronous phasor data, introduces an attention mechanism to highlight key features, and fuses global and local feature information. A dynamic weighted feature loss function is combined for optimization, and finally accurate feature expression is generated. The core advantage is that it can effectively distinguish the feature differences between different channels and highlight key fault information, thereby improving the accuracy and response speed of fault detection, overcoming the limitations of traditional methods in handling multi-dimensional complex features, and providing an innovative solution for efficient fault diagnosis of power systems.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring and fault diagnosis technology for power systems, and in particular to a method for feature extraction of synchronous phasor data, a model training method, a medium, and a device. Background Technology

[0002] In power systems, synchronization phasor data refers to the electrical quantities such as voltage and current collected by the phasor measurement unit (PMU). However, traditional fault feature extraction methods often struggle to effectively capture the characteristic differences and key fault information between different channels when processing multidimensional synchronization phasor data, thus affecting the accuracy and response speed of subsequent fault detection. Summary of the Invention

[0003] Therefore, it is necessary to provide methods, model training methods, media, and equipment to address the problem that traditional fault feature extraction methods often struggle to effectively capture feature differences and key fault information between different channels when processing multidimensional synchronous phasor data, thus affecting the accuracy and response speed of subsequent fault detection.

[0004] A training method for an adaptive feature channel convolutional network, wherein the adaptive feature channel convolutional network includes a dynamically connected convolutional layer, a local feature extraction layer, a global feature fusion layer, and a fully connected layer, wherein the dynamically connected convolutional layer is provided with multiple convolutional kernels of sequentially increasing size, and the method includes:

[0005] Obtain a sample set; wherein the sample set includes multiple samples, and each sample includes multiple sample synchronization phasor data within a preset time period;

[0006] The samples are sequentially input into the adaptive feature channel convolutional network. The fluctuation amplitude of the synchronous phasor data of each sample is calculated through the dynamic convolutional layer. The corresponding convolutional kernel is selected according to the fluctuation amplitude for convolution processing to obtain multiple convolutional feature maps. The fluctuation amplitude is positively correlated with the size of the corresponding selected convolutional kernel.

[0007] The attention mechanism of the local feature extraction layer is used to extract key features from the multiple feature maps to obtain weighted local features;

[0008] The global feature fusion layer captures the global features of the weighted local features, and then fuses the global features with the weighted local features to obtain the fused features.

[0009] The fully connected layer converts the fused features into feature vectors of the synchronous phasor data. The feature vectors of the synchronous phasor data are then input into a preset dynamic weighted feature loss function to calculate the loss value. The adaptive feature channel convolutional network is iteratively adjusted using the loss value until the adaptive feature channel convolutional network converges.

[0010] In one embodiment, the step of calculating the fluctuation amplitude of the synchronization phasor data of each sample in the sample through the dynamic convolutional layer, and selecting the corresponding convolution kernel for convolution processing according to the fluctuation amplitude, includes:

[0011] The fluctuation amplitude of the synchronous phasor data for each sample in the sample is calculated and expressed as:

[0012]

[0013] In the above formula, X represents the fluctuation amplitude of the synchronized phasor data of the i-th sample; T represents the duration of the preset time period; X i (t) represents the synchronization phasor data of the i-th sample at time t; Represents the mean of the synchronized phasor data of the i-th sample within a preset time period;

[0014] The mapping size of the synchronous phasor data for each sample in the sample is calculated based on the fluctuation amplitude and is expressed as follows:

[0015]

[0016] In the above formula, k i k represents the mapping size of the synchronized phasor data for the i-th sample. min k represents the minimum size of all convolutional kernels in the dynamic convolutional layer. max σ represents the maximum size of all convolutional kernels in the dynamic convolutional layer. 2 max σ represents the maximum fluctuation amplitude of all sample synchronized phasor data. 2 min This represents the minimum fluctuation range of the synchronized phasor data for all samples;

[0017] Based on the mapping size, the convolution kernel corresponding to the synchronous phasor data of each sample is selected for convolution processing using the rounding allocation method.

