Subsynchronous oscillation tracing method and system based on high-dimensional deterministic learning modeling

Abstract

The invention provides a subsynchronous oscillation traceability method and system based on high-dimensional deterministic learning modeling, and belongs to the technical field of power system stabilization, and the method comprises the steps: constructing graph structure data based on a power signal of a power grid, and employing a graph theory method to extract topological information from the graph structure data; for each node in the graph structure data, carrying out distributed dynamic modeling on a power signal corresponding to the node by adopting a deterministic learning mechanism, and constructing a pattern library; establishing a distributed dynamic estimator by using the neural network weight stored in the pattern library, inputting a to-be-detected power signal to the distributed dynamic estimator in a parallel mode at each topological node, and comparing the to-be-detected power signal with all patterns in the pattern library through a distributed dynamic pattern recognition mechanism to obtain a recognition residual error at each node; and carrying out real-time aggregation on the topological information of the identification residual generated at each topological node by adopting a graph convolution method to obtain a traceability result. According to the invention, the accuracy and reliability of subsynchronous oscillation detection are improved.

Description

A method and system for tracing the source of subsynchronous oscillations based on high-dimensional deterministic learning modeling Technical Field

[0001] This invention belongs to the field of power system stability technology, and particularly relates to a subsynchronous oscillation source tracing method and system based on high-dimensional deterministic learning modeling. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Subsynchronous oscillations (SSOs) are a significant aspect of power system stability, posing a serious threat to the safe and stable operation of generator shaft systems and the power system itself. With the large-scale grid connection of new energy sources such as wind farms, the penetration rate of renewable energy in the power system is continuously increasing, making SSOs a critical issue threatening power system stability. SSOs not only threaten the safe and stable operation of the power system but can also lead to damage to generating equipment and a decline in power quality. Therefore, accurately identifying and locating the source of SSOs through effective source tracing methods is crucial for providing decision-making support for system operators, enabling the implementation of effective suppression measures, and ensuring the safe operation of the power system.

[0004] Existing methods for tracing the source of subsynchronous oscillations mainly include model-based methods and measurement data-based methods. Model-based methods require complete system parameters and accurate mathematical models. While they can provide good analysis of physical mechanisms, the confidentiality of models by equipment manufacturers often makes accurate models difficult to obtain in practical applications. Furthermore, if applied to large-scale power grids, the high-dimensional system leads to an enormous computational burden for tracing tasks, making them unsuitable for online tasks. Measurement data-based methods are sensitive to data quality and noise, and have limitations when processing strongly time-varying signals. Machine learning methods learn from power system measurement data, automatically extracting discriminative features and learning discriminative models through training data, thereby achieving the identification and location of subsynchronous oscillation sources. However, these methods lack interpretability, severely hindering their practical application in power systems with extremely high safety and reliability requirements. To improve interpretability, existing technologies require very high-dimensional power system measurement data, causing the parameters of the neural network used for modeling to grow exponentially. Summary of the Invention

[0005] In response to the challenges posed by the ever-expanding scale of power grids, the high dimensionality of power system measurement data, the dramatic increase in neural network parameters resulting from deterministic learning modeling of high-dimensional data, the difficulty of traditional deterministic learning methods in handling high-dimensional measurement data from wind power systems, and the inability of existing dynamic pattern recognition methods to effectively address the unique characteristics of wind power system graph structure data, this invention proposes a subsynchronous oscillation source tracing method and system based on high-dimensional deterministic learning modeling. This method effectively processes high-dimensional data, reduces algorithm complexity, and utilizes topological information to achieve accurate identification and location of subsynchronous oscillation sources, thereby improving the accuracy and reliability of subsynchronous oscillation detection and providing technical support for the safe and stable operation of power systems.

[0006] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions: Firstly, the present invention discloses a subsynchronous oscillation tracing method based on high-dimensional deterministic learning modeling, comprising: constructing graph structure data based on preprocessed power grid signals, and extracting topological information from the graph structure data using graph theory methods; performing distributed dynamic modeling of the power signal corresponding to each node in the graph structure data using a deterministic learning mechanism and constructing a pattern library; building a distributed dynamic estimator using the neural network weights stored in the pattern library; inputting the power signal to be measured into the distributed dynamic estimator in parallel at each topological node, comparing it with all patterns in the pattern library through a distributed dynamic pattern recognition mechanism to obtain the recognition residual at each node; and aggregating the topological information of the recognition residual generated at each topological node in real time using a graph convolution method to obtain the tracing result.

