Power station resonance multi-mode monitoring and distributed control method of path aggregation impedance

Abstract

The invention relates to the technical field of power grids, and discloses a power station resonance multi-modal monitoring and distributed control method for path aggregation impedance, and the method comprises the following steps: constructing a high-dimensional tensor through obtaining time-varying impedance data of a power station, and carrying out the decomposition to extract a resonance mode; constructing a stable manifold by using modal features, and describing a stable boundary of the system; the resonance risk is quantified by calculating the geometric distance between a future state point and the stable manifold, and finally an optimal preventive control strategy is generated according to a risk assessment result; the system comprises a data acquisition and construction module, a resonance mode decoupling module, a stable manifold construction module, a resonance risk assessment module and a distributed cooperative control module. According to the method, the resonance mode is accurately identified through tensor decomposition, the nonlinear stable boundary is constructed by using manifold learning, the prospective quantitative evaluation of the resonance risk and the intelligent preventive control are realized, and the active safety of power station operation is remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of power grid technology, specifically to a method for multimodal monitoring and distributed control of power plant resonance based on path aggregation impedance. Background Art

[0002] With the rapid development of renewable energy power generation technology, the scale of new energy power plants centered on grid-connected inverters is increasing day by day. The large-scale integration of these power electronic devices, while providing the grid with flexible adjustment capabilities, also leads to interactions between their complex dynamic characteristics and the grid, making the power plant system face increasingly severe challenges in resonance stability, especially in wide frequency bands such as subsynchronous / supersynchronous.

[0003] Currently, impedance analysis is the mainstream technique for assessing such resonance problems. However, its effectiveness faces challenges in practical applications. The operating conditions of a power plant, such as illumination intensity and load levels, are constantly and dynamically changing, causing the system's impedance characteristics to vary over time. Analysis results based on specific steady-state conditions cannot fully reflect the system's true stability at different operating times. Furthermore, in large power plants containing numerous inverters, multiple potential resonance modes are often intertwined and coupled. Traditional frequency domain scanning analysis struggles to effectively separate these modes, thus failing to accurately pinpoint the root cause of the resonance and its spatial impact range within the power plant.

[0004] Furthermore, existing stability criteria largely rely on fixed linear thresholds. This approach simplifies the stability boundary of complex nonlinear systems to a static critical line, making it difficult to adapt the accuracy of the evaluation results to the dynamic evolution of system operating conditions. At the control level, traditional centralized control architectures face severe challenges. This architecture depends on centrally transmitting all station data to a central controller for decision-making, resulting in significant communication delays and bandwidth pressure. Simultaneously, the central controller serves as a single point of failure risk, and its computational load increases exponentially with the scale of the power plant, severely restricting the real-time performance, robustness, and scalability of large-scale new energy power plant control systems.

[0005] Therefore, this invention proposes a method for multimodal monitoring and distributed control of power plant resonance based on path aggregation impedance to address the shortcomings of existing technologies. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a method for multimodal monitoring and distributed control of power plant resonance based on path aggregation impedance. This method solves the problem of how to accurately monitor, proactively assess, and preventively control complex coupled resonance phenomena in new energy power plants under dynamically changing operating conditions.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a method for multimodal monitoring and distributed control of power plant resonance based on path aggregation impedance, the method comprising the following steps:

[0008] S1. Obtain time-varying dynamic impedance data of preset nodes in the power plant, and construct a data tensor containing frequency, spatial and time dimensions based on the time-varying dynamic impedance data.

[0009] S2. Decompose the data tensor to obtain the factor matrix characterizing the resonance mode;

[0010] S3. Define a modal coordinate system using the time dimension characteristics of the factor matrix, and construct a stable manifold in the modal coordinate system to characterize the set of stable operating states of the power station.

[0011] S4. Determine the future state point based on the preset operation of the power station, and assess the resonance risk by calculating the geometric relationship between the future state point and the stable manifold.

[0012] S5. Based on the assessed resonance risk, a preventive control strategy for the power plant is generated through a distributed collaborative control architecture.

[0013] Preferably, in step S1, the step of acquiring time-varying dynamic impedance data of preset nodes within the power station and constructing a data tensor containing frequency, spatial, and time dimensions based on the time-varying dynamic impedance data includes:

[0014] A weighted directed graph model of the power plant electrical network is constructed, wherein the nodes of the weighted directed graph model represent electrical equipment, the edges represent physical connections, and the weight of the edges is the line complex impedance; time-varying dynamic impedance data of multiple nodes are collected; and based on the collected time-varying dynamic impedance data, a third-order impedance data tensor is constructed as the data tensor.

