A power grid synchronous broadband forced oscillation detection and positioning method, system and medium

Abstract

The application discloses a power grid synchronous broadband forced oscillation detection and positioning method and system and a medium. The method comprises the following steps: obtaining oscillation data sequences and a power system topology graph of a power grid, and inputting the pre-trained multi-task graph sequence network to obtain an oscillation source positioning detection result. The oscillation data sequences are input into a convolution layer, a pooling layer and an output feature obtained by a created embeddable sequence autoregressive module. The power system topology graph is input into a multi-task graph sequence network to perform information transmission, and a node feature of each node is obtained through a gated recurrent unit (GRU). Then, the method is input into a graph-level detection branch and a node-level detection branch to realize event type classification of an oscillation source and positioning of the oscillation source. The application aims to reduce the interference of other information on broadband oscillation information, dynamically analyze data from two dimensions of time and space, and quickly and accurately realize broadband oscillation source positioning.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of power system, in particular to a power grid synchronous wide frequency forced oscillation detection and positioning method, system and medium. BACKGROUND

[0002] In modern power systems, the increasing integration of renewable energy, high-voltage direct current links and power electronic devices greatly increases the possibility of wide frequency oscillations, ranging from low frequency oscillations of a few hertz to high frequency oscillations of hundreds of hertz, thereby posing a major challenge to traditional stability analysis methods. Accurate detection and positioning of such oscillations is crucial for implementing targeted mitigation strategies to prevent equipment damage and large-scale system failures. At the same time, advances in wide-area monitoring technology, such as waveform measurement units, enable higher precision synchronization and significantly enhance monitoring capabilities in a wider frequency range and finer time resolution. However, the time-varying, nonlinear and dynamic characteristics of wide frequency oscillation propagation continue to hinder the effectiveness of real-time positioning work. To address the challenges of oscillation detection and positioning, existing methods can be broadly divided into three categories: physical-based, energy flow-based and data-driven methods. Physical-based methods use critical oscillation parameters to identify the source, however, a key limitation of this method is that they are difficult to accurately identify forced sources under resonance conditions, where the most severe oscillations may not occur near the actual source; secondly, to reduce the uncertainty of oscillation localization, energy flow-based methods are proposed, where dissipated energy flow is the most widely used method, however, under low damping conditions and high noise levels, this method still struggles with complex situations; considering the large amount of measurement data, data-driven methods have been developed, however, this method can cause information loss during inter-stage interaction. Other emerging methods, including graph convolutional neural networks and long short-term memory networks, have also been proposed to locate oscillations by integrating signal processing techniques with energy flow methods. Nevertheless, most methods require additional noise-resistant modules before positioning, thereby increasing the complexity of the model; secondly, purely physical and data-driven methods exhibit limited accuracy and real-time performance when dealing with high-speed dynamic oscillations; finally, traditional models usually assume that oscillations have been detected and only focus on positioning. Therefore, how to efficiently and accurately solve the problem of finding the source of complex wide frequency forced oscillations based on wide-area synchronous measurement data in modern power systems has become a key technical problem that needs to be solved. SUMMARY

[0003] The technical problems to be solved by the present application: in view of the above problems of the prior art, a power grid synchronous wide frequency forced oscillation detection and positioning method, system and medium are provided, the present application aims to efficiently and accurately solve the source finding problem of complex wide frequency forced oscillation in modern power system based on wide area synchronous measurement data, reduce the interference of other information on wide frequency oscillation information, dynamically analyze the data from two dimensions of time and space, and quickly and accurately realize wide frequency oscillation source positioning.

[0004] In order to solve the above technical problems, the technical scheme adopted by the present application is:

[0005] A power grid synchronous wide frequency forced oscillation detection and positioning method, comprising the following steps: obtaining an oscillation data sequence of a power grid and a power system topology graph, the oscillation data sequence comprising voltage, frequency and frequency change rate of the power grid, the nodes in the power system topology graph being generators or loads in the power grid, the edges being lines, and the oscillation characteristics of the nodes comprising voltage, frequency and frequency change rate of the nodes; inputting the oscillation data sequence of the power grid and the power system topology graph into a pre-trained multi-task graph sequence network to obtain an oscillation source positioning detection result, comprising:

[0006] S101, extracting key features from the oscillation data sequence through a convolution layer and a pooling layer ;

[0007] S102, inputting the key features into an embeddable sequence autoregressive module, the embeddable sequence autoregressive module performing embedded feature filtering on the key features through a plurality of stacked ARMA(1,0) filters, and summing up the responses of the plurality of stacked ARMA(1,0) filters to obtain final output features ;

[0008] S103, taking the output features and the power system topology graph as inputs of the multi-task graph sequence network, performing information transmission through the multi-task graph sequence network, and obtaining node features of each node through a gated recurrent unit (GRU) ;

[0009] S104, inputting the node features of each node into a graph-level detection branch and a node-level detection branch respectively, realizing event type classification of the oscillation source through the graph-level detection branch, and realizing positioning of the oscillation source through the node-level detection branch.