[0007] Secondly, this invention discloses a subsynchronous oscillation tracing system based on high-dimensional deterministic learning modeling, comprising: a data acquisition module, used to construct graph structure data based on preprocessed power grid signals, and extract topological information from the graph structure data using graph theory methods; a distributed deterministic learning module, used to perform distributed dynamic modeling of the power signal corresponding to each node in the graph structure data using a deterministic learning mechanism and construct a pattern library; a distributed dynamic recognition module, used to build a distributed dynamic estimator using the neural network weights stored in the pattern library, inputting the power signal to be measured into the distributed dynamic estimator in parallel at each topological node, comparing it with all patterns in the pattern library through a distributed dynamic pattern recognition mechanism to obtain the recognition residual at each node; and an aggregation tracing module, used to aggregate the topological information of the recognition residual generated at each topological node in real time using a graph convolution method to obtain the tracing result.

[0008] Thirdly, the present invention discloses an electronic device, including a memory and a processor, and computer instructions stored in the memory and running on the processor, wherein the computer instructions, when run by the processor, complete the steps of the above-mentioned subsynchronous oscillation source tracing method based on high-dimensional deterministic learning modeling.

[0009] Fourthly, the present invention discloses a computer-readable storage medium for storing computer instructions, which, when executed by a processor, complete the steps of the above-described subsynchronous oscillation source tracing method based on high-dimensional deterministic learning modeling.

[0010] Compared with existing technologies, the beneficial effects of this invention are as follows: The subsynchronous oscillation tracing method based on high-dimensional deterministic learning modeling and recognition in this invention creatively constructs a three-in-one subsynchronous oscillation tracing system of "physical topology-driven, distributed learning, and coupled information decision-making". Compared with model methods that rely solely on data-driven approaches, this invention considers the physical topology connection of the power grid while leveraging the advantages of existing data methods, providing a subsynchronous oscillation tracing method with a certain degree of physical understanding.

[0011] This invention innovates a distributed deterministic learning and dynamic pattern recognition architecture, enabling the construction of node-level RBF neural networks. This invention effectively solves the parameter explosion problem encountered by traditional deterministic learning methods in high-dimensional data modeling by transforming centralized processing into distributed learning, thereby reducing computational complexity.

[0012] This invention effectively addresses the coupling relationships within power systems by deeply integrating distributed training and topology information fusion. It establishes a dynamic correlation between node oscillation patterns and network-wide oscillations through topology information structures, and provides the final oscillation source location result at the decision-making level, thus significantly improving source tracing accuracy.

[0013] The method of this invention adopts a distributed framework that enables local deployment at the node level. It can choose to centrally model and diagnose at the main detection node or to deploy and integrate communication decisions at the edge, effectively reducing the computational cost of centralized processing.

[0014] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0015] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0016] Figure 1 is a flowchart of the subsynchronous oscillation source tracing method based on high-dimensional deterministic learning modeling as described in Embodiment 1 of the present invention.

[0017] Figure 2 is a schematic diagram of the topology structure described in Embodiment 1 of the present invention.

[0018] Figure 3 is a schematic diagram of the distributed dynamic pattern library described in Embodiment 1 of the present invention. Detailed Implementation

[0019] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0020] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of the present invention.

[0021] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0022] As should be understood in Example 1, deterministic learning theory, as an effective modeling and learning method for complex nonlinear systems, has attracted widespread attention from researchers. Utilizing limited time-series data, deterministic learning can achieve locally accurate modeling of the dynamics of unknown systems. Compared to general machine learning methods, deterministic learning has significant advantages in interpretability, providing clear explanations of physical mechanisms. Furthermore, dynamic pattern recognition algorithms designed based on deterministic learning methods have shown great advantages in various dynamic scenarios, such as human gait recognition, aircraft engine surge detection, and heartbeat classification. However, high-dimensional data modeling has always been a key issue restricting the development of deterministic learning methods, as it leads to an exponential increase in the number of neural network parameters used in deterministic modeling. Currently, the dimensionality of measurement data in power systems is often very high, and the graph structure data generated by the physical topology of power systems presents new challenges to dynamic pattern recognition methods designed for Euclidean data.

