A method and system for detecting broadband oscillation parameters of a power system

Abstract

The present invention discloses a method and system for detecting broadband oscillation parameters of an electric power system, the method comprising: obtaining a broadband oscillation signal to be detected; using a preset deep residual network to perform group convolution on the broadband oscillation signal to be detected, and obtaining the oscillation modal features corresponding to the oscillation signals of different frequency bands in the broadband oscillation signal to be detected; using a preset timing generative adversarial network to perform oscillation parameter identification on the oscillation modal features, and obtaining the oscillation parameters corresponding to the oscillation signals of different frequency bands. The present invention can adapt to the multi-modal features of the broadband oscillation signal by using a deep residual network to perform group convolution on the broadband oscillation signal to be detected, so as to extract the oscillation modal features of the oscillation signals of different frequency bands; considering the strong time-varying characteristics of the broadband oscillation signal, the timing features of the oscillation parameters can be further extracted on the basis of the oscillation modal features through the timing generative adversarial network, so as to accurately obtain the oscillation parameters of the oscillation signals of different frequency bands.

Description

Technical Field

[0001] The present invention relates to the technical field of power systems, and in particular to a method and system for detecting broadband oscillation parameters of a power system. Background Art

[0002] With the widespread access of a high proportion of new energy and power electronic equipment to the power system, many devices and components in the power system, such as transformers, motors, rectifiers and inverters, show significant nonlinear characteristics, making the broadband oscillation signal of the power system present strong time-varying, strong nonlinear and multi-modal characteristics. Since in the broadband oscillation signal, multiple oscillation mode signals such as low-frequency oscillation, subsynchronous oscillation, supersynchronous oscillation, medium-frequency oscillation and high-frequency oscillation may exist simultaneously and overlap with each other, and the existing oscillation parameter detection method cannot overcome the influence of modal aliasing, and it is easy to have the problem of missing modes or false modes in the process of oscillation parameter identification, so the existing oscillation parameter detection method cannot accurately and effectively detect the broadband oscillation parameters of the power system. Summary of the invention

[0003] The present invention provides a method and system for detecting broadband oscillation parameters of an electric power system, so as to solve the technical problem that the existing oscillation parameter detection method cannot accurately and effectively detect the broadband oscillation parameters of the electric power system.

[0004] In order to solve the above technical problems, a first aspect of an embodiment of the present invention provides a method for detecting broadband oscillation parameters of a power system, comprising:

[0005] Acquire a broadband oscillation signal to be detected;

[0006] Using a preset deep residual network to perform group convolution on the broadband oscillation signal to be detected, so as to obtain oscillation modal features corresponding to oscillation signals of different frequency bands in the broadband oscillation signal to be detected;

[0007] A preset timing generative adversarial network is used to identify oscillation parameters of the oscillation modal characteristics, so as to obtain oscillation parameters corresponding to the oscillation signals in different frequency bands.

[0008] As a preferred solution, the method of using a preset deep residual network to perform group convolution on the broadband oscillation signal to be detected to obtain oscillation modal features corresponding to oscillation signals in different frequency bands in the broadband oscillation signal to be detected specifically includes:

[0009] Based on the deep residual network, feature extraction is performed on the broadband oscillation signal to be detected to obtain initial signal features;

[0010] Performing group convolution on the initial signal features to obtain signal local features;

[0011] Average pooling is performed on the local features of the signal to obtain oscillation modal features corresponding to the oscillation signals in different frequency bands.

[0012] As a preferred solution, based on the deep residual network, feature extraction is performed on the broadband oscillation signal to be detected to obtain initial signal features, specifically including:

[0013] A convolution layer and a maximum pooling layer are used to respectively perform feature extraction and feature dimension reduction on the broadband oscillation signal to be detected to obtain the initial signal features.

[0014] As a preferred solution, performing group convolution on the initial signal features to obtain signal local features specifically includes:

[0015] Performing feature grouping on the initial signal features to obtain a plurality of group feature graphs;

[0016] Based on a plurality of parallel residual convolution paths in the deep residual network, residual convolution is performed on a plurality of the grouped feature maps to obtain local features corresponding to each of the grouped feature maps; wherein the residual convolution path is formed by stacking a plurality of residual modules;

[0017] The local features are aggregated, and the aggregated local features are residually connected with the initial signal features to obtain the signal local features.

[0018] As a preferred solution, the residual convolution path includes several first residual modules and several second residual modules, the size of the convolution layer and the number of output channels in the first residual module are 3×3 and 64 respectively, and the size of the convolution layer and the number of output channels in the second residual module are 3×3 and 128 respectively; the number of the residual convolution paths is equal to the number of the grouped feature maps.