[0015] Preferably, based on the weighted directed graph model, a path transfer matrix is ​​constructed to describe any specified path within the power station, and the path aggregate impedance of the specified path is calculated based on the path transfer matrix.

[0016] Preferably, in step S2, the step of decomposing the data tensor to obtain the factor matrix characterizing the resonance mode includes:

[0017] The decomposition of the data tensor is performed by solving an optimization objective function, which includes topological information calculated based on the weighted directed graph model, in order to separate the physically meaningful resonant modes from the data tensor.

[0018] Preferably, the decomposition is performed by constructing a specific mathematical expression for the optimization objective function; the mathematical expression includes a data fidelity term and a physical topology regularization term; the data fidelity term measures the error between the tensor reconstructed from the factor matrix obtained by the decomposition and the original data tensor constructed; the physical topology regularization term uses the topological information of the weighted directed graph model as a constraint to ensure that the decomposed resonant modes conform to physical connectivity characteristics in spatial distribution; the mathematical expression for the optimization objective function is:

[0019]

[0020] In the formula, Let A be the data tensor; A, B, and C be the frequency, spatial, and temporal factor matrices to be solved, respectively; a r ,b r ,c r Let be the column vectors corresponding to the r-th resonant mode in the factor matrix; R be the total number of resonant modes; ° be the cross product; ||·|| F γ is the Frobenius norm; γ is the regularization coefficient; tr(·) is the trace of the matrix; B H L is the conjugate transpose of the spatial factor matrix B; L is the Laplace matrix calculated by the weighted directed graphical model.

[0021] Preferably, in step S3, the step of defining a modal coordinate system using the time dimension characteristics of the factor matrix, and constructing a stable manifold representing the set of stable operating states of the power plant in the modal coordinate system includes:

[0022] The column vectors of the time factor matrix in the factor matrix obtained by decomposition are defined as the basis of the modal coordinate system; multiple modal state vectors corresponding to the historical stable operating conditions of the power station are collected, which are composed of the row vectors of the time factor matrix; and a manifold learning algorithm is applied to the collected multiple modal state vectors to construct the stable manifold.

[0023] Preferably, in step S4, the step of determining the future state point based on the preset operation of the power plant and assessing the resonance risk by calculating the geometric relationship between the future state point and the stable manifold includes:

[0024] A prediction model is established based on the preset operation of the power station to determine the future state point in the modal coordinate system; the geodesic distance between the future state point and the boundary of the stable manifold is calculated; and the reciprocal of the geodesic distance is used as a risk index to quantify the resonance risk.

[0025] Preferably, step S5, which involves generating a preventative control strategy for the power plant based on the assessed resonance risk, includes:

[0026] The controllable equipment in the power station is defined as multiple control units, and each control unit is configured with a deep reinforcement learning sub-agent to form a multi-agent control system.

[0027] The local measurement information corresponding to each sub-agent, as well as the global resonant mode information and risk assessment results obtained from the analysis, are used as the composite state input;

[0028] Through information interaction and collaborative decision-making within the multi-agent system, the optimal control adjustment is output, which together constitutes the preventive control strategy, thereby realizing distributed control of the power plant.

[0029] Preferably, the offline training steps of the multi-agent system include:

[0030] Construct a simulation environment consistent with the dynamic characteristics of the power plant;

[0031] In the simulation environment, a training paradigm of centralized training and decentralized execution is adopted. That is, during the training process, a centralized evaluation module with access to global information is used to evaluate the behavior of each sub-agent and generate guidance signals to optimize the strategy of each sub-agent.

[0032] Based on the acquired reward signal, and using a reinforcement learning algorithm to iteratively update the internal parameters of each sub-agent until they converge to the optimal cooperative strategy, after training, each sub-agent can be deployed and executed in a distributed manner.

[0033] This invention also provides a multimodal monitoring and distributed control system for power plant resonance based on path aggregation impedance, the system comprising the following modules:

[0034] The data acquisition and construction module is used to acquire time-varying dynamic impedance data of preset nodes in the power plant, and construct a data tensor containing frequency, spatial and time dimensions based on the time-varying dynamic impedance data.

[0035] The resonant mode decoupling module is used to decompose the data tensor to obtain the factor matrix characterizing the resonant mode;

[0036] A stable manifold construction module is used to define a modal coordinate system using the time dimension features of the factor matrix, and to construct a stable manifold in the modal coordinate system that represents the set of stable operating states of the power plant.