[0023] This invention proposes a distributed deterministic learning and pattern recognition strategy to solve the problem of high-dimensional data modeling. Furthermore, it creatively introduces a topological structure in the identification stage to aggregate and fuse distributed identification information, so that the identification results can contain full-dimensional information and efficiently locate the source of subsynchronous oscillations.

[0024] Topological processing is performed using graph neural networks (GNNs). A GNN is a neural network framework specifically designed for processing graph structures (i.e., topological data). It can efficiently learn complex topological relationships in non-Euclidean data and is an important means of uncovering hidden information behind these relationships. Based on whether the convolution operation operates in the spectral or spatial domain, GNN methods can be divided into two categories: spectral methods and spatial methods. This invention uses graph convolution to process topological information. The principle of graph convolution is as follows: A graph consists of nodes and edges, which can be represented by an adjacency matrix. and node feature matrix It means that among them Indicates the number of nodes. This represents the feature dimension of each node. In the adjacency matrix... Represents a node and nodes Are there any edges connecting them? The core idea of ​​graph convolution is to update the features of the current node by aggregating information from neighboring nodes. The GCN of a layer can be represented by the following formula: (1) Among them, This represents the normalized adjacency matrix. Self-loops are added to allow each node to participate in the aggregation as well. At the same time, the features of neighboring nodes are weighted and averaged to prevent the aggregation results from being biased due to different node degrees. It is the first The feature vector of the layer, These are neural network parameters. This represents the activation function.

[0025] In one or more embodiments, the present invention discloses a subsynchronous oscillation source tracing method based on high-dimensional deterministic learning modeling. It explores new approaches to deterministic learning modeling and recognition from the perspective of knowledge utilization. By utilizing the idea of ​​topological information aggregation in the distributed deterministic learning and knowledge utilization stages, a dynamic pattern recognition framework based on topological information fusion is proposed, which transforms the complexity of the algorithm from exponential growth to linear growth. On this basis, the subsynchronous oscillation source is located, as shown in Figure 1, including the following steps: Step S1: Acquire and preprocess the power grid signal.

[0026] Step S1-1: Acquire power grid signals: Acquire measurement data including voltage signals, current signals, reactive power and active power signals at the power system output.

[0027] The collected data needs to include measurement data of various subsynchronous oscillations, which cover the subsynchronous oscillation characteristics of the system under different operating conditions.

[0028] Step S1-2, Data Filtering: Use a suitable filter to filter the acquired power signal, remove the power frequency component and high-frequency noise, and retain the subsynchronous component to reduce the interference of the power frequency signal on the extraction of subsynchronous oscillation features.

[0029] The filter consists of two parts: a pre-filter with a 50Hz notch filter and a post-filter with a low-pass filter to simultaneously filter out the power frequency component and high-frequency noise signals. It is also necessary to ensure that signals in the subsynchronous oscillation frequency band are not filtered out. The purpose of filtering the power signal to be diagnosed is to remove noise interference from the signal.

[0030] A cascaded "notch filter + low-pass filter" scheme is adopted: First, the power frequency component is filtered out using a band-stop filter / notch filter. A Butterworth filter is used to design a band-stop filter specifically to filter out the 50Hz power frequency component. Then, a low-pass filter with a cutoff frequency of 45Hz is designed to suppress high-frequency noise.

[0031] Step S1-3, Dimensionless Parameter Processing: The power signal after filtering out the power frequency component is processed into a dimensionless parameter. A sliding window normalization method is used to adjust the signal amplitude to a fixed range, converting it into a dimensionless form to facilitate subsequent modeling and analysis.

[0032] For the filtered power signal To handle the real-time data stream during the tracing process, the following operations are performed to obtain the results.

[0033] (2) Among them, , This represents the mean and standard deviation of all data within the current sliding window. This represents the scaling factor, which is taken here. .

[0034] It should be understood that although the waveform of an electrical signal changes after it is converted into dimensionless data, its inherent laws and characteristics remain unchanged.