[0019] As a preferred solution, the method of using a preset timing generative adversarial network to identify oscillation parameters of the oscillation modal features to obtain oscillation parameters corresponding to the oscillation signals in different frequency bands specifically includes:

[0020] Based on the time series generative adversarial network, extracting time series features from the oscillation modal features to obtain oscillation parameter feature representation;

[0021] Oscillation parameter identification is performed on the oscillation parameter characteristic representation to obtain the oscillation parameters corresponding to the oscillation signals in different frequency bands; wherein the oscillation parameters include an oscillation frequency and an attenuation factor.

[0022] As a preferred solution, the time series feature extraction of the oscillation modal features based on the time series generative adversarial network to obtain the oscillation parameter feature representation specifically includes:

[0023] Utilizing several GRU layers in the timing generative adversarial network, extracting timing features from the oscillation modal features to obtain timing features of oscillation parameters;

[0024] The fully connected layer in the timing generative adversarial network is used to perform nonlinear transformation on the timing characteristics of the oscillation parameters to obtain the characteristic representation of the oscillation parameters.

[0025] As a preferred solution, performing oscillation parameter identification on the oscillation parameter characteristic representation to obtain the oscillation parameters corresponding to the oscillation signals in different frequency bands specifically includes:

[0026] Based on the output layer in the timing generative adversarial network, the probability distribution of the oscillation frequency and the probability distribution of the attenuation factor corresponding to the oscillation parameter characteristic representation are calculated by an activation function;

[0027] According to the oscillation frequency probability distribution and the attenuation factor probability distribution, the oscillation frequency and the attenuation factor corresponding to the oscillation signal in different frequency bands are obtained respectively.

[0028] As a preferred solution, the method specifically trains the deep residual network and the temporal generative adversarial network through the following steps:

[0029] The ResNeXt network is trained using a broadband oscillation simulation signal to obtain the deep residual network; wherein the oscillation mode of the broadband oscillation simulation signal includes at least two of low-frequency oscillation, subsynchronous oscillation, supersynchronous oscillation, medium-frequency oscillation and high-frequency oscillation;

[0030] The oscillation modal characteristics corresponding to the wide-band oscillation simulation signal output by the ResNeXt network during the training process are used to train the time series generative adversarial network to be trained, so as to iteratively optimize the parameters of the generator and the parameters of the discriminator in the time series generative adversarial network to be trained, and obtain the time series generative adversarial network: wherein the generator includes an input layer, a hidden layer and an output layer, and the hidden layer consists of 2 GRU layers and 1 fully connected layer.

[0031] A second aspect of an embodiment of the present invention provides a power system broadband oscillation parameter detection system, comprising:

[0032] A signal acquisition module, used to acquire a broadband oscillation signal to be detected;

[0033] An oscillation modal feature extraction module, used to perform group convolution on the broadband oscillation signal to be detected using a preset deep residual network, so as to obtain oscillation modal features corresponding to oscillation signals of different frequency bands in the broadband oscillation signal to be detected;

[0034] The oscillation parameter identification module is used to use a preset timing generative adversarial network to perform oscillation parameter identification on the oscillation modal characteristics to obtain oscillation parameters corresponding to the oscillation signals in different frequency bands.

[0035] Compared with the prior art, the beneficial effect of the embodiments of the present invention lies in that, by utilizing a deep residual network to perform group convolution on the wide-band oscillation signal to be detected, it can effectively adapt to the multimodal characteristics of the wide-band oscillation signal, thereby being able to extract the oscillation modal characteristics of oscillation signals in different frequency bands; taking into account the strong time-varying characteristics of the wide-band oscillation signal, the timing characteristics of the oscillation parameters can be further extracted on the basis of the oscillation modal characteristics through the timing generative adversarial network, thereby being able to accurately obtain the oscillation parameters of oscillation signals in different frequency bands, effectively avoiding the problems of missing modes and false modes. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 1 is a flow chart of a method for detecting wide-band oscillation parameters of a power system according to an embodiment of the present invention;

[0037] Figure 2 It is a schematic diagram of the fitting between the oscillation frequency detected by using a deep residual network and a temporal generative adversarial network and the actually set oscillation frequency in an embodiment of the present invention;

[0038] Figure 3 It is a schematic diagram of the fitting between the attenuation factor obtained by using the deep residual network and the temporal generative adversarial network detection in an embodiment of the present invention and the actually set attenuation factor;

[0039] Figure 4 It is a structural schematic diagram of a power system broadband oscillation parameter detection system in an embodiment of the present invention. DETAILED DESCRIPTION

[0040] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0041] See also Figure 1 According to a first aspect of an embodiment of the present invention, there is provided a method for detecting broadband oscillation parameters of a power system, comprising the following steps S1 to S3:

[0042] Step S1, obtaining a broadband oscillation signal to be detected;

[0043] Step S2, using a preset deep residual network to perform group convolution on the broadband oscillation signal to be detected, to obtain oscillation modal features corresponding to oscillation signals of different frequency bands in the broadband oscillation signal to be detected;

[0044] Step S3, using a preset timing generative adversarial network to identify oscillation parameters of the oscillation modal characteristics, and obtaining oscillation parameters corresponding to the oscillation signals in different frequency bands.