[0018] In one embodiment, the step of calculating the fluctuation amplitude of the synchronization phasor data of each sample in the sample through the dynamic convolutional layer, and selecting the corresponding convolution kernel for convolution processing according to the fluctuation amplitude, includes:

[0019] The fluctuation amplitude of the synchronous phasor data for each sample in the sample is calculated and expressed as:

[0020] A i =max(X i (t))-min(X i (t)), t∈[1,T]

[0021] In the above formula, A i The value represents the fluctuation amplitude of the synchronized phasor data of the i-th sample; T represents the duration of the preset time period; max(X) i (t) represents the maximum value of the synchronized phasor data of the i-th sample within a preset time period; min(X) i (t) represents the minimum value of the synchronized phasor data of the i-th sample within a preset time period;

[0022] The mapping size of the synchronous phasor data for each sample in the sample is calculated based on the fluctuation amplitude and is expressed as follows:

[0023]

[0024] In the above formula, k i k represents the mapping size of the synchronized phasor data for the i-th sample. min k represents the minimum size of all convolutional kernels in the dynamic convolutional layer. max A represents the maximum size of all convolutional kernels in the dynamic convolutional layer. min A represents the minimum fluctuation range of all sample synchronized phasor data. max This represents the maximum fluctuation amplitude of the synchronized phasor data for all samples;

[0025] Based on the mapping size, the convolution kernel corresponding to the synchronous phasor data of each sample is selected for convolution processing using the rounding allocation method.

[0026] In one embodiment, the step of selecting the convolution kernel corresponding to each sample synchronization phasor data for convolution processing based on the mapping size using a rounding allocation method includes:

[0027] Mapping size k i The synchronized phasor data is allocated to a floor(k) size. i The convolution kernel is used for convolution processing;

[0028] If at least two synchronized phasor data are assigned to the same convolution kernel, then the synchronized phasor data whose size difference is not the smallest is assigned to the idle convolution kernel with the smallest corresponding size difference for convolution processing; where the size difference is the absolute value of the difference between the mapping size and the convolution kernel size.

[0029] In one embodiment, the key features are extracted from the multiple feature maps through the attention mechanism of the local feature extraction layer to obtain weighted local features, as shown below:

[0030]

[0031]

[0032] In the above formula, Z Pw (t) represents the weighted local feature, and I represents the total number of synchronous phasor data in each sample. Z represents the fluctuation amplitude of the synchronous phasor data of the i-th sample. i The feature map represents the synchronized phasor data of the i-th sample.

[0033] In one embodiment, the global features of the weighted local features are captured by the global feature fusion layer, and the global features are fused with the weighted local features to obtain fused features, represented as follows:

[0034] Z Pf (t)=Z Pw (t)+W global ·Z Pg

[0035] Z Pg =GlobalAvgPool(Z Pw )

[0036] In the above formula, Z Pf (t) represents the fusion feature, W global Z represents the weight matrix of the global features. Pg The global feature is represented by GlobalAvgPool(), which represents the global average pooling operation.

[0037] In one embodiment, the dynamically weighted feature loss function is expressed as:

[0038]

[0039] In the above formula, L DWFL F′ represents the loss value. Pn F represents the expected eigenvector of the synchronized phasor data in the nth feature channel; Pn This represents the feature vector extracted from the nth feature channel of the synchronized phasor data; w n The weight of the nth feature channel is represented by λ; the regularization coefficient is represented by W. conv The parameter matrix represents the convolution kernel.

[0040] A method for feature extraction of synchronized phasor data, the method comprising:

[0041] Acquire multiple measured synchronous phasor data within a preset time period;

[0042] The multiple measured synchronous phasor data are input into an adaptive feature channel convolutional network to obtain the feature vectors of the multiple measured synchronous phasor data; wherein, the adaptive feature channel convolutional network is trained by the above-described adaptive feature channel convolutional network training method.

[0043] A computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the above-described adaptive feature channel convolutional network training method and the synchronous phasor data feature extraction method.

[0044] A terminal device includes a memory and a processor. The memory stores a computer program, which, when executed by the processor, causes the processor to perform the steps of the above-described adaptive feature channel convolutional network training method and the synchronous phasor data feature extraction method.