[0010] ​​Optionally, the process of acquiring the oscillation data sequence of the power grid and the power system topology map includes acquiring the voltage time series of the power grid, filtering the voltage time series to eliminate the DC component to extract the frequency of the power grid, and calculating the frequency change rate based on the frequency of the power grid and the following formula:

[0011] ;

[0012] in, and Sampling points and sampling points frequency, The sampling interval is... The voltage sampling frequency is used, and the obtained voltage, frequency, and rate of change of the power grid are normalized to form an oscillation data sequence.

[0013] Optionally, when acquiring the oscillation data sequence of the power grid and the power system topology map, the functional expression of the obtained power system topology map is:

[0014] ;

[0015] in, This is a power system topology diagram. For a set of nodes, It is an unordered set of edges. The oscillation characteristics of a node include its voltage, frequency, and rate of change of frequency, and the unordered edge set. Through the adjacency matrix This represents the line connection relationship between different nodes. The node and the first When nodes are directly connected by a line, the adjacency matrix The Middle Line number Column elements ;otherwise , indicating the first The node and the first The nodes have no direct connection.

[0016] Optionally, in step S101, key features are extracted from the oscillating data sequence through convolutional and pooling layers. The function expression for the convolution operation of the convolutional layer is:

[0017] ;

[0018] In the above formula, This is the output of the convolution operation. The number of types of oscillating data. for The first moment The input signal for the oscillation data; convolution kernel The offset time that is applied, and The first moment Input signal of oscillation data Time axis alignment; convolution kernel For the first Such oscillation data in The convolution kernel corresponding to each time step.

[0019] Optionally, in step S102, the... The final output feature is obtained by summing the responses of the stacked ARMA(1,0) filters. The function expression is:

[0020] ;

[0021] in, Let be the number of stacking levels of the ARMA(1,0) filter. For activation function, For the first The output characteristics of an ARMA(1,0) filter. For the first Input characteristics of an ARMA(1,0) filter. and Let be the trainable weight parameters of the k-th ARMA(1,0) filter. As a key feature, The normalized Laplace matrix is ​​symmetric; the ARMA(1,0) filter is approximated using a first-order recursive method as shown in the following equation:

[0022] ;

[0023] ;

[0024] in, and These represent the output characteristics and the oscillation characteristics of the input nodes of the ARMA(1,0) filter, respectively. and For coefficients, For symmetric normalized Laplace matrix The intermediate coefficients related to the eigenvalues, and These are the symmetric normalized Laplace matrices. Maximum and minimum eigenvalues It is an identity matrix.

[0025] Optionally, in step S103, the output features will be... The power system topology graph is used as input to the multi-task graph sequence network. Information is transmitted through the multi-task graph sequence network, and the node characteristics of each node are obtained through the gated cyclic unit (GRU). At that time, the functional expression for information transmission in the multi-task graph sequence network is:

[0026] ;

[0027] in, and The first The output characteristics and the oscillation characteristics of the input nodes during the next transmission. For activation function, Adjacency matrix and identity matrix The sum of adjacency matrices To represent the unordered set of edges in a power system topology graph The adjacency matrix, for The degree matrix of the diagonal nodes, and The first The node feature matrix and trainable parameter matrix are obtained during each transmission, and the node features of each node are obtained through a gated recurrent unit (GRU). The function expression is:

[0028] ;

[0029] ;

[0030] in, For the softmax function, The normalized adjacency matrix, For activation function, The output of the gated recurrent unit (GRU) layer. and is a trainable parameter matrix.

[0031] Optionally, the loss function used by the multi-task graph sequence network during training has the following expression:

[0032] ;

[0033] ;

[0034] ;

[0035] ;

[0036] in, This is the loss function used during training of the multi-task graph sequence network. This is the location detection result of the oscillation source in the node-level detection branch. These are the model parameters for a multi-task graph sequence network. and These are the weight hyperparameters for the node-level and graph-level detection branches learned by the softplus function, respectively. To consider the weighted binary cross-entropy loss function that takes into account node mask loss, The cross-entropy loss is used for graph classification tasks in the graph-level detection branch. For the first The mask of each node, which serves as a flag indicating whether a node is valid. For the first The weight of each node, This represents the total number of measurement nodes. This represents the number of generator nodes. Let cross-entropy be the loss function. and These are the detection results and ground truth values ​​for the graph classification task performed by the graph-level detection branch.