[0045] Specifically, since the broadband oscillation signal has a multimodal feature, that is, it may contain multiple oscillation modes, such as low-frequency oscillation, subsynchronous oscillation, supersynchronous oscillation, medium-frequency oscillation and high-frequency oscillation, illustratively, the frequency band of low-frequency oscillation is 0.1Hz~2.5Hz, the frequency band of sub / supersynchronous oscillation is 2.5Hz~100Hz, the frequency band of medium-frequency oscillation is 100Hz~1000Hz, and the frequency band of high-frequency oscillation is above 1000Hz. Therefore, in view of this multimodal feature, this embodiment uses a deep residual network to perform group convolution on the acquired broadband oscillation signal to be detected, so as to extract the features of the broadband oscillation signal to be detected in groups, thereby realizing the extraction of complex features under multiple frequency components and avoiding the problems of missing modes and false modes.

[0046] Furthermore, since the broadband oscillation signal has a strong time-varying characteristic, the present embodiment inputs the acquired oscillation modal features into a timing generative adversarial network for further timing feature extraction, thereby effectively capturing the dynamic changes of the broadband oscillation signal to be detected in the time dimension, and determining the timing characteristics of the oscillation parameters. The oscillation parameters can be identified by utilizing the timing characteristics of the oscillation parameters, thereby obtaining the oscillation parameters corresponding to the oscillation signals of different frequency bands.

[0047] The method for detecting wide-band oscillation parameters of a power system provided by an embodiment of the present invention can effectively adapt to the multi-modal characteristics of the wide-band oscillation signal by utilizing a deep residual network to perform group convolution on the wide-band oscillation signal to be detected, thereby being able to extract the oscillation modal characteristics of the oscillation signals in different frequency bands; taking into account the strong time-varying characteristics of the wide-band oscillation signal, the timing characteristics of the oscillation parameters can be further extracted on the basis of the oscillation modal characteristics through a timing generative adversarial network, thereby being able to accurately obtain the oscillation parameters of the oscillation signals in different frequency bands, effectively avoiding the problems of missing modes and false modes.

[0048] As a preferred solution, the method of using a preset deep residual network to perform group convolution on the broadband oscillation signal to be detected to obtain oscillation modal features corresponding to oscillation signals in different frequency bands in the broadband oscillation signal to be detected specifically includes:

[0049] Based on the deep residual network, feature extraction is performed on the broadband oscillation signal to be detected to obtain initial signal features;

[0050] Performing group convolution on the initial signal features to obtain signal local features;

[0051] Average pooling is performed on the local features of the signal to obtain oscillation modal features corresponding to the oscillation signals in different frequency bands.

[0052] Specifically, this embodiment first extracts the initial signal features of the broadband oscillation signal to be detected, and then performs group convolution on the initial signal features to convert the features of each group to obtain local features. Since the features of each group are processed separately, it is possible to effectively extract the important features of the oscillation signals of different frequency bands in the broadband oscillation signal to be detected, thereby increasing the expressive power of the model and reducing the amount of calculation. Furthermore, the local features of the signal obtained by the group convolution are subjected to average pooling processing. It can be understood that the average pooling processing retains the averaged overall features by calculating the average value of the local area, and at the same time helps to remove noise, and finally obtains the oscillation modal features corresponding to the oscillation signals of different frequency bands.

[0053] As a preferred solution, based on the deep residual network, feature extraction is performed on the broadband oscillation signal to be detected to obtain initial signal features, specifically including:

[0054] A convolution layer and a maximum pooling layer are used to respectively perform feature extraction and feature dimension reduction on the broadband oscillation signal to be detected to obtain the initial signal features.

[0055] Specifically, this embodiment first uses a convolutional layer with a size of 7×7, an output channel number of 64, and a step size of 2 to perform preliminary feature extraction on the broadband oscillation signal to be detected, and then uses a maximum pooling layer with a size of 3×3 and a step size of 2 to perform feature dimension reduction on the preliminary extracted features, reducing the size of the feature map. It can be understood that the maximum pooling layer can highlight the important features in the data, and enhance the recognition ability of the main features of the local area by retaining the strongest feature response of the area, and finally obtain the initial signal features. This embodiment achieves the purpose of focusing on different types of features through maximum pooling and average pooling, captures more comprehensive feature information, strengthens the balance between feature enhancement and feature stability, and enables the overall network to adapt to different types of input data, thereby improving the performance of the network.