[0045] This invention provides a feature extraction method, model training method, medium, and device for synchronization phasor data. It utilizes dynamic convolutional kernels to adaptively capture multidimensional features of synchronization phasor data, introduces an attention mechanism to highlight key features, and fuses global and local feature information. Optimization is achieved through a dynamically weighted feature loss function, ultimately generating accurate feature representations. Its core advantage lies in its ability to effectively distinguish feature differences between different channels and highlight key fault information, thereby improving the accuracy and response speed of fault detection. It overcomes the limitations of traditional methods in handling multidimensional and complex features, providing an innovative solution for efficient fault diagnosis in power systems. Attached Figure Description

[0046] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0047] in:

[0048] Figure 1 A flowchart illustrating the training method for adaptive feature channel convolutional networks;

[0049] Figure 2 A flowchart for feature extraction of synchronized phasor data;

[0050] Figure 3 A flowchart illustrating the feature extraction method for synchronized phasor data;

[0051] Figure 4 This is a structural block diagram of the terminal device. Detailed Implementation

[0052] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0053] The terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or apparatuses.

[0054] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0055] like Figure 1 As shown, Figure 1 This is a flowchart illustrating a training method for an adaptive feature channel convolutional network in one embodiment. The adaptive feature channel convolutional network includes a dynamically connected convolutional layer, a local feature extraction layer, a global feature fusion layer, and a fully connected layer. The dynamically connected layer has multiple convolutional kernels with progressively increasing sizes. The training method for the adaptive feature channel convolutional network in this embodiment includes the following steps:

[0056] S101, Obtain the sample set.

[0057] The sample set comprises multiple samples, each containing synchronized phasor data from multiple samples within a preset time period. The sample set forms the basis for model training and fault feature extraction, providing input data to support algorithm analysis.

[0058] In this embodiment, as Figure 2 As shown, the sample synchronization phasor data within each sample includes voltage, current, frequency, reactive power, and active power, denoted as:

[0059] X=[V(t), I(t), F(t), Q(t), P(t)], t=1, 2, ..., T

[0060] S102, the samples are sequentially input into the adaptive feature channel convolutional network. The fluctuation amplitude of the synchronous phasor data of each sample is calculated through the dynamic convolutional layer, and the corresponding convolution kernel is selected according to the fluctuation amplitude for convolution processing to obtain multiple convolutional feature maps.

[0061] The fluctuation amplitude refers to the intensity of fluctuation in the sample data, which can be measured by statistics such as variance, standard deviation, and the difference between the maximum and minimum values. In this embodiment, the fluctuation amplitude is set to be positively correlated with the size of the selected convolution kernel.

[0062] This is because for synchronous phasor data with large fluctuations, larger convolutional kernels are needed to capture the range of changes and fully extract global contextual features. For example, in synchronous phasors, drastic changes in voltage or current within a specific time period may characterize the overall pattern of fault occurrence. Therefore, after calculating the fluctuation range of the data, the dynamic convolutional layer assigns a convolutional kernel with a larger kernel size for convolution processing. Conversely, for synchronous phasor data with small fluctuations, after calculating the fluctuation range of the data, the dynamic convolutional layer assigns a convolutional kernel with a smaller kernel size for convolution processing. This allows the focus to be on extracting local details and avoids mixing subtle features with irrelevant context.

[0063] By calculating the fluctuation amplitude to select the convolution kernel and dynamically adjusting the convolution operation, the model's adaptability to multi-scale features is enhanced, and the accuracy and diversity of feature extraction are significantly improved.

[0064] S103 extracts key features from multiple feature maps through the attention mechanism of the local feature extraction layer to obtain weighted local features.

[0065] In the local feature extraction layer, the attention mechanism is used to dynamically allocate weights according to the importance of features, enhance key features through weighted summation, and weaken redundant or noisy features to obtain weighted local features, which provides a foundation for subsequent global feature fusion.

[0066] Optional, such as Figure 2 As shown, in S103, the attention mechanism of the local feature extraction layer extracts key features from multiple feature maps to obtain weighted local features, which are represented as follows:

[0067]

[0068] In the above formula, Z Pw (t) represents the weighted local feature, and I represents the total number of synchronous phasor data within each sample. Z represents the fluctuation amplitude of the synchronous phasor data of the i-th sample. i The feature map represents the synchronized phasor data of the i-th sample.

[0069] This step uses the attention mechanism of the local feature extraction layer to weight the feature map, which can effectively extract key features and highlight important information.