[0037] The present invention also provides a power grid synchronization broadband forced oscillation detection and location system, including a microprocessor and a memory interconnected thereto, wherein the microprocessor is programmed or configured to execute the power grid synchronization broadband forced oscillation detection and location method.

[0038] The present invention also provides a computer-readable storage medium storing a computer program or instructions that are programmed or configured to execute the power grid synchronization broadband forced oscillation detection and location method by a processor.

[0039] The present invention also provides a computer program product, including a computer program or instructions, which are programmed or configured to execute the power grid synchronization broadband forced oscillation detection and location method by a processor.

[0040] Compared with existing technologies, the present invention mainly achieves the following beneficial effects: The power grid synchronous broadband forced oscillation detection and location method of the present invention can make full use of the combination of autoregressive moving average model and multi-task graph sequence network and multi-task mechanism, which can effectively reduce the interference of other information on broadband oscillation information, efficiently and accurately solve the problem of finding the source of complex broadband forced oscillations based on wide-area synchronous measurement data in modern power systems, reduce the interference of other information on broadband oscillation information, and perform dynamic analysis of data from both time and space dimensions to quickly and accurately locate the broadband oscillation source. Attached Figure Description

[0041] Figure 1 This is a schematic diagram of the basic process of the method in an embodiment of the present invention.

[0042] Figure 2 This is the original frequency data of the generator based on the IEEE 39-bus system in this embodiment of the invention.

[0043] Figure 3 The frequency data is the generator normalized data based on the IEEE 39-bus system in this embodiment of the invention.

[0044] Figure 4 This is a schematic diagram of a network structure in an embodiment of the present invention that can be embedded with a sequence autoregressive module.

[0045] Figure 5 This is a comparison of the signal spectrum diagrams of the input and output of the embeddable sequence autoregressive module in the embodiments of the present invention. Detailed Implementation

[0046] To enable those skilled in the art to better understand the technical solution of the present invention, the technical solution of the present invention will be further described in detail below with reference to the accompanying drawings in the embodiments of the present invention.

[0047] like Figure 1 As shown, the method for detecting and locating synchronous broadband forced oscillations in a power grid in this embodiment includes the following steps: acquiring the oscillation data sequence and power system topology map of the power grid, wherein the oscillation data sequence includes the voltage, frequency, and frequency change rate of the power grid, and the nodes in the power system topology map are generators or loads in the power grid, the edges are lines, and the oscillation characteristics of the nodes include the voltage, frequency, and frequency change rate of the nodes; inputting the oscillation data sequence and power system topology map of the power grid into a pre-trained multi-task graph sequence network to obtain the oscillation source location detection result, including:

[0048] S101 extracts key features from the oscillating data sequence through convolutional and pooling layers. ;

[0049] S102, key features Input can be embedded into a sequence autoregressive module, the embedded sequence autoregressive module through Stacked ARMA(1,0) filters for key features Perform embedded feature filtering and... The final output feature is obtained by summing the responses of the stacked ARMA(1,0) filters. ;

[0050] S103 will output features The power system topology graph is used as input to the multi-task graph sequence network. Information is transmitted through the multi-task graph sequence network, and the node characteristics of each node are obtained through the gated cyclic unit (GRU). ;

[0051] S104, The node characteristics of each node. Input graph-level detection branch and node-level detection branch respectively. Use graph-level detection branch to classify the event type of the oscillation source and use node-level detection branch to locate the oscillation source.

[0052] In this embodiment, acquiring the oscillation data sequence of the power grid and the power system topology map includes acquiring the voltage time series of the power grid and collecting the oscillation data source of the synchronous motor excitation system, as well as the voltage of the injected broadband forced oscillation. The function expression in the time domain is as follows:

[0053] ;

[0054] In the above formula, , , and These represent the oscillation amplitude, damping coefficient, angular frequency, and initial phase, respectively. and angular frequency As two key parameters, once their values ​​are determined, they can uniquely define the oscillation characteristic curve; initial phase This determines the onset time of oscillation, a characteristic similar to the phase convergence phenomenon at the zero-crossing point of the power frequency; the acquired voltage The data contains wideband oscillation information and a large amount of redundant information. Therefore, the voltage time series can be filtered to eliminate the DC component and extract the grid frequency. As an optional implementation, to remove the long-term trend in the original data, this embodiment uses a zero-phase high-pass Butterworth filter to perform DC filtering on the original data, extracting and highlighting the frequency information of the wideband oscillation. In this embodiment, the frequency change rate is calculated based on the grid frequency and the following formula:

[0055] ;

[0056] in, and Sampling points and sampling points frequency, The sampling interval is... The sampling frequency of the voltage, where the sampling interval is typically... It can be set to 1 to 5, where , Indicates the number of sampling points. The sampling time interval is [not specified]. The obtained grid voltage [is then used]. ,frequency and frequency change rate After normalization, the resulting oscillating data sequence can be represented as a three-dimensional eigenvector:

[0057] .

[0058] Figure 2 This embodiment uses the raw frequency data of the generator based on the IEEE 39-bus system. Figure 3 This embodiment presents the normalized frequency data of generators based on the IEEE 39-bus system. Figure 2 and Figure 3 Curve G30 represents an oscillating generator, while the other curves represent non-oscillating generators. (Comparison) Figure 2 and Figure 3 It can be seen that, Figure 2 The frequencies of each generator, including both oscillating and non-oscillating generators, are almost identical, making it difficult to distinguish subtle variations; however, as Figure 3 As shown, after data preprocessing through filtering and normalization, the frequency difference between the oscillating source generator and the non-oscillating source generator becomes larger, and the oscillation waveform becomes distinguishable, providing stronger nodal characteristics for broadband oscillation localization.

[0059] The functional expression for the power system topology obtained in this embodiment is:

[0060] ;

[0061] in, This is a power system topology diagram. For a set of nodes, It is an unordered set of edges. The oscillation characteristics of a node include its voltage, frequency, and rate of change of frequency, and the unordered edge set. Through the adjacency matrix This represents the line connection relationship between different nodes. The node and the first When nodes are directly connected by a line, the adjacency matrix The Middle Line number Column elements ;otherwise , indicating the first The node and the first The nodes have no direct connection.

[0062] In step S101 of this embodiment, key features are extracted from the oscillating data sequence through convolutional and pooling layers. The function expression for the convolution operation of the convolutional layer is:

[0063] ;

[0064] In the above formula, This is the output of the convolution operation. The number of types of oscillating data. for The first moment The input signal for the oscillation data; convolution kernel The offset time that is applied, and The first moment Input signal of oscillation data Time axis alignment; convolution kernel For the first Such oscillation data in The convolution kernel corresponding to each time step.

[0065] Considering the long duration of low-frequency oscillations and the short duration of high-frequency oscillations, this embodiment further constructs an embeddable sequence autoregressive module. The embeddable sequence autoregressive module is a neural network component specifically designed for processing spatiotemporal sequence data. Its core idea is to combine the classic autoregressive moving average model from signal processing with a graph neural network to extract features from broadband oscillation information. The embeddable sequence autoregressive module uses an autoregressive moving average model for embedded feature filtering to extract oscillation-related features and physical features, simultaneously capturing the short-term and long-term dynamic characteristics of oscillations, corresponding to the modes of high-frequency and low-frequency oscillations, respectively. Figure 4 As shown, in this embodiment, a sequence autoregression module can be embedded through... Stacked ARMA(1,0) filters for key features Perform embedded feature filtering and... The final output feature is obtained by summing the responses of the stacked ARMA(1,0) filters. In step S102, for The final output feature is obtained by summing the responses of the stacked ARMA(1,0) filters. The function expression is:

[0066] ;

[0067] in, Let be the number of stacking levels of the ARMA(1,0) filter. For activation function, For the first The output characteristics of an ARMA(1,0) filter. For the first Input characteristics of an ARMA(1,0) filter. and Let be the trainable weight parameters of the k-th ARMA(1,0) filter. As a key feature, For a symmetric normalized Laplacian matrix, the above function expression can be structurally implemented using skip connection layers in a neural network.