[0056] As a preferred solution, performing group convolution on the initial signal features to obtain signal local features specifically includes:

[0057] Performing feature grouping on the initial signal features to obtain a plurality of group feature graphs;

[0058] Based on a plurality of parallel residual convolution paths in the deep residual network, residual convolution is performed on a plurality of the grouped feature maps to obtain local features corresponding to each of the grouped feature maps; wherein the residual convolution path is formed by stacking a plurality of residual modules;

[0059] The local features are aggregated, and the aggregated local features are residually connected with the initial signal features to obtain the signal local features.

[0060] Specifically, in order to extract the oscillation modal features of oscillation modes such as low-frequency oscillation, subsynchronous oscillation, supersynchronous oscillation, medium-frequency oscillation and high-frequency oscillation, this embodiment first uses a convolution layer of size 1×1 to perform feature grouping on the initial signal features to obtain several group feature maps. Exemplarily, the number of groups is set to 32 groups according to experience. In actual application, it can also be adjusted according to the complexity of the task. This embodiment is not specifically limited here. Furthermore, using several parallel residual convolution paths, parallel residual convolution is performed on several of the group feature maps, that is, the processing process of each group feature map is independent of each other, and the topological structure of each residual convolution path is the same, and finally the local features of each group feature map are extracted. It is worth noting that the residual convolution path is formed by stacking several residual modules, so that the depth of the deep residual network can be deepened, so that the model can simultaneously capture the shallow simple features and deep complex features of the broadband oscillation signal, and realize the hierarchical extraction of all important features. Each residual module contains a convolutional layer, a batch normalization layer and a ReLU activation function, where the ReLU activation function is a nonlinear activation function. When the input value is less than 0, the output is 0, and when the input value is greater than or equal to 0, the output value is the input value. This piecewise characteristic enables the ReLU function to introduce nonlinearity, because it is not linear in the entire definition domain, and the activation mode of the ReLU function is nonlinear. Different input values ​​will cause different neurons to be activated or not activated. This combination of activation modes can produce nonlinear outputs, thereby being able to adapt to the nonlinear characteristics of broadband oscillation signals.

[0061] Furthermore, each of the local features is aggregated, and a residual connection is performed between the aggregated local features and the initial signal features, so that the information of the initial signal features can be retained, which helps to avoid the gradient vanishing problem.

[0062] As a preferred solution, the residual convolution path includes several first residual modules and several second residual modules, the size of the convolution layer and the number of output channels in the first residual module are 3×3 and 64 respectively, and the size of the convolution layer and the number of output channels in the second residual module are 3×3 and 128 respectively; the number of the residual convolution paths is equal to the number of the grouped feature maps.

[0063] As a preferred solution, the method of using a preset timing generative adversarial network to identify oscillation parameters of the oscillation modal features to obtain oscillation parameters corresponding to the oscillation signals in different frequency bands specifically includes:

[0064] Based on the time series generative adversarial network, extracting time series features from the oscillation modal features to obtain oscillation parameter feature representation;

[0065] Oscillation parameter identification is performed on the oscillation parameter characteristic representation to obtain the oscillation parameters corresponding to the oscillation signals in different frequency bands; wherein the oscillation parameters include an oscillation frequency and an attenuation factor.

[0066] Specifically, the time series generative adversarial network in this embodiment is an improved generative adversarial network, which maps high-dimensional time series data to a low-dimensional latent space, reduces the dimension of the generator learning space, and retains the important features of the time series data, thereby improving the accuracy of parameter identification. By using a time series generative adversarial network, time series feature extraction can be performed based on the oscillation modal features. It can be understood that the oscillation frequency is the number of periodic changes completed by the oscillation in a unit time, which determines the speed of the signal change, and the attenuation factor refers to the proportion of the signal attenuation in each cycle, that is, as the broadband oscillation signal changes dynamically in the time dimension, the oscillation parameters also change. Therefore, through time series feature extraction, the oscillation parameter feature representation can be obtained.

[0067] As a preferred solution, the time series feature extraction of the oscillation modal features based on the time series generative adversarial network to obtain the oscillation parameter feature representation specifically includes:

[0068] Utilizing several GRU layers in the timing generative adversarial network, extracting timing features from the oscillation modal features to obtain timing features of oscillation parameters;

[0069] The fully connected layer in the timing generative adversarial network is used to perform nonlinear transformation on the timing characteristics of the oscillation parameters to obtain the characteristic representation of the oscillation parameters.