[0070] S104 captures the global features of the weighted local features through the global feature fusion layer, and fuses the global features with the weighted local features to obtain the fused features.

[0071] In this step, the global feature fusion layer is used to extract global information of the input features (i.e., the overall representation of all features) and fuse them with local features to form a more comprehensive and unified fused feature, providing more accurate feature input for subsequent tasks (such as classification and detection).

[0072] Optional, such as Figure 2 As shown, in S104, the global features of the weighted local features are captured through a global feature fusion layer. The global features and the weighted local features are then fused to obtain the fused features, which are represented as follows:

[0073] Z Pf (t)=Z Pw (t)+W global ·Z Pg

[0074] Z Pg =GlobalAvgPool(Z Pw )

[0075] In the above formula, Z Pf (t) represents the fused feature, which includes both overall information and retains key details. W global Z represents the weight matrix of global features. Pg Representing global features, GlobalAvgPool() represents a global average pooling operation, which compresses the global information of each channel into a single value.

[0076] This step extracts global features through a global feature fusion layer and fuses them with local features, which enhances the model's comprehensive understanding of the input data.

[0077] S105 converts the fused features into feature vectors of the synchronous phasor data through a fully connected layer. The feature vectors of the synchronous phasor data are then input into a preset dynamic weighted feature loss function to calculate the loss value. The adaptive feature channel convolutional network is iteratively adjusted based on the loss value until the adaptive feature channel convolutional network converges.

[0078] In this step, fully connected layers are used to achieve high-dimensional compression or mapping of features, thereby outputting fixed-length feature vectors. The dynamically weighted feature loss function is then used to dynamically assign weights based on the importance of different feature vectors, thus guiding the model to focus more on learning key features.

[0079] Optionally, the dynamically weighted feature loss function is expressed as:

[0080]

[0081] In the above formula, L DWFL Indicates the loss value; F′ Pn F represents the expected eigenvector of the synchronized phasor data in the nth feature channel; Pn This represents the feature vector extracted from the nth feature channel of the synchronized phasor data; w n The weight of the nth feature channel is represented by λ; the regularization coefficient is represented by W. conv The parameter matrix represents the convolution kernel.

[0082] In the aforementioned dynamically weighted feature loss function, w is dynamically adjusted. n This can enhance the focus on important features while reducing overfitting to secondary features. Regarding the convolution kernel parameter W... conv L2 regularization can suppress overfitting while ensuring the model's generalization ability. λ controls the strength of regularization; a value that is too large will affect feature extraction ability, while a value that is too small may lead to overfitting.

[0083] We calculate the gradient derivative of the loss function with respect to the network parameters (such as convolutional kernel parameters and weights). Then, we use optimization algorithms (such as gradient descent or Adam) to adjust the network parameters based on the gradient, thereby reducing the loss value. By repeatedly performing these steps and iterating multiple times to continuously optimize the network parameters, we eventually bring the loss function value to its minimum, and the network reaches its optimal state.

[0084] The aforementioned feature extraction method for synchronous phasor data adaptively captures the multidimensional features of the data using dynamic convolutional kernels, introduces an attention mechanism to highlight key features, and integrates global and local feature information. It is then optimized using a dynamically weighted feature loss function to ultimately generate accurate feature representations. Its core advantage lies in its ability to effectively distinguish feature differences between different channels and highlight key fault information, thereby improving the accuracy and response speed of fault detection. This overcomes the limitations of traditional methods in handling multidimensional and complex features, providing an innovative solution for efficient fault diagnosis in power systems.

[0085] Optionally, in S102, the fluctuation amplitude of the synchronous phasor data of each sample in the sample is calculated through a dynamic convolutional layer, and the corresponding convolution kernel is selected for convolution processing based on the fluctuation amplitude. This specifically includes the following steps:

[0086] A1. Calculate the fluctuation amplitude of the synchronous phasor data for each sample in the sample, expressed as:

[0087]

[0088] In the above formula, X represents the fluctuation amplitude of the synchronized phasor data of the i-th sample; T represents the duration of the preset time period; X i (t) represents the synchronization phasor data of the i-th sample at time t; Represents the mean of the synchronized phasor data of the i-th sample within a preset time period;

[0089] A2. Calculate the mapping size of the synchronous phasor data for each sample in the sample based on the fluctuation amplitude, expressed as:

[0090]

[0091] In the above formula, k i k represents the mapping size of the synchronized phasor data for the i-th sample. min k represents the minimum size of all convolutional kernels in a dynamic convolutional layer. max σ represents the maximum size of all convolutional kernels in a dynamic convolutional layer. 2 max σ represents the maximum fluctuation amplitude of all sample synchronized phasor data. 2 min Let k represent the minimum fluctuation range of all sample synchronized phasor data; optionally, let k min Take 3, k max Take 9.