[0035] The purpose of dimensionless normalization in this embodiment is to standardize signal data within a certain range, facilitating the subsequent determination of neuron placement during the learning and modeling process. Simultaneously, this normalization process eliminates the influence of dimensional differences between different signals, reducing model complexity and improving model stability and generalization ability.

[0036] Step S2: Construct graph structure data based on the preprocessed power grid signal, and extract topological information from the graph structure data using graph theory methods.

[0037] As shown in Figure 2, the processed high-dimensional power data is abstracted into graph structure data. Graph theory methods and the topological connection relationships of each power plant in the power grid are used to construct nodes in the data. Each power plant is regarded as a node in the topological graph, and the edges between nodes represent the interconnection relationships between power plants. The topological information of the power system is extracted to provide a foundation for subsequent distributed deterministic learning and recognition.

[0038] Step S3: For each node in the graph structure data, a deterministic learning mechanism is used to perform distributed dynamic modeling of the power signal corresponding to the node and construct a pattern library.

[0039] Step S3-1: Distributed dynamics modeling means that each node is modeled individually, rather than centrally, which helps reduce computational complexity. This reduces computational complexity from... Reduced to .

[0040] For high-dimensional power systems, the following nonlinear dynamics can be used to represent them: (3) Among them, A vector composed of state variables in a power system, such as voltage, current, active power, and reactive power. For the internal parameters of this nonlinear dynamic system, The dynamic equations of a nonlinear system This is the time derivative of the system state variable.

[0041] After abstracting into a graph structure, the dynamic relationships at different nodes of the above high-dimensional nonlinear dynamic system can be represented by the following distributed system: (4) Among them, For the node subsystem, For the subsystem index, Represents its nonlinear dynamic function, It is a subsystem Given the neighborhood set, the state equations of each node subsystem after decoupling are as follows: (5) Among them, Let represent the state vector of the nonlinear dynamic system (4), and These represent the two phases and active power of the current output to the grid from a subsystem (power plant node). Subscript 1 represents the current dynamics of the d-axis, subscript 2 represents the current dynamics of the q-axis, and subscript 3 represents the dynamics of the system's active power.

[0042] This represents the intrinsic dynamics of a power signal within a power system. Represents the system parameter vector. For subsystem State variables in the middle, For subsystem Internal parameters, They are nodes The current dynamics representation of three-phase currents along the d and q axes after Park transformation. This is the dynamic representation of the system's active power.

[0043] To address the high-dimensional characteristics of the overall power grid system, deterministic learning algorithms are deployed at each power plant node to achieve node-level dynamic modeling. By effectively decoupling the complex coupling relationships between various variables in the power system, efficient deterministic learning modeling is performed using a distributed computing framework.

[0044] With the first Current state variables in each subsystem For example, we can use the Euler model for discretization: (6) In order to further accurately model and represent the internal dynamics of the power system, a dynamic neural network identifier is then constructed: (7) Among them, It represents the state of the neural network identifier, indicating an estimate of the current state. It is the gain of the identifier to be designed. This represents an RBF neural network used to learn system dynamics. It is the weight vector of the neural network.

[0045] The weights of the RBF neural network are optimized according to the following update rule: (8) Among them, It is the learning rate used to train the network, and the tracking error. Defined as .

[0046] By utilizing the aforementioned dynamic neural network identifier and the neural network weight update law based on energy descent design, the internal unknown dynamics of the decoupled subsystems of the power system can be learned along the trajectory of the power signal. Furthermore, this dynamic knowledge can be stored in a space-time invariant form in a constant RBF neural network. middle: (9) Among them, The mean of the neural network at a node over a period of time after convergence, and the error term. Indicates modeling error. It is the regression vector composed of Gaussian radial basis functions.

[0047] That is, for the entire system, it can be written as: (10) Among them, This represents the system dynamics at node n after modeling. This represents the weights at node n after the neural network converges, stored in a time-invariant form. It is the regression vector composed of Gaussian radial basis functions.