[0037] The resonance risk assessment module is used to determine the future state point based on the preset operation of the power plant, and to assess the resonance risk by calculating the geometric relationship between the future state point and the stable manifold.

[0038] The distributed collaborative control module is used to deploy a multi-agent control system. Based on the assessed resonance risk and by integrating global state information and local measurement data, the system generates and executes a distributed preventive control strategy through collaborative decision-making among the agents.

[0039] This invention provides a method for multimodal monitoring and distributed control of power plant resonance based on path aggregation impedance. It offers the following advantages:

[0040] 1. This invention integrates time-varying dynamic impedance data from multiple nodes within a power plant into a high-dimensional tensor and applies constraints to the tensor decomposition process using a graph Laplace matrix containing physical topology information of the power plant. This achieves precise decoupling and physical tracing of multiple resonant modes. Compared to existing technologies that isolate the impedance characteristics of a single measurement point or frequency point, this invention overcomes the limitation of failing to fully capture the spatiotemporal coupling correlation of resonant modes, thus improving the accuracy and physical interpretability of resonant mode identification.

[0041] 2. This invention utilizes the time evolution characteristics of resonant modes extracted from actual operating data to construct a nonlinear geometric model characterizing the system's stability domain through a manifold learning algorithm. This achieves a global and refined characterization of the power plant's stability boundary. Unlike existing methods that commonly use fixed linear thresholds for stability assessment, this invention overcomes the limitation of these methods in adapting to the complex and variable operating conditions of power plants, thereby effectively reducing the misjudgment and missed detection rates of resonance risk.

[0042] 3. This invention proposes a forward-looking risk assessment mechanism. It determines the future state point under preset operations through a predictive model and calculates the geodesic distance between this point and the stable manifold boundary, thereby quantifying the risk. This achieves a shift from passive monitoring to proactive prediction. It overcomes the limitation of most existing technologies that only respond after resonance phenomena occur, gaining a valuable time window for implementing preventative control and improving the proactive safety of system operation.

[0043] 4. This invention employs a multi-agent distributed cooperative control architecture to generate preventative control strategies. The global control task is decomposed into multiple cooperating sub-agents. Through an advanced paradigm of "centralized training and distributed execution," it learns a nonlinear mapping from a state that integrates global and local information to the optimal control action. Compared to existing control methods that rely on fixed models or single centralized agents, this invention overcomes the problems of single-point failure risk, communication bandwidth bottlenecks, and high computational complexity in large-scale systems, achieving a more intelligent, efficient, and reliable preventative intervention against resonance risks. Attached Figure Description

[0044] Figure 1 This is a flowchart of the method of the present invention;

[0045] Figure 2 This is a system architecture diagram of the present invention. DETAILED DESCRIPTION

[0046] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0047] Please see Figure 1 This invention provides a method for multimodal monitoring and distributed control of power plant resonance based on path aggregation impedance. The method includes the following steps:

[0048] S1. Obtain time-varying dynamic impedance data of preset nodes in the power plant, and construct a data tensor containing frequency, spatial and time dimensions based on the time-varying dynamic impedance data.

[0049] In this embodiment, the specific implementation of obtaining time-varying dynamic impedance data of preset nodes within the power station in step S1, and constructing a data tensor containing frequency, spatial, and time dimensions based on the time-varying dynamic impedance data, is as follows:

[0050] In one possible implementation, to provide a systematic mathematical description of the complex electrical network of a power station, a weighted directed graph model of the power station's electrical network is first constructed. This graph model is an abstract representation of the power station's physical topology, providing a structured framework for subsequent impedance data acquisition and analysis.

[0051] Specifically, the construction of the weighted directed graph model includes defining nodes and edges. The nodes of the model represent key electrical equipment within the power plant. In some embodiments, these devices are the objects of interest in resonance analysis, such as grid-connected inverters, transformers, AC or DC filters, and high-power loads. The edges of the model represent the physical electrical connections between these devices, such as connecting cables or busbars. The directionality of the edges can represent the typical direction of power flow or signal propagation.

[0052] To quantify the characteristics of electrical connections, a weight is assigned to each edge of the graph model. This weight is defined as the complex impedance value of the corresponding physical connection line. The complex impedance value comprehensively describes the resistance and reactance characteristics of the line and is an important physical parameter for subsequent resonance analysis.