[0092] This maps the fluctuation amplitude to a specified range of convolution kernel sizes, ensuring that the convolution kernel can adapt to the degree of variation in different sample features.

[0093] A3. Based on the mapping size, the convolution kernel corresponding to the synchronous phasor data of each sample is selected by the rounding allocation method for convolution processing.

[0094] The fluctuation amplitude calculation and mapping mechanism in this embodiment enables each sample to be assigned a suitable convolution kernel, avoiding the information loss caused by the fixed convolution kernel size in traditional methods.

[0095] Optionally, in S102, the fluctuation amplitude of the synchronous phasor data of each sample in the sample is calculated through a dynamic convolutional layer, and the corresponding convolution kernel is selected for convolution processing based on the fluctuation amplitude. This specifically includes the following steps:

[0096] B1. Calculate the fluctuation amplitude of the synchronous phasor data for each sample in the sample, expressed as:

[0097] A i=max(X i (t))-min(X i (t)), t∈[1,T]

[0098] In the above formula, A i The value represents the fluctuation amplitude of the synchronized phasor data of the i-th sample; T represents the duration of the preset time period; max(X) i (t) represents the maximum value of the synchronized phasor data of the i-th sample within a preset time period; min(X) i (t) represents the minimum value of the synchronized phasor data of the i-th sample within a preset time period.

[0099] B2. Calculate the mapping size of the synchronous phasor data for each sample in the sample based on the fluctuation amplitude, expressed as:

[0100]

[0101] In the above formula, k i k represents the mapping size of the synchronized phasor data for the i-th sample. min k represents the minimum size of all convolutional kernels in a dynamic convolutional layer. max A represents the maximum size of all convolutional kernels in a dynamic convolutional layer. min A represents the minimum fluctuation range of all sample synchronized phasor data. max This represents the maximum fluctuation range of the synchronous phasor data for all samples.

[0102] This step uses linear interpolation to measure the fluctuation amplitude A. i Mapped to the kernel size range [k min k max This ensures that the convolution kernel can adapt to the varying degrees of change in the features of different samples.

[0103] B3. Based on the mapping size, the convolution kernel corresponding to the synchronous phasor data of each sample is selected by the rounding allocation method for convolution processing.

[0104] Similarly, the calculation and mapping mechanism of fluctuation amplitude in this embodiment can also ensure that each sample is assigned a suitable convolution kernel, avoiding the information loss caused by the fixed convolution kernel size in traditional methods.

[0105] Optionally, the convolutional kernel corresponding to each sample's synchronized phasor data is selected for convolution processing based on the mapping size using a rounding allocation method. This specifically includes the following steps:

[0106] C1. Mapping size k i The synchronized phasor data is allocated to a floor(k) size. i The convolution kernel is used for convolution processing.

[0107] C2. If at least two synchronized phasor data are assigned to the same convolution kernel, then the synchronized phasor data whose size difference is not the smallest is assigned to the idle convolution kernel with the smallest corresponding size difference for convolution processing.

[0108] Here, the size difference is the absolute value of the difference between the mapping size and the convolution kernel size.

[0109] For example, assuming the convolution kernel size ranges from [3, 9], the following mapping sizes exist for the synchronization phasor data:

[0110] k1=4.8, k2=4.3, k3=6.7, k4=4.9

[0111] Preliminary allocation phase:

[0112] The integer parts of k1 and k2 are then allocated to the convolution kernel size of 4.

[0113] Rounding down to k3, the result is allocated to a kernel size of 6.

[0114] Rounding down to k4, the result is allocated to a convolution kernel size of 4.

[0115] Repeated adjustment phase:

[0116] Since k1, k2, and k4 are all assigned to a kernel size of 4, they need to be adjusted based on the size difference.