[0070] Specifically, the temporal generative adversarial network in this embodiment is composed of a generator and a discriminator. The generator includes an input layer, a hidden layer and an output layer. The hidden layer is composed of several GRU (Gated Recurrent Unit) layers and a fully connected layer. The input layer is used to transmit the oscillation modal features to the hidden layer. The GRU layer in the hidden layer is responsible for extracting features in the time dimension. Through the GRU gating mechanism, that is, the update gate and the reset gate, it can effectively capture the dynamic changes of broadband oscillation signals in the time dimension, remember the information of long time spans, adapt to the dynamic changes of broadband oscillation signals in time, and provide real-time oscillation parameter identification capabilities. The update gate and the reset gate allow the model to simultaneously handle the long-term and short-term dependency problems in the signal, retain the long-term dependency information while capturing short-term changes, thereby obtaining the temporal features of the broadband oscillation signal, and then use the fully connected layer to integrate the temporal features extracted by the GRU layer, and realize the flexible transformation of the input feature dimension to the output feature dimension, so as to map the high-dimensional temporal feature space to the low-dimensional oscillation parameter recognition space.

[0071] As a preferred solution, performing oscillation parameter identification on the oscillation parameter characteristic representation to obtain the oscillation parameters corresponding to the oscillation signals in different frequency bands specifically includes:

[0072] Based on the output layer in the timing generative adversarial network, the probability distribution of the oscillation frequency and the probability distribution of the attenuation factor corresponding to the oscillation parameter characteristic representation are calculated by an activation function;

[0073] According to the oscillation frequency probability distribution and the attenuation factor probability distribution, the oscillation frequency and the attenuation factor corresponding to the oscillation signal in different frequency bands are obtained respectively.

[0074] Specifically, based on the output layer of the generator in the time-series generative adversarial network, the activation function can be used to calculate the probabilities of different oscillation frequencies and different attenuation factors corresponding to the current oscillation parameter characteristic representation, thereby determining the oscillation frequency probability distribution and the attenuation factor probability distribution. According to the oscillation frequency and attenuation factor with the highest probability in the oscillation frequency probability distribution and the attenuation factor probability distribution, the oscillation frequency and attenuation factor corresponding to the oscillation signals in different frequency bands can be output.

[0075] As a preferred solution, the method specifically trains the deep residual network and the temporal generative adversarial network through the following steps:

[0076] The ResNeXt network is trained using a broadband oscillation simulation signal to obtain the deep residual network; wherein the oscillation mode of the broadband oscillation simulation signal includes at least two of low-frequency oscillation, subsynchronous oscillation, supersynchronous oscillation, medium-frequency oscillation and high-frequency oscillation;

[0077] The oscillation modal characteristics corresponding to the wide-band oscillation simulation signal output by the ResNeXt network during the training process are used to train the time series generative adversarial network to be trained, so as to iteratively optimize the parameters of the generator and the parameters of the discriminator in the time series generative adversarial network to be trained, and obtain the time series generative adversarial network: wherein the generator includes an input layer, a hidden layer and an output layer, and the hidden layer consists of 2 GRU layers and 1 fully connected layer.

[0078] Specifically, the deep residual network in this embodiment is specifically a ResNeXt network. The ResNeXt network draws on the grouping idea of ​​the Inception structure, proposes a group convolution module, and combines it with the residual module to finally obtain a network structure with a simple structure and fewer parameters, and its performance is better than that of the ResNet network. It not only retains the advantage of the ResNet network in avoiding the problem of vanishing model gradients caused by network stacking, but also incorporates the multi-scale idea, dividing the input data into multiple data sets and inputting them into each path of the model to extract features. Each feature extraction channel in the model is stacked through the same convolutional neural network structure, which reduces the hyperparameters in the network compared to other multi-scale networks. In order to enable the ResNeXt network to learn the ability to extract oscillation modal features of different oscillation modes, this embodiment first simulates to obtain a broadband oscillation simulation signal whose oscillation modes include at least two of low-frequency oscillation, subsynchronous oscillation, supersynchronous oscillation, medium-frequency oscillation and high-frequency oscillation. Preferably, the oscillation mode of the broadband oscillation simulation signal simultaneously includes low-frequency oscillation, sub / supersynchronous oscillation, medium-frequency oscillation and high-frequency oscillation, and then uses the broadband oscillation simulation signal to train the original ResNeXt network. Exemplarily, the broadband oscillation simulation signal is divided into a training set and a test set in a ratio of 7:3, and the training set is used for training, and the test set is used to verify the performance of the network after each training iteration. The parameters of the ResNeXt network are optimized through iterative training, and finally the desired deep residual network is obtained.