[0053] After constructing the weighted directed graph model, time-varying dynamic impedance data is collected at multiple preset nodes within the model. These preset nodes are typically selected at critical locations in the system, such as the AC output of the inverter and the point of common coupling (PCC). The purpose of collecting time-varying dynamic impedance data is to capture the dynamic process of system impedance characteristics changing with operating conditions. Generally, the operating conditions of the power plant, such as the output level of new energy sources, load start-up and shutdown, and the switching on and off of reactive power compensation equipment, will cause changes in system impedance.

[0054] In one possible implementation, the acquisition of the time-varying dynamic impedance data can be achieved through online perturbation measurement. For example, at a predetermined node, a known perturbation signal with a wide spectral component is injected into the system via a signal generating device, while the voltage and current response signals of that node are acquired at high speed. By performing Fourier transform on the acquired voltage and current signals, their components in the frequency domain are obtained, and then the complex impedance values ​​at a series of frequency points are calculated.

[0055] Based on the acquired time-varying dynamic impedance data, a third-order impedance data tensor is constructed. The construction of this data tensor aims to integrate scattered data collected from different dimensions (frequency, space, time) into a unified mathematical object, facilitating subsequent multidimensional signal processing and mode decomposition. The third-order data tensor... The three dimensions are specifically defined as: frequency dimension, spatial dimension, and time dimension.

[0056] The frequency dimension corresponds to the set of frequency points covered by the impedance scan measurement. The spatial dimension corresponds to the preset set of nodes for data acquisition. The temporal dimension corresponds to the sequence of different time snapshots for data acquisition. Therefore, any element in the tensor... This represents the complex impedance value measured at the j-th spatial node and the i-th frequency point in the k-th time snapshot.

[0057] Furthermore, in this embodiment, the weighted directed graph model can also be used for more refined path aggregation impedance analysis. Specifically, for any specified signal propagation path in the power plant network with clearly defined start and end points, a path transfer matrix can be constructed. The path transfer matrix is ​​used to describe the transformation relationship of voltage or current disturbance signals as they propagate along the specified path, and the elements of the matrix are jointly determined by the impedance parameters of each node and branch along the path.

[0058] After constructing the path transfer matrix, the path aggregate impedance of the specified path can be calculated based on this matrix. The path aggregate impedance is a representation of the overall equivalent impedance of the path, comprehensively reflecting the impedance characteristics of all series and parallel components on the path. The calculation of this path aggregate impedance provides a powerful tool for analyzing the interaction between specific devices, such as between one inverter and another, and for identifying critical resonant loops.

[0059] S2. Decompose the data tensor to obtain the factor matrix characterizing the resonance mode;

[0060] In this embodiment, the specific implementation of decomposing the data tensor in step S2 to obtain the factor matrix representing the resonance mode is as follows:

[0061] After obtaining the third-order impedance data tensor in step S1, this step aims to separate and extract the individual resonant modes from the high-dimensional data using mathematical decomposition methods. Generally, the original data tensor is a mixture of multiple potential resonant modes and other dynamic processes linearly superimposed, making direct analysis difficult. Therefore, this embodiment employs a tensor decomposition-based technique to decouple these modes.

[0062] In one possible implementation, the decomposition of the data tensor is performed by solving an optimization objective function. The objective function is designed to ensure that the decomposed resonant modes not only mathematically accurately reconstruct the original data, but more importantly, have a clear and logical physical meaning. To this end, the objective function includes topological information calculated based on the weighted directed graph model constructed in step S1, to guide the decomposition process.

[0063] Specifically, to perform the decomposition, an optimization objective function is constructed as follows, the mathematical expression of which can be named the Physical Constraint Tensor Decomposition Objective Function:

[0064]

[0065] In the formula: The third-order impedance data tensor constructed in step S1; A, B, and C are the frequency factor matrix, spatial factor matrix, and time factor matrix to be solved, respectively, and the goal of the decomposition is to solve these three matrices; a r ,b r ,c r These are the column vectors corresponding to the r-th resonant mode in the frequency factor matrix A, spatial factor matrix B, and time factor matrix C, respectively; R is the preset total number of resonant modes to be decomposed, which is also the rank of the tensor decomposition. The symbol represents the cross product operation. Generate a rank-one tensor with the same dimension as the original tensor, representing a single resonant mode; ||·|| F γ is the Frobenius norm of a matrix or tensor, used to measure the magnitude of the error between two tensors; γ is a positive regularization coefficient used to adjust the strength of physical topological constraints in the optimization objective; tr(·) is the trace operation of the matrix; B H L is the conjugate transpose of the spatial factor matrix B; L is the graph Laplacian matrix calculated by the weighted directed graph model.