[0048] This embodiment employs a distributed deterministic learning mechanism to model high-dimensional power data, achieving efficient extraction of subsynchronous oscillation characteristics. Addressing the high-dimensional characteristics of the overall power grid system, deterministic learning algorithms are deployed at each power plant for node-level dynamic modeling. This decouples the complex coupling relationships between power system variables through distributed computation, effectively improving modeling efficiency and accuracy. By using distributed modeling and parallel training of high-dimensional data, the training data for deterministic learning is distributed across different computing nodes, eliminating dependence on global power grid signals and effectively improving modeling efficiency. This dependency is then reused during the knowledge utilization phase.

[0049] Step S3-2: For each topology node, as shown in Figure 3, construct a distributed subsynchronous oscillation pattern library. Store the modeling weights obtained from the above distributed deterministic learning into the distributed dynamic pattern library. The pattern library is constructed as follows: (11) Among them, Represents a node The pattern library at the location, Indicates the first The first in the class There are several patterns. The number of patterns in each of the above pattern libraries is not required to be the same, which can handle situations where the categories are imbalanced.

[0050] This embodiment acquires representative dynamic patterns of power signals through training, and stores the dynamic knowledge of power grid signals learned by distributed deterministic learning in the form of constant RBF neural network weights in a distributed pattern library, so as to enable efficient reuse of knowledge in subsequent subsynchronous oscillation tracing processes.

[0051] This embodiment employs a distributed modeling method that fully considers the independence and interrelationship of each node, enabling more accurate capture of the dynamic behavior of each power plant during subsynchronous oscillations. After distributed modeling, each node generates a pattern library containing rich subsynchronous oscillation information, which provides crucial reference templates for subsequent fault diagnosis.

[0052] Step S4: At each topology node, the power grid signal to be processed is compared with all patterns in the pattern library through a distributed dynamic pattern recognition mechanism to obtain the recognition result at each node.

[0053] Step S4-1: Design a distributed dynamic estimator.

[0054] During the real-time source tracing phase, a distributed schema library is utilized. A distributed dynamic estimator is constructed using the RBF neural network weights stored in the database. The input to the dynamic estimator is the power signal to be identified, and the output is the state variable of the dynamic estimator. Specifically, at each distributed node, the RBF neural network is reconstructed using all the pattern weights in the pattern library of that node, and the dynamic estimator is constructed based on this.

[0055] For each distributed node, the RBF neural network is reconstructed using the weights of all patterns in the pattern library, and a dynamic estimator is built based on this. The dynamic estimator is constructed as follows: (12) Among them, nodes The input to the dynamic estimator is the power signal to be identified. The output is the state variables of the dynamic estimator. The output of this dynamic estimator reflects the effect of the measured power signal on the node. The first in the pattern library The first in the class The response of an intrinsic dynamic mode within a mode. It is the dynamic estimator gain, a hyperparameter less than 1. This represents the dynamic knowledge stored in the model library in step S3. This represents the sensor sampling time interval.

[0056] Therefore, the output of each dynamic estimator corresponds to the response of the measured power data to the intrinsic dynamic pattern of each power signal in the pattern library. Accordingly, the dynamic estimators deployed on each node can reproduce the dynamic knowledge obtained from distributed deterministic learning modeling in parallel, enabling efficient real-time source tracing tasks.

[0057] Step S4-2: For each power network topology node, the power signal of the power grid to be tested is input into the distributed dynamic estimator in step S4-1 in parallel. The signal is compared with all test patterns in the pattern library through the distributed dynamic pattern recognition mechanism to obtain the recognition residual at each node.

[0058] The difference between the output generated by the dynamic estimator and the power signal pattern to be diagnosed is used to obtain the recognition residuals for different training patterns in the pattern library: (13) Among them, It is the first in the pattern library The first in the class The time corresponding to each mode Residual identification at any given moment.

[0059] The magnitude and characteristics of the residual can reflect the degree of deviation between the measured signal and the normal pattern in the pattern library. The smaller the dynamic difference, the smaller the residual, which provides a quantitative indicator for the identification and localization of subsynchronous oscillation sources.

[0060] Step S5: The identification residuals generated at each topology node are aggregated in real time using graph convolution to obtain the source tracing result.