[0068] ARMA(1,0) filter is an ARMA ( p , q One type of model, ARMA ( p , q The model can be approximated as:

[0069] ;

[0070] in, For ARMA ( p , q The transfer function of the model. These are characteristic roots. and These are the weighting coefficients. For order, and These are the orders of the moving average component and the autoregressive component, respectively. Transforming the above equation to the measurement domain, we obtain the domain expression:

[0071] ;

[0072] in, The output characteristics of the ARMA(1,0) filter are given. The oscillation characteristics of the node at the input of the ARMA(1,0) filter. For symmetric normalized Laplace matrix. ARMA( p , q The model will be based on the transfer function. In the symmetric normalized Laplace matrix Oscillation characteristics of nodes on the eigenvector basis Modulation is applied. The transfer function of the autoregressive moving average model is expressed as follows:

[0073] ;

[0074] In the above formula, This is the moving average portion, which can be used to control the zero point; This is the autoregressive part, which can be used to control the poles. and These are the orders of the moving average component and the autoregressive component, respectively. and The weighting coefficients are used. This autoregressive moving average model can adjust the gain at zeros and poles, and this structure can produce a filter-like effect. In this embodiment, the moving average and autoregressive moving average models have been incorporated into the power system topology structure. The autoregressive moving average model can model stochastic processes, corresponding to the propagation process of broadband oscillations. This gives the ARMA(1,0) filter both noise immunity, long-range dependency capture capability, and global graph structure awareness capability. By tracking measurement data and calculating poles, the oscillation frequency can be derived. With damping ratio The calculation formula is:

[0075] ;

[0076] ;

[0077] In the above formula, and express The i For conjugate eigenvalues, The sampling time and eigenvalues ​​determine the stability and frequency selectivity of the power system. To avoid delays in oscillation localization, this embodiment employs a recursive calculation method for the ARMA(1,0) filter. Specifically, the ARMA(1,0) filter is approximated using a first-order recursion as shown in the following formula:

[0078] ;

[0079] ;

[0080] in, and These represent the output characteristics and the oscillation characteristics of the input nodes of the ARMA(1,0) filter, respectively. and For coefficients, For symmetric normalized Laplace matrix The intermediate coefficients related to the eigenvalues, and These are the symmetric normalized Laplace matrices. Maximum and minimum eigenvalues It is an identity matrix. and This can be simplified to fixing the hyperparameters 1 and 0, making the intermediate coefficients... Approximately equal to the symmetric normalized Laplace matrix ;coefficient and While following the ARMA(1,0) model, and with the transfer function There is a mathematical connection; and Let represent the residuals and poles of a rational function, respectively. Coefficients and The parameters are optimized to be trainable to adapt to different oscillation characteristics; simultaneously, the autoregressive module runs iteratively and synchronously with the convolutional and pooling layers, forming an embeddable sequence autoregressive module, which facilitates dynamic feature extraction and solves the problem of tracing the source of multimodal oscillations. Furthermore, when applied to graph-based oscillation signals, the ARMA(1,0) filter function can be regarded as a low-pass filter. The embeddable sequence autoregressive module constructs a filter bank by stacking multiple ARMA(1,0) filters to adapt to various oscillation modes and achieve dynamic oscillation feature extraction. In this embodiment, the spectra of the input and output signals of the embeddable sequence autoregressive module are visualized, and the results are as follows: Figure 5 As shown, comparison Figure 5 The spectra of the input and output signals show that the features preserved by the embedded sequence autoregressive module are mainly concentrated in the low-frequency range, which effectively verifies the filtering performance of the embedded sequence autoregressive module.

[0081] In step S103 of this embodiment, the output feature will be... The power system topology graph is used as input to the multi-task graph sequence network. Information is transmitted through the multi-task graph sequence network, and the node characteristics of each node are obtained through the gated cyclic unit (GRU). At that time, the functional expression for information transmission in the multi-task graph sequence network is:

[0082] ;

[0083] in, and The first The output characteristics and the oscillation characteristics of the input nodes during the next transmission. For activation function, Adjacency matrix and identity matrix The sum of adjacency matrices To represent the unordered set of edges in a power system topology graph The adjacency matrix, for The degree matrix of the diagonal nodes, and The first The node feature matrix and trainable parameter matrix at each transmission are used. The purpose is to achieve symmetric normalization, thereby avoiding distortion of the feature vector scale.

[0084] In this embodiment, the gated recurrent unit (GRU) enhances the model's learning ability by selectively memorizing and forgetting information to model sequence measurement data. The GRU obtains the node features of each node. The function expression is:

[0085] ;

[0086] ;

[0087] in, For the softmax function, The normalized adjacency matrix, For activation function, The output of the gated recurrent unit (GRU) layer. and is a trainable parameter matrix.