[0079] Furthermore, during the training process of the ResNeXt network, the extracted oscillation modal features are selected and dimensionally reduced to screen out unnecessary features and reduce the data dimension. Then, the pre-trained oscillation modal features are used to train the time series generative adversarial network to be trained, so as to iteratively optimize the parameters of the generator and the parameters of the discriminator in the time series generative adversarial network to be trained. In each round of training, the parameters of the generator are first fixed, the discriminator is trained, the real time series samples and the time series samples generated by the generator in the previous round are input into the discriminator for scoring, and the loss function (usually using binary cross entropy loss) is calculated based on the scoring to update the parameters of the discriminator. Then, the parameters of the discriminator are fixed, the generator is trained, the generator generates the corresponding time series features based on the input oscillation modal features, and the features are input into the discriminator for scoring. The difference between the score and the real label (usually 1) is used as the loss function, and the parameters of the generator are updated through back propagation to minimize the loss function. It is worth noting that the hidden layer of the generator in this embodiment includes 2 GRU layers, so as to ensure the detection accuracy of the oscillation parameters while ensuring the fast speed of temporal feature extraction.

[0080] After the training of the ResNeXt network and the time-series generative adversarial network is completed, the independent validation set and the above test set are used to evaluate the oscillation parameter detection performance of the trained deep residual network and the time-series generative adversarial network. The evaluation indicators are the mean absolute error and the root mean square error. The test set is input into the trained deep residual network and tested in combination with the time-series generative adversarial network. The oscillation frequency and attenuation factor in the test results are evaluated respectively. The following Table 1 shows the average analysis results of 300 tests.

[0081] Table 1 Detection and evaluation results of oscillation frequency and attenuation factor

[0082]

[0083] The fitting between the oscillation frequency and attenuation factor detected by the trained deep residual network and time series generative adversarial network and the actual set oscillation frequency and attenuation factor are shown in the figure below. Figure 2 and Figure 3 As shown in Table 1, Figure 2 and Figure 3 It can be seen that the deep residual network and timing generative adversarial network in this embodiment have good detection performance for oscillation parameters, and the oscillation frequency and attenuation factor obtained by detection have a good fitting effect with the actual oscillation frequency and attenuation factor, and have good detection accuracy.

[0084] See also Figure 4 A second aspect of an embodiment of the present invention provides a power system broadband oscillation parameter detection system, comprising:

[0085] A signal acquisition module 11 is used to acquire a broadband oscillation signal to be detected;

[0086] An oscillation modal feature extraction module 12 is used to perform group convolution on the broadband oscillation signal to be detected using a preset deep residual network to obtain oscillation modal features corresponding to oscillation signals of different frequency bands in the broadband oscillation signal to be detected;

[0087] The oscillation parameter identification module 13 is used to identify the oscillation parameters of the oscillation modal characteristics by using a preset timing generative adversarial network, so as to obtain the oscillation parameters corresponding to the oscillation signals in different frequency bands.

[0088] As a preferred solution, the oscillation modal feature extraction module 12 is used to use a preset deep residual network to perform group convolution on the broadband oscillation signal to be detected, and obtain the oscillation modal features corresponding to the oscillation signals of different frequency bands in the broadband oscillation signal to be detected, specifically including:

[0089] Based on the deep residual network, feature extraction is performed on the broadband oscillation signal to be detected to obtain initial signal features;

[0090] Performing group convolution on the initial signal features to obtain signal local features;

[0091] Average pooling is performed on the local features of the signal to obtain oscillation modal features corresponding to the oscillation signals in different frequency bands.

[0092] As a preferred solution, the oscillation mode feature extraction module 12 is used to extract features of the broadband oscillation signal to be detected based on the deep residual network to obtain initial signal features, specifically including:

[0093] A convolution layer and a maximum pooling layer are used to respectively perform feature extraction and feature dimension reduction on the broadband oscillation signal to be detected to obtain the initial signal features.

[0094] As a preferred solution, the oscillation mode feature extraction module 12 is used to perform group convolution on the initial signal features to obtain signal local features, specifically including:

[0095] Performing feature grouping on the initial signal features to obtain a plurality of group feature graphs;

[0096] Based on a plurality of parallel residual convolution paths in the deep residual network, residual convolution is performed on a plurality of the grouped feature maps to obtain local features corresponding to each of the grouped feature maps; wherein the residual convolution path is formed by stacking a plurality of residual modules;

[0097] The local features are aggregated, and the aggregated local features are residually connected with the initial signal features to obtain the signal local features.

[0098] As a preferred solution, the residual convolution path includes several first residual modules and several second residual modules, the size of the convolution layer and the number of output channels in the first residual module are 3×3 and 64 respectively, and the size of the convolution layer and the number of output channels in the second residual module are 3×3 and 128 respectively; the number of the residual convolution paths is equal to the number of the grouped feature maps.

[0099] As a preferred solution, the oscillation parameter identification module 13 is used to identify the oscillation parameters of the oscillation modal features using a preset time-series generative adversarial network to obtain the oscillation parameters corresponding to the oscillation signals in different frequency bands, specifically including:

[0100] Based on the time series generative adversarial network, extracting time series features from the oscillation modal features to obtain oscillation parameter feature representation;

[0101] Oscillation parameter identification is performed on the oscillation parameter characteristic representation to obtain the oscillation parameters corresponding to the oscillation signals in different frequency bands; wherein the oscillation parameters include an oscillation frequency and an attenuation factor.