[0061] In the topology information fusion and identification stage, the identification residuals generated by the power signals at each topology node are fused in real time. First, a sliding window is set to process the output residual results. The norm is used to fuse the residual results corresponding to the same category of dynamic modes. Then, graph convolution is applied to the fused residuals to take into account the topological relationships between nodes in the power system topology graph, so as to effectively cluster the identification results at different nodes. The clustered results... The fusion residual with the smallest norm indicates that the power signal dynamics pattern is most similar to the category in the corresponding pattern library. By aggregating topological information in this way, the source tracing results after aggregation are provided in real time.

[0062] Step S5-1: Based on the recognition residuals generated in the above steps, calculate the output results. The norm is used to fuse the residuals corresponding to the same category of dynamic modes. By fusing the residual identification results of various categories, the fused result... The results corresponding to each pattern are unified into a single pattern: (14) Among them, Indicates at time step The smallest one selected A set of indexes for each pattern. Let represent the category recognition residuals after fusion at the i-th node. To improve the effectiveness of recognition, an average averaging process is first applied to it. Instead of directly using recognition error to make decisions, we use the norm. It is a preset time window length, which is less than the time length of the residual sequence.

[0063] The residual matrix at each node is represented as follows: (15) Wherein, matrix The row vector represents the node The residual sequences after fusion of various classes are given, and the column vectors represent the residuals of all class recognitions at time t. The residual value at that point.

[0064] The topological feature matrices of the identification residuals at all the above nodes are as follows: (16) By reconstructing the identification residual results, Represents the node feature matrix, where each node is in Time and a The eigenvectors of the dimension are correlated. Let n be the type a residual of node n at time step k.

[0065] Step S5-2: Using graph convolution to identify the residual topology feature matrix, create the system's adjacency matrix based on the topological connectivity between power grid nodes. Then, the graph convolution method is used to update the features of the current node by aggregating the information of neighboring nodes. After the update... The graph convolutional network of layers is represented as follows: (17) Among them, Indicates the first The topological feature matrix of the layer Indicates the first The input of the layer. It is the activation function of the network. These are learnable parameters. It is the renormalized adjacency matrix. It is a degree matrix, representing the number of edges connecting a given node to other nodes. It is an N-order identity matrix.

[0066] Preferably, a supervised graph convolutional neural network is trained using labeled data, and the weighted model is saved for testing the source localization of power data. This invention considers a learning-free domain aggregation method. To preserve the interpretability of deterministic learning and dynamic pattern recognition, a second aggregation scheme is proposed by simplifying the above operations to effectively aggregate the recognition results at different nodes.

[0067] (18) Among them, Topological feature matrix after The result after topological aggregation.

[0068] Furthermore, the row vector of a certain node after aggregation is taken as the final recognition residual.

[0069] (19) Among them, The operation involves obtaining the eigenvectors of a specified row of the aggregated residual matrix at each time step. The aggregated identification residuals... The smallest fusion residual category indicates that the power signal dynamics pattern is most similar to the category in the corresponding pattern library, thus obtaining the source tracing result. By aggregating topological information in this way, the aggregated source tracing result is given in real time.

[0070] From a knowledge utilization perspective, this embodiment addresses the high-dimensional modeling problem of power systems by topologically aggregating the distributed identification results during the identification phase. Using graph convolution to aggregate topological information enhances the global perception of subsynchronous oscillations. The source-tracing signal processed by the graph neural network integrates the topological information of the power nodes themselves, neighboring nodes, and the network, providing a more comprehensive and accurate basis for the source-tracing analysis of subsynchronous oscillations.

[0071] Preferably, the topological information fusion (aggregation) method of the present invention includes using a general graph neural network algorithm, as well as a learning-free method, using a standardized adjacency matrix to perform graph convolution operation, and taking the feature of one of the aggregated nodes as a discriminative identification residual.

[0072] Through the above technical solution, this invention first acquires power system measurement data containing various subsynchronous oscillation modes. Secondly, these measurement data are abstracted into a graph structure based on the topological connections in the actual power grid, with each power plant considered a node and edges between nodes representing topological connections within the grid. Then, for each node in the graph structure, a deterministic learning mechanism is used to perform distributed dynamic modeling of the power data corresponding to that node. The distributed modeling results generate a pattern library containing various subsynchronous oscillation information for each node. Furthermore, a dynamic estimator is constructed at each topological node in the graph structure. The power signal to be diagnosed is filtered and normalized before being input into the dynamic estimator. The dynamic estimator compares and analyzes the output results with the dynamic patterns in the node pattern library one by one to generate identification residuals. Finally, multi-node topological information fusion is performed using a graph convolution method. This fully utilizes the topological relationships between nodes in the graph structure to effectively aggregate and fuse the identification results at different nodes, and provides a real-time source tracing signal for each power node.