[0117] Step 1: Adjust k1, k2, and k4 based on the size difference:

[0118] Size difference of k1: |4.8 - 4| = 0.8

[0119] k2 dimension difference: |4.3-4|=0.3

[0120] Size difference of k4: |4.9 - 4| = 0.9

[0121] Therefore, k2 is retained in convolution kernel 4, and other data are adjusted.

[0122] Step 2: Adjust k1 and k4 to other idle convolutional kernels based on the size difference:

[0123] Size difference of k1: |4.8-5|=0.2

[0124] Size difference of k4: |4.9-5|=0.1

[0125] Therefore, k4 is retained in convolution kernel 5, and other data are adjusted.

[0126] Step 3: Adjust k1 to other idle convolutional kernels based on the size difference:

[0127] Kernel sizes 3 and 7 are empty; adjustments are made based on the size difference.

[0128] The difference in size k1 is 1: |4.8-3| = 1.8

[0129] Size difference 2 for k1: |4.8-7| = 2.2

[0130] Therefore, k1 is retained at kernel size 3.

[0131] In this embodiment, by adjusting the repeatedly allocated data, the rationality of convolution kernel allocation is ensured, avoiding feature conflicts or loss when multiple synchronized phasor data share the same convolution kernel. Furthermore, after optimizing the allocation process, the convolution kernel is closer to the ideal size, improving feature extraction performance, especially when processing multidimensional synchronized phasor data, making it easier to capture key fault information.

[0132] like Figure 3 As shown, this invention also proposes a feature extraction method for synchronized phasor data, which includes the following steps:

[0133] S301, acquire multiple measured synchronous phasor data within a preset time period.

[0134] That is, multiple synchronous phasor data are collected within a preset time period. These data include measured values ​​of voltage, current, power, frequency, etc., which reflect the operating status of the power system.

[0135] S302, inputs multiple measured synchronous phasor data into an adaptive feature channel convolutional network to obtain feature vectors of multiple measured synchronous phasor data.

[0136] The adaptive feature channel convolutional network is trained using the method described above. This adaptive network processes the input data and outputs feature vectors representing key characteristics of the synchronization phasor data. These feature vectors can be used for subsequent analysis, such as fault detection or classification tasks.

[0137] Figure 4 An internal structural diagram of a terminal device in one embodiment is shown. Figure 4 As shown, the terminal device includes a processor, a memory, and a network interface connected via a system bus. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and may also store a computer program. When executed by the processor, this computer program enables the processor to implement a feature extraction method for synchronization phasor data. The internal memory may also store a computer program, which, when executed by the processor, enables the processor to implement the feature extraction method for synchronization phasor data. Those skilled in the art will understand that… Figure 4The structure shown is merely a block diagram of a portion of the structure related to the solution of this application and does not constitute a limitation on the terminal device to which the solution of this application is applied. A specific terminal device may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0138] A computer-readable storage medium storing a computer program, which, when executed by a processor, performs the following steps: acquiring a sample set; wherein the sample set includes multiple samples, each sample including multiple sample synchronization phasor data within a preset time period; sequentially inputting the samples into an adaptive feature channel convolutional network, calculating the fluctuation amplitude of the synchronization phasor data of each sample through the dynamic convolutional layer, and selecting a corresponding convolutional kernel for convolution processing based on the fluctuation amplitude to obtain multiple convolutional feature maps; wherein the fluctuation amplitude is positively correlated with the size of the corresponding selected convolutional kernel; The attention mechanism of the local feature extraction layer extracts key features from the multiple feature maps to obtain weighted local features. The global feature fusion layer captures the global features of the weighted local features and fuses them with the weighted local features to obtain fused features. The fully connected layer converts the fused features into feature vectors of the synchronous phasor data. The feature vectors of the synchronous phasor data are input into a preset dynamic weighted feature loss function to calculate the loss value. The adaptive feature channel convolutional network is iteratively adjusted using the loss value until the adaptive feature channel convolutional network converges.

[0139] The following steps are implemented: acquiring multiple measured synchronous phasor data within a preset time period; inputting the multiple measured synchronous phasor data into an adaptive feature channel convolutional network to obtain the feature vector of the multiple measured synchronous phasor data; wherein, the adaptive feature channel convolutional network is trained using the above-mentioned training method for adaptive feature channel convolutional networks.