[0102] As a preferred solution, the oscillation parameter identification module 13 is used to extract the time series features of the oscillation modal features based on the time series generative adversarial network to obtain the oscillation parameter feature representation, which specifically includes:

[0103] Utilizing several GRU layers in the timing generative adversarial network, extracting timing features from the oscillation modal features to obtain timing features of oscillation parameters;

[0104] The fully connected layer in the timing generative adversarial network is used to perform nonlinear transformation on the timing characteristics of the oscillation parameters to obtain the characteristic representation of the oscillation parameters.

[0105] As a preferred solution, the oscillation parameter identification module 13 is used to identify the oscillation parameters of the oscillation parameter characteristic representation to obtain the oscillation parameters corresponding to the oscillation signals in different frequency bands, specifically including:

[0106] Based on the output layer in the timing generative adversarial network, the probability distribution of the oscillation frequency and the probability distribution of the attenuation factor corresponding to the oscillation parameter characteristic representation are calculated by an activation function;

[0107] According to the oscillation frequency probability distribution and the attenuation factor probability distribution, the oscillation frequency and the attenuation factor corresponding to the oscillation signal in different frequency bands are obtained respectively.

[0108] As a preferred solution, the system further includes a model training module for:

[0109] The ResNeXt network is trained using a broadband oscillation simulation signal to obtain the deep residual network; wherein the oscillation mode of the broadband oscillation simulation signal includes at least two of low-frequency oscillation, subsynchronous oscillation, supersynchronous oscillation, medium-frequency oscillation and high-frequency oscillation;

[0110] The oscillation modal characteristics corresponding to the wide-band oscillation simulation signal output by the ResNeXt network during the training process are used to train the time series generative adversarial network to be trained, so as to iteratively optimize the parameters of the generator and the parameters of the discriminator in the time series generative adversarial network to be trained, and obtain the time series generative adversarial network: wherein the generator includes an input layer, a hidden layer and an output layer, and the hidden layer consists of 2 GRU layers and 1 fully connected layer.

[0111] The power system broadband oscillation parameter detection system provided by the embodiment of the present invention can effectively adapt to the multi-modal characteristics of broadband oscillation signals by using a deep residual network to perform group convolution on the broadband oscillation signals to be detected, thereby being able to extract the oscillation modal characteristics of oscillation signals in different frequency bands; taking into account the strong time-varying characteristics of broadband oscillation signals, the oscillation parameter timing characteristics can be further extracted on the basis of the oscillation modal characteristics through a timing generative adversarial network, thereby being able to accurately obtain the oscillation parameters of oscillation signals in different frequency bands, effectively avoiding the problems of missing modes and false modes.

[0112] The above is a preferred embodiment of the present invention. It should be pointed out that a person skilled in the art can make several improvements and modifications without departing from the principle of the present invention. These improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A method for detecting broadband oscillation parameters of a power system, characterized in that: include: Acquire a broadband oscillation signal to be detected; Using a preset deep residual network to perform group convolution on the broadband oscillation signal to be detected, so as to obtain oscillation modal features corresponding to oscillation signals of different frequency bands in the broadband oscillation signal to be detected; Using a preset time-series generative adversarial network to identify oscillation parameters of the oscillation modal characteristics, and obtaining oscillation parameters corresponding to the oscillation signals in different frequency bands; The method of using a preset deep residual network to perform group convolution on the broadband oscillation signal to be detected to obtain oscillation modal features corresponding to oscillation signals in different frequency bands in the broadband oscillation signal to be detected specifically includes: Based on the deep residual network, feature extraction is performed on the broadband oscillation signal to be detected to obtain initial signal features; Performing group convolution on the initial signal features to obtain signal local features; Performing average pooling processing on the local features of the signal to obtain oscillation modal features corresponding to the oscillation signals in different frequency bands; The performing group convolution on the initial signal features to obtain signal local features specifically includes: Performing feature grouping on the initial signal features to obtain a plurality of group feature graphs; Based on a plurality of parallel residual convolution paths in the deep residual network, residual convolution is performed on a plurality of the grouped feature maps to obtain local features corresponding to each of the grouped feature maps; wherein the residual convolution path is formed by stacking a plurality of residual modules; The local features are aggregated, and the aggregated local features are residually connected with the initial signal features to obtain the signal local features.

2. The method for detecting broadband oscillation parameters of a power system according to claim 1, characterized in that: The extracting features of the broadband oscillation signal to be detected based on the deep residual network to obtain initial signal features specifically includes: A convolution layer and a maximum pooling layer are used to respectively perform feature extraction and feature dimension reduction on the broadband oscillation signal to be detected to obtain the initial signal features.