[0073] The method for tracing the source of power grid signals in this invention is based on a deterministic learning mechanism. Deterministic learning theory provides an effective method for addressing the challenges of knowledge acquisition, representation, and utilization in dynamic systems. Under the condition of continuous excitation (PE), the deterministic learning mechanism provides a solution for the accurate local identification or modeling of nonlinear dynamic systems. The modeled dynamics can be represented and stored by a constant RBF neural network, and this stored knowledge can be effectively used for tasks in dynamic environments, such as dynamic pattern recognition and intelligent control.

[0074] Example 2 discloses a subsynchronous oscillation tracing system based on high-dimensional deterministic learning modeling in one or more embodiments. Specifically, it includes: a data acquisition module for constructing graph structure data based on preprocessed power grid signals and extracting topological information from the graph structure data using graph theory methods; a distributed deterministic learning module for performing distributed dynamic modeling of the power signal corresponding to each node in the graph structure data using a deterministic learning mechanism and constructing a pattern library; a distributed dynamic recognition module for building a distributed dynamic estimator using neural network weights stored in the pattern library, inputting the power signal to be measured into the distributed dynamic estimator in parallel at each topological node, comparing it with all patterns in the pattern library through a distributed dynamic pattern recognition mechanism to obtain the recognition residual at each node; and an aggregation tracing module for aggregating the topological information of the recognition residuals generated at each topological node in real time using a graph convolution method to obtain the tracing result.

[0075] Example 3 This example provides an electronic device, including a memory and a processor, as well as computer instructions stored in the memory and running on the processor. When the computer instructions are executed by the processor, they complete the steps of the above-described subsynchronous oscillation tracing method based on high-dimensional deterministic learning modeling.

[0076] Example 4 This example provides a computer-readable storage medium for storing computer instructions, which, when executed by a processor, complete the steps of the above-described subsynchronous oscillation source tracing method based on high-dimensional deterministic learning modeling.

[0077] 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, create means for implementing the functions specified in one or more blocks of the flowchart illustrations and / or one or more blocks of the block diagrams.

[0078] These 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 function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means that implement the functions specified in one or more flowcharts and / or one or more block diagrams.

[0079] These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to perform a series of operational steps on the computer or other programmable apparatus to produce a computer-implemented process, such that the instructions, which execute on the computer or other programmable apparatus, provide steps for implementing the functions specified in one or more flowcharts and / or one or more block diagrams.

[0080] The descriptions of each embodiment in the above embodiments have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0081] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A subsynchronous oscillation source tracing method based on high-dimensional deterministic learning modeling, characterized in that, include: Graph structure data is constructed based on the preprocessed power grid signals, and topological information is extracted from the graph structure data using graph theory methods. For each node in the graph structure data, a deterministic learning mechanism is used to perform distributed dynamic modeling of the power signal corresponding to the node and construct a pattern library. A distributed dynamic estimator is built using the neural network weights stored in the pattern library. At each topology node, the power signal to be measured is input into the distributed dynamic estimator in parallel. The signal is compared with all patterns in the pattern library through a distributed dynamic pattern recognition mechanism to obtain the recognition residual at each node. The recognition residual generated at each topology node is aggregated in real time using a graph convolution method to obtain the source tracing result.

2. The subsynchronous oscillation source tracing method based on high-dimensional deterministic learning modeling as described in claim 1, characterized in that, Prior to distributed dynamics modeling, a dynamic neural network discriminator is constructed: in, It refers to the state of the neural network identifier. It is the gain of the identifier to be designed. This represents an RBF neural network used to learn system dynamics. It is the weight vector of the neural network.

3. The subsynchronous oscillation source tracing method based on high-dimensional deterministic learning modeling as described in claim 1, characterized in that, The distributed dynamics modeling expression is: in, This represents the system dynamics at node n after modeling. This represents the weights at node n after the neural network converges, stored in a time-invariant form. It is the regression vector composed of Gaussian radial basis functions.