[0140] A terminal device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it performs the following steps: acquiring a sample set; wherein the sample set includes multiple samples, each sample including multiple sample synchronization phasor data within a preset time period; sequentially inputting the samples into an adaptive feature channel convolutional network; calculating the fluctuation amplitude of the synchronization phasor data of each sample through the dynamic convolutional layer; and selecting a corresponding convolutional kernel based on the fluctuation amplitude for convolution processing to obtain multiple convolutional feature maps; wherein the fluctuation amplitude is related to the size of the corresponding selected convolutional kernel. The features are positively correlated; key features are extracted from the multiple feature maps through the attention mechanism of the local feature extraction layer to obtain weighted local features; the global feature fusion layer captures the global features of the weighted local features, and the global features are fused with the weighted local features to obtain fused features; the fused features are converted into feature vectors of synchronous phasor data through the fully connected layer, and the feature vectors of synchronous phasor data are input into a preset dynamic weighted feature loss function to calculate the loss value, and the adaptive feature channel convolutional network is iteratively adjusted through the loss value until the adaptive feature channel convolutional network converges.

[0141] The following steps are implemented: acquiring multiple measured synchronous phasor data within a preset time period; inputting the multiple measured synchronous phasor data into an adaptive feature channel convolutional network to obtain the feature vector of the multiple measured synchronous phasor data; wherein, the adaptive feature channel convolutional network is trained using the above-mentioned training method for adaptive feature channel convolutional networks.

[0142] It should be noted that the above-mentioned feature extraction method, model training method, device, and computer-readable storage medium for synchronized phasor data belong to the same general inventive concept, and the contents of the embodiments of the feature extraction method, model training method, device, and computer-readable storage medium for synchronized phasor data are applicable to each other.

[0143] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. This program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.

[0144] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0145] The above embodiments merely illustrate several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A training method for an adaptive feature channel convolutional network, characterized in that, The adaptive feature channel convolutional network includes a dynamically connected convolutional layer, a local feature extraction layer, a global feature fusion layer, and a fully connected layer. The dynamically connected layer has multiple convolutional kernels with progressively increasing sizes. The method includes: Obtain a sample set; wherein the sample set includes multiple samples, each sample including multiple sample synchronization phasor data within a preset time period, the sample synchronization phasor data including voltage, current, frequency, reactive power, and active power, denoted as: ； The samples are sequentially input into the adaptive feature channel convolutional network. The fluctuation amplitude of the synchronous phasor data of each sample is calculated through the dynamic convolutional layer. The corresponding convolutional kernel is selected according to the fluctuation amplitude for convolution processing to obtain multiple convolutional feature maps. The fluctuation amplitude is positively correlated with the size of the corresponding selected convolutional kernel. The attention mechanism of the local feature extraction layer is used to extract key features from multiple feature maps to obtain weighted local features; The global feature fusion layer captures the global features of the weighted local features, and then fuses the global features with the weighted local features to obtain the fused features. The fused features are converted into feature vectors of synchronous phasor data through the fully connected layer. The feature vectors of synchronous phasor data are input into a preset dynamic weighted feature loss function to calculate the loss value. The adaptive feature channel convolutional network is iteratively adjusted using the loss value until the adaptive feature channel convolutional network converges. The feature vectors of synchronous phasor data are used for fault detection or classification tasks.

2. The method according to claim 1, characterized in that, The step of calculating the fluctuation amplitude of the synchronous phasor data of each sample in the sample through the dynamic convolutional layer, and selecting the corresponding convolution kernel for convolution processing according to the fluctuation amplitude, includes: The fluctuation amplitude of the synchronous phasor data for each sample in the sample is calculated and expressed as: In the above formula, This represents the fluctuation amplitude of the synchronized phasor data of the i-th sample; Indicates the duration of the preset time period; This represents the synchronization phasor data of the i-th sample at time t; Represents the mean of the synchronized phasor data of the i-th sample within a preset time period; The mapping size of the synchronous phasor data for each sample in the sample is calculated based on the fluctuation amplitude and is expressed as follows: In the above formula, This represents the mapping size of the synchronized phasor data for the i-th sample. This represents the minimum size of all convolutional kernels in the dynamic convolutional layer. This represents the maximum size of all convolutional kernels in the dynamic convolutional layer. This represents the maximum fluctuation amplitude of the synchronous phasor data across all samples. This represents the minimum fluctuation range of the synchronized phasor data for all samples; Based on the mapping size, the convolution kernel corresponding to the synchronous phasor data of each sample is selected for convolution processing using the rounding allocation method.