[0088] In this embodiment, the loss function used by the multi-task graph sequence network during training has the following expression:

[0089] ;

[0090] ;

[0091] ;

[0092] ;

[0093] in, This is the loss function used during training of the multi-task graph sequence network. This is the location detection result of the oscillation source in the node-level detection branch. These are the model parameters for a multi-task graph sequence network. and These are the weight hyperparameters for the node-level and graph-level detection branches learned by the softplus function, respectively. To consider the weighted binary cross-entropy loss function that takes into account node mask loss, The cross-entropy loss is used for graph classification tasks in the graph-level detection branch. For the first The mask of each node, which serves as a flag indicating whether a node is valid. For the first The weight of each node, This represents the total number of measurement nodes. This represents the number of generator nodes. Let cross-entropy be the loss function. and These represent the detection results and ground truth values ​​for the graph classification task performed by the graph-level detection branch. In this embodiment, the loss function (multi-task loss function) includes the weighted binary cross-entropy loss function. Cross-entropy loss and a regularization term The weights of each loss term are not fixed hyperparameters but dynamically adaptive parameters, automatically learned and updated during the neural network training process. The multi-task loss function achieves dynamic updates of weights and parameters, with the loss values ​​of the two sub-branches calculated separately based on their respective performance, ensuring a balance between tasks during optimization and preventing a single task from dominating the training process. In this embodiment, the multi-task loss function introduces a loss weighting mechanism and node-level loss masking technology to focus the learning process on effective samples. Loss weighting mechanism (…) The calculation of loss masks assigns greater weights to classes with smaller sample sizes, thereby achieving a lower loss penalty. Node-level loss masking techniques ( The calculation of the loss function ensures that only valid nodes participate in the loss calculation during training. The main advantage of this masking mechanism is that when calculating the loss function, it ensures that only valid nodes participate in backpropagation and parameter optimization, prevents noise from interfering with irrelevant nodes, and avoids unlabeled nodes affecting gradient updates. This embodiment achieves subgraph and subnode recognition by introducing trainable parameters and a weight system. Each branch is designed to support a specific learning objective: the graph-level detection branch focuses on event type classification, while the node-level detection branch is dedicated to solving the problem of locating synchronous broadband oscillation sources. Information is transmitted through the multi-task graph sequence network to obtain the node features of each node; the node with the largest predicted value is the synchronous broadband oscillation source.

[0094] To verify the performance of the proposed Multi-Task Graph Sequence Network (MGSNet) in the power grid synchronous broadband forced oscillation detection and localization method of this embodiment, an ablation experiment was designed based on the IEEE 39-node system as the experimental object. The "node-graph convolutional network" represents modeling each node as an independent graph. The "masked graph convolutional network-gated recurrent unit model" and the "sequence graph sequence network" are both single-task frameworks, corresponding to a gated recurrent unit architecture with added masking mechanism and an architecture integrating an embeddable sequence autoregressive module, respectively. Based on the IEEE 39-node system, this embodiment selected three state-of-the-art existing methods and compared them with the proposed method. The experiments summarized six performance indicators: accuracy, F1 score, precision, area under the curve (AUC), event detection capability, and single-sample computation time. During the test, a noise level of 40 dB was superimposed on the original voltage and frequency oscillation measurement data. The results are shown in Table 1.

[0095] Table 1 Comparison of different existing methods based on the IEEE 39-bus system

[0096]

[0097] In Table 1, "-" indicates that the method does not have event type detection capability. As shown in Table 1, the multi-task graph sequence network proposed in this embodiment outperforms other methods in terms of accuracy, precision, area under the curve (AUC), and F1 score. Furthermore, its average runtime on a single task is only 11.2 ms, demonstrating its real-time localization capability.

[0098] Furthermore, experiments were conducted on the multi-task graph sequence network proposed in this embodiment based on the IEEE 39-node system. Different methods were used to locate and analyze the oscillations of different levels of noise, and the results are shown in Table 2.

[0099] Table 2. Experimental results of oscillation source localization using different methods and noise levels based on the IEEE 39-bus system.

[0100]

[0101] Table 2 compares the F1 score and TOP-1 accuracy of different methods under different noise levels based on the IEEE 39-node system. The F1 score, by coordinating precision and recall, serves as a comprehensive evaluation index of localization quality; the TOP-1 accuracy directly reflects the correctness of the highest confidence prediction. As shown in Table 2, the multi-task graph sequence network proposed in this embodiment outperforms other methods in broadband oscillation localization tasks under different noise levels, achieving state-of-the-art performance and realizing the optimal balance in distinguishing between oscillation sources and non-oscillation sources.