[0066] The optimization objective function mainly consists of two parts: a data fidelity term and a physical topology regularization term.

[0067] The data fidelity item, namely This term measures the error between the tensor reconstructed from the factor matrix obtained from the decomposition and the original data tensor. By minimizing this term, we can ensure that the decomposition result retains the information of the original data to the greatest extent possible, thus ensuring the accuracy of the decomposition.

[0068] The physical topological regularization term, namely γ·tr(B H The graph Laplace matrix L is key to introducing prior physical knowledge. The graph Laplace matrix L is calculated from the adjacency matrix and degree matrix of the weighted directed graph model constructed in step S1. This matrix encodes the physical connections and line impedance information between nodes in the power plant network. Applying this regularization term to the spatial factor matrix B ensures that the spatial distribution of the decomposed resonant modes conforms to the physical characteristics of the network. Specifically, minimizing this term ensures that physically directly connected nodes exhibit a higher degree of participation (i.e., spatial factor vector b) in the same resonant mode. r The values ​​of corresponding elements in the model tend to be close. This constraint effectively suppresses the generation of pseudo-patterns that may exist mathematically but are meaningless physically, ensuring that each decomposed pattern has a clear physical correspondence.

[0069] By solving the above objective function, for example using iterative algorithms such as alternating least squares, the optimal factor matrices A, B, and C can be obtained. The column vectors of these three matrices describe the characteristics of each separated resonant mode in terms of frequency, space, and time, respectively.

[0070] The column vector a of the frequency factor matrix A r This describes the spectral characteristics of the r-th resonant mode, from which the mode's resonant frequency, peak impedance, and bandwidth can be identified.

[0071] The column vector b of the spatial factor matrix B r The spatial distribution characteristics of the r-th resonant mode are described, revealing the degree of participation and influence of this mode at different physical nodes of the power plant.

[0072] The column vector c of the time factor matrix C r The dynamic trajectory of the intensity and phase of the r-th resonant mode as a function of time is described, reflecting how the mode is affected by changes in the operating conditions of the power plant.

[0073] S3. Define a modal coordinate system using the time dimension characteristics of the factor matrix, and construct a stable manifold in the modal coordinate system to characterize the set of stable operating states of the power station.

[0074] In this embodiment, the specific implementation of step S3, which uses the time dimension features of the factor matrix to define a modal coordinate system and constructs a stable manifold representing the set of stable operating states of the power plant in the modal coordinate system, is as follows:

[0075] After decomposing the data tensor in step S2, the core task of this step is to transform the decomposed time factor matrix, which describes the dynamic characteristics of each resonance mode, into a geometric model that can intuitively characterize the system's stability boundary. This provides a benchmark for subsequent quantitative assessment of resonance risk.

[0076] In one possible implementation, a modal coordinate system is first defined using the time factor matrix C obtained in step S2. Each column vector c of the time factor matrix C... r (where r ranges from 1 to R) describes the evolution of the amplitude of an independent resonant mode over time. These column vectors form a basis for describing the dynamic behavior of the system. Therefore, these R column vectors {c1, c2, ..., c...} R Using a set of orthogonal or approximately orthogonal bases, an R-dimensional coordinate space is constructed, which is the modal coordinate system. In this coordinate system, each coordinate axis uniquely corresponds to a specific resonant mode.

[0077] Specifically, the overall operating state of the power plant system at any time snapshot t can be represented as a state point in this modal coordinate system. The coordinates of this state point are determined by the t-th row vector of the time factor matrix C, which can be represented as p. t =[C t1 C t2 ,...,C tR This vector is called the modal state vector, and each of its components C tR This represents the activation level or amplitude of the r-th resonant mode at time snapshot t.

[0078] To construct a model that can characterize the set of stable operating states of a power plant, it is necessary to collect a large number of modal state vectors corresponding to the historical stable operating conditions of the power plant. Generally, by monitoring the power plant's operating records, several time periods in which the system is in a clearly stable state can be identified. All modal state vectors {p} corresponding to these time periods will be... t The data is extracted to form a high-dimensional dataset. In the modal coordinate system, this dataset is represented as a point cloud distribution.

[0079] In one possible implementation, the point cloud is not randomly distributed in the modal coordinate system, but rather tends to cluster on or near a lower-dimensional nonlinear geometric structure. This geometric structure is the stable manifold. The stable manifold geometrically represents the set of all stable operating states.

[0080] To learn and construct this stable manifold from point cloud data, this embodiment employs a manifold learning algorithm. Specifically, algorithms