3. The method according to claim 1, characterized in that, The step of calculating the fluctuation amplitude of the synchronous phasor data of each sample in the sample through the dynamic convolutional layer, and selecting the corresponding convolution kernel for convolution processing according to the fluctuation amplitude, includes: The fluctuation amplitude of the synchronous phasor data for each sample in the sample is calculated and expressed as: In the above formula, This represents the fluctuation amplitude of the synchronized phasor data of the i-th sample; Indicates the duration of the preset time period; This represents the maximum value of the synchronized phasor data of the i-th sample within a preset time period; This represents the minimum value of the synchronized phasor data of the i-th sample within a preset time period; The mapping size of the synchronous phasor data for each sample in the sample is calculated based on the fluctuation amplitude and is expressed as follows: In the above formula, This represents the mapping size of the synchronized phasor data for the i-th sample. This represents the minimum size of all convolutional kernels in the dynamic convolutional layer. This represents the maximum size of all convolutional kernels in the dynamic convolutional layer. This represents the minimum fluctuation range of all sample synchronized phasor data. This represents the maximum fluctuation amplitude of the synchronized phasor data for all samples; Based on the mapping size, the convolution kernel corresponding to the synchronous phasor data of each sample is selected for convolution processing using the rounding allocation method.

4. The method according to claim 2 or 3, characterized in that, The step of selecting the convolution kernel corresponding to each sample's synchronized phasor data for convolution processing based on the mapping size using a rounding allocation method includes: Mapping size is The synchronized phasor data is allocated to a size of Convolution is performed using convolution kernels; If at least two synchronized phasor data are assigned to the same convolution kernel, then the synchronized phasor data whose size difference is not the smallest is assigned to the idle convolution kernel with the smallest corresponding size difference for convolution processing; where the size difference is the absolute value of the difference between the mapping size and the convolution kernel size.

5. The method according to claim 1, characterized in that, The attention mechanism of the local feature extraction layer is used to extract key features from the multiple feature maps to obtain weighted local features, which are represented as follows: In the above formula, This represents the weighted local feature. This represents the total number of sample synchronization phasor data within each sample. This represents the fluctuation amplitude of the synchronized phasor data for the i-th sample. The feature map represents the synchronized phasor data of the i-th sample.

6. The method according to claim 5, characterized in that, The global features of the weighted local features are captured through the global feature fusion layer, and the global features are fused with the weighted local features to obtain the fused features, which are represented as follows: In the above formula, This indicates the fusion feature. The weight matrix represents the global features. This represents the global feature. This indicates a global average pooling operation.

7. The method according to claim 1, characterized in that, The dynamically weighted feature loss function is expressed as: In the above formula, This represents the loss value; This represents the expected eigenvector of the synchronized phasor data in the nth feature channel; This represents the feature vector extracted from the nth feature channel of the synchronized phasor data; This represents the weight of the nth feature channel; Represents the regularization coefficient; The parameter matrix represents the convolution kernel.

8. A method for feature extraction of synchronized phasor data, characterized in that, The method includes: Acquire multiple measured synchronous phasor data within a preset time period. The synchronous phasor data includes voltage, current, frequency, reactive power, and active power, denoted as: ； The plurality of measured synchronous phasor data are input into an adaptive feature channel convolutional network to obtain feature vectors of the plurality of measured synchronous phasor data. The feature vectors of the synchronous phasor data are used for fault detection or classification tasks. The adaptive feature channel convolutional network is trained by the method described in claim 1.

9. A computer-readable storage medium, characterized in that, The system contains a computer program that, when executed by a processor, causes the processor to perform the steps of the method as described in any one of claims 1 or 8.

10. A terminal device, characterized in that, It includes a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the steps of the method as described in any one of claims 1 or 8.

Images