3. The method for detecting broadband oscillation parameters of a power system according to claim 1, wherein: The residual convolution path includes several first residual modules and several second residual modules, the size of the convolution layer and the number of output channels in the first residual module are 3×3 and 64 respectively, and the size of the convolution layer and the number of output channels in the second residual module are 3×3 and 128 respectively; the number of the residual convolution paths is equal to the number of the grouped feature maps.

4. The method for detecting broadband oscillation parameters of a power system according to claim 1, wherein: The step of using a preset time-series generative adversarial network to identify oscillation parameters of the oscillation modal features to obtain oscillation parameters corresponding to the oscillation signals in different frequency bands specifically includes: Based on the time series generative adversarial network, extracting time series features from the oscillation modal features to obtain oscillation parameter feature representation; Oscillation parameter identification is performed on the oscillation parameter characteristic representation to obtain the oscillation parameters corresponding to the oscillation signals in different frequency bands; wherein the oscillation parameters include an oscillation frequency and an attenuation factor.

5. The method for detecting broadband oscillation parameters of a power system according to claim 4, characterized in that: The step of extracting time series features from the oscillation modal features based on the time series generative adversarial network to obtain oscillation parameter feature representation specifically includes: Utilizing several GRU layers in the timing generative adversarial network, extracting timing features from the oscillation modal features to obtain timing features of oscillation parameters; The fully connected layer in the timing generative adversarial network is used to perform nonlinear transformation on the timing characteristics of the oscillation parameters to obtain the characteristic representation of the oscillation parameters.

6. The method for detecting broadband oscillation parameters of a power system according to claim 4, characterized in that: The performing oscillation parameter identification on the oscillation parameter characteristic representation to obtain the oscillation parameters corresponding to the oscillation signals in different frequency bands specifically includes: Based on the output layer in the timing generative adversarial network, the probability distribution of the oscillation frequency and the probability distribution of the attenuation factor corresponding to the oscillation parameter characteristic representation are calculated by an activation function; According to the oscillation frequency probability distribution and the attenuation factor probability distribution, the oscillation frequency and the attenuation factor corresponding to the oscillation signal in different frequency bands are obtained respectively.

7. The method for detecting broadband oscillation parameters of a power system according to claim 1, wherein: The method specifically obtains the deep residual network and the temporal generative adversarial network by training through the following steps: The ResNeXt network is trained using a broadband oscillation simulation signal to obtain the deep residual network; wherein the oscillation mode of the broadband oscillation simulation signal includes at least two of low-frequency oscillation, subsynchronous oscillation, supersynchronous oscillation, medium-frequency oscillation and high-frequency oscillation; The oscillation modal characteristics corresponding to the wide-band oscillation simulation signal output by the ResNeXt network during the training process are used to train the time series generative adversarial network to be trained, so as to iteratively optimize the parameters of the generator and the parameters of the discriminator in the time series generative adversarial network to be trained, and obtain the time series generative adversarial network: wherein the generator includes an input layer, a hidden layer and an output layer, and the hidden layer consists of 2 GRU layers and 1 fully connected layer.

8. A power system broadband oscillation parameter detection system, characterized in that: include: A signal acquisition module, used to acquire a broadband oscillation signal to be detected; An oscillation modal feature extraction module, used to perform group convolution on the broadband oscillation signal to be detected using a preset deep residual network, so as to obtain oscillation modal features corresponding to oscillation signals of different frequency bands in the broadband oscillation signal to be detected; An oscillation parameter identification module, used to identify the oscillation parameters of the oscillation modal features using a preset timing generative adversarial network, and obtain oscillation parameters corresponding to the oscillation signals in different frequency bands; The oscillation modal feature extraction module is used to perform group convolution on the broadband oscillation signal to be detected using a preset deep residual network to obtain oscillation modal features corresponding to oscillation signals of different frequency bands in the broadband oscillation signal to be detected, specifically including: Based on the deep residual network, feature extraction is performed on the broadband oscillation signal to be detected to obtain initial signal features; Performing group convolution on the initial signal features to obtain signal local features; Performing average pooling processing on the local features of the signal to obtain oscillation modal features corresponding to the oscillation signals in different frequency bands; The oscillation mode feature extraction module is used to perform group convolution on the initial signal features to obtain local signal features, specifically including: Performing feature grouping on the initial signal features to obtain a plurality of group feature graphs; Based on a plurality of parallel residual convolution paths in the deep residual network, residual convolution is performed on a plurality of the grouped feature maps to obtain local features corresponding to each of the grouped feature maps; wherein the residual convolution path is formed by stacking a plurality of residual modules; The local features are aggregated, and the aggregated local features are residually connected with the initial signal features to obtain the signal local features.

Images