[0102] In summary, the method in this embodiment achieves end-to-end processing from raw measurement data to event detection and source localization using a multi-task graph sequence network (MGSNet), eliminating the need for additional preprocessing modules, simplifying the process, and improving real-time performance. The method introduces an embeddable sequence autoregressive (SA) block, combined with the implementation of an autoregressive moving average model in the graph domain, effectively filtering noise, extracting key oscillation features, and maintaining high accuracy at noise levels of 20–60 dB. The method simultaneously performs event type classification and oscillation source localization through multi-task learning, and introduces a customized loss function and masking mechanism, significantly improving the model's discriminative ability and generalization performance. Based on a graph convolutional network (GCN), combined with a gated recurrent unit (GRU) and an embeddable sequence autoregressive block, the method effectively integrates spatiotemporal features and topological information, enhancing the ability to model oscillation propagation in complex power grids. The method achieves autoregressive moving average filtering through recursive approximation, combined with a lightweight graph neural network structure, requiring only 11.2 ms (3 seconds of data) for a single sample processing time, demonstrating good engineering applicability. This embodiment's method can fully capture the temporal and spatial information of electrical parameters at power grid nodes, constructing a sequence-based autoregressive module and an autoregressive moving average model. This allows for the simultaneous capture of short-term and long-term dynamic characteristics, extracting and integrating oscillation-related features and physical features to achieve noise suppression and significant feature enhancement. The autoregressive moving average model is embedded within the sequence autoregressive module, possessing both embedded and pluggable characteristics. Furthermore, the autoregressive module iteratively runs synchronously with convolutional and pooling layers, forming an embeddable sequence autoregressive module that facilitates dynamic feature extraction and solves the challenge of tracing the source of multimodal oscillations. A multi-task mechanism enables simultaneous processing of oscillation detection and localization tasks, and an improved multi-task loss function minimizes the objective function value while adaptively balancing performance to achieve oscillation source tracing. This embodiment's method effectively reduces interference from other information on broadband oscillation information, effectively solves the challenge of multi-task broadband oscillation localization, and achieves rapid and accurate localization of broadband oscillation sources, possessing significant engineering practical value.

[0103] This embodiment also provides a power grid synchronous broadband forced oscillation detection and location system, including a microprocessor and a memory interconnected, wherein the microprocessor is programmed or configured to execute the power grid synchronous broadband forced oscillation detection and location method.

[0104] This embodiment also provides a computer-readable storage medium storing a computer program or instructions that are programmed or configured to execute the power grid synchronization broadband forced oscillation detection and location method by a processor.

[0105] This embodiment also provides a computer program product, including a computer program or instructions, which are programmed or configured to execute the power grid synchronization broadband forced oscillation detection and location method via a processor.

[0106] Those skilled in the art will understand that the technical solutions provided by this invention may take the form of a method, system, or computer program product. Therefore, this invention may take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this invention may take the form of a computer program product embodied on one or more computer-readable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, produce an implementation of the flowchart. Figure 1 One or more processes and / or boxes Figure 1 The computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The functions specified in one or more boxes. These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable apparatus for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0107] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principle of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A method for detecting and locating broadband forced oscillations in a power grid, characterized in that, The process includes the following steps: acquiring an oscillation data sequence of the power grid and a power system topology map, wherein the oscillation data sequence includes the voltage, frequency, and rate of change of the power grid, and the nodes in the power system topology map are generators or loads in the power grid, the edges are lines, and the oscillation characteristics of the nodes include the voltage, frequency, and rate of change of the nodes. The oscillation data sequence of the power grid and the power system topology map are input into a pre-trained multi-task graph sequence network to obtain the oscillation source localization and detection results, including: S101 extracts key features from the oscillating data sequence through convolutional and pooling layers. ; S102, key features Input can be embedded into a sequence autoregressive module, the embedded sequence autoregressive module through Stacked ARMA(1,0) filters for key features Perform embedded feature filtering and... The final output feature is obtained by summing the responses of the stacked ARMA(1,0) filters. ; S103 will output features The power system topology graph is used as input to the multi-task graph sequence network. Information is transmitted through the multi-task graph sequence network, and the node characteristics of each node are obtained through the gated cyclic unit (GRU). ; S104, The node characteristics of each node. Input graph-level detection branch and node-level detection branch respectively. Use graph-level detection branch to classify the event type of the oscillation source and use node-level detection branch to locate the oscillation source. In step S102, for The final output feature is obtained by summing the responses of the stacked ARMA(1,0) filters. The function expression is: ； in, Let be the number of stacking levels of the ARMA(1,0) filter. For activation function, For the first The output characteristics of an ARMA(1,0) filter. For the first Input characteristics of an ARMA(1,0) filter. and For the first Trainable weight parameters of an ARMA(1,0) filter. As a key feature, The normalized Laplace matrix is ​​symmetric; the ARMA(1,0) filter is approximated using a first-order recursive method as shown in the following equation: ； ； in, and These represent the output characteristics and the oscillation characteristics of the input nodes of the ARMA(1,0) filter, respectively. and For coefficients, For symmetric normalized Laplace matrix The intermediate coefficients related to the eigenvalues, and These are the symmetric normalized Laplace matrices. Maximum and minimum eigenvalues It is the identity matrix; The loss function used during training of the multi-task graph sequence network is expressed as follows: ； ； ； ； in, This is the loss function used during training of the multi-task graph sequence network. This is the location detection result of the oscillation source in the node-level detection branch. These are the model parameters for a multi-task graph sequence network. and These are the weight hyperparameters for the node-level and graph-level detection branches learned by the softplus function, respectively. To consider the weighted binary cross-entropy loss function that takes into account node mask loss, The cross-entropy loss is used for graph classification tasks in the graph-level detection branch. For the first The mask of each node, which serves as a flag indicating whether a node is valid. For the first The weight of each node, This represents the total number of measurement nodes. This represents the number of generator nodes. Let cross-entropy be the loss function. and These are the detection results and ground truth values ​​for the graph classification task performed by the graph-level detection branch.