4. The subsynchronous oscillation source tracing method based on high-dimensional deterministic learning modeling as described in claim 1, characterized in that, For each topology node, a distributed subsynchronous oscillation pattern library is constructed. The distributed deterministic learning modeling weights are stored in the distributed dynamic pattern library. The pattern library is constructed as follows: in, Represents a node The pattern library at the location, Indicates the first The first in the class Each mode.

5. The subsynchronous oscillation source tracing method based on high-dimensional deterministic learning modeling as described in claim 1, characterized in that, For each distributed node, the RBF neural network is reconstructed using the weights of all patterns in the pattern library, and a dynamic estimator is built based on this. The dynamic estimator is constructed as follows: in, For nodes The power signal to be identified is input to the dynamic estimator. For the state variables of the dynamic estimator, It is the dynamic estimator gain. This represents the dynamic knowledge stored in the pattern library. The regression vector is composed of Gaussian radial basis functions. This represents the sensor sampling time interval.

6. The subsynchronous oscillation source tracing method based on high-dimensional deterministic learning modeling as described in claim 1, characterized in that, Residual calculation for identification Norm, and simultaneously perform fusion processing on the residuals corresponding to the same category of dynamic modes; by fusing the residual identification results of various categories, the fused result is... The results corresponding to each pattern are unified into a single pattern: in, Indicates at time step The smallest one selected A set of indexes for each pattern. This represents the residuals for each category after fusion. It is the preset time window length. For the first in the pattern library The first in the class The time corresponding to each mode Residual identification at any given moment.

7. The subsynchronous oscillation source tracing method based on high-dimensional deterministic learning modeling as described in claim 6, characterized in that, The obtained single-mode residual results are combined into an identification residual matrix at each node, where the row vectors of the matrix represent the nodes. The residual sequences after fusion of various classes are given, and the column vectors represent the residuals of all class recognitions at time t. The residual value at each node; constructing the identification residual topological feature matrix from the identification residual matrices at all nodes; specifically, calculating the output result based on the identification residual. Norm, and simultaneously perform fusion processing on the residuals corresponding to the same category of dynamic modes; by fusing the residual identification results of various categories, the fused result is... The results corresponding to each pattern are normalized into a single pattern; based on the single pattern, an identification residual matrix is ​​constructed at each node, and the row vectors of the identification residual matrix represent the nodes. The residual sequences after various fusions are given, and the column vector represents the residuals of all categories at time 1. The residual value at the point; using graph convolution to create the system's adjacency matrix based on the topological connections between power grid nodes, the topological feature matrix is ​​updated by aggregating information from adjacent nodes using graph convolution, thus updating the topological feature matrix to obtain the topological feature matrix. The result after topological aggregation; take the row vector of a certain node after aggregation as the final identification residual, and the category in the pattern library corresponding to the smallest fusion residual category in the final identification residual is the source tracing result after aggregation.

8. A subsynchronous oscillation source tracing system based on high-dimensional deterministic learning modeling, characterized in that, include: The data acquisition module is used to construct graph structure data based on the preprocessed power grid power signal, and to extract topological information from the graph structure data using graph theory methods. The distributed deterministic learning module is used to perform distributed dynamic modeling of the power signal corresponding to each node in the graph structure data using a deterministic learning mechanism and to build a pattern library. The distributed dynamic recognition module is used to build a distributed dynamic estimator using the neural network weights stored in the pattern library. At each topology node, the power signal to be measured is input into the distributed dynamic estimator in parallel. The signal is compared with all patterns in the pattern library through the distributed dynamic pattern recognition mechanism to obtain the recognition residual at each node. The aggregation and tracing module is used to aggregate the topological information of the identification residuals generated at each topological node in real time using graph convolution to obtain the tracing result.

9. An electronic device, characterized in that, It includes a memory and a processor, as well as computer instructions stored in the memory and running on the processor, which, when executed by the processor, perform the subsynchronous oscillation source tracing method based on high-dimensional deterministic learning modeling as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, Used to store computer instructions, which, when executed by a processor, complete the subsynchronous oscillation source tracing method based on high-dimensional deterministic learning modeling as described in any one of claims 1-7.

Images