2. The method for detecting and locating broadband forced oscillations in power grids according to claim 1, characterized in that, The process of acquiring the oscillation data sequence and power system topology map of the power grid includes acquiring the voltage time series of the power grid, filtering the voltage time series to eliminate the DC component to extract the frequency of the power grid, and calculating the frequency change rate based on the frequency of the power grid and the following formula: ； in, and Sampling points and sampling points frequency, The sampling interval is... The voltage sampling frequency is used, and the obtained voltage, frequency, and rate of change of the power grid are normalized to form an oscillation data sequence.

3. The method for detecting and locating broadband forced oscillations in power grids according to claim 1, characterized in that, When acquiring the oscillation data sequence of the power grid and the power system topology map, the functional expression of the obtained power system topology map is: ； in, This is a power system topology diagram. For a set of nodes, It is an unordered set of edges. The oscillation characteristics of a node include its voltage, frequency, and rate of change of frequency, and the unordered edge set. Through the adjacency matrix This represents the line connection relationship between different nodes. The node and the first When nodes are directly connected by a line, the adjacency matrix The Middle Line number Column elements ;otherwise , indicating the first The node and the first The nodes have no direct connection.

4. The method for detecting and locating broadband forced oscillations in power grids according to claim 1, characterized in that, In step S101, key features are extracted from the oscillating data sequence through convolutional and pooling layers. The function expression for the convolution operation of the convolutional layer is: ； In the above formula, This is the output of the convolution operation. The number of types of oscillating data. for The first moment The input signal for the oscillation data; convolution kernel The offset time that is applied, and The first moment Input signal of oscillation data Time axis alignment; convolution kernel For the first Such oscillation data in The convolution kernel corresponding to each time step.

5. The method for detecting and locating broadband forced oscillations in power grids according to claim 1, characterized in that, In step S103, the output features will be... The power system topology graph is used as input to the multi-task graph sequence network. Information is transmitted through the multi-task graph sequence network, and the node characteristics of each node are obtained through the gated cyclic unit (GRU). At that time, the functional expression for information transmission in the multi-task graph sequence network is: ； in, and The first The output characteristics and the oscillation characteristics of the input nodes during the next transmission. For activation function, Adjacency matrix and identity matrix The sum of adjacency matrices To represent the unordered set of edges in a power system topology graph The adjacency matrix, for The degree matrix of the diagonal nodes, and The first The node feature matrix and trainable parameter matrix are obtained during each transmission, and the node features of each node are obtained through a gated recurrent unit (GRU). The function expression is: ； ； in, For the softmax function, The normalized adjacency matrix, For activation function, The output of the gated recurrent unit (GRU) layer. and is a trainable parameter matrix.

6. A power grid synchronous broadband forced oscillation detection and positioning system, comprising a microprocessor and a memory interconnected, characterized in that, The microprocessor is programmed or configured to execute the power grid synchronization broadband forced oscillation detection and location method according to any one of claims 1 to 5.

7. A computer-readable storage medium storing a computer program or instructions, characterized in that, The computer program or instructions are programmed or configured to execute the power grid synchronization broadband forced oscillation detection and location method according to any one of claims 1 to 5 via a processor.

8. A computer program product, comprising a computer program or instructions, characterized in that, The computer program or instructions are programmed or configured to execute the power grid synchronization broadband forced oscillation detection and location method according to any one of claims 1 to 5 via a processor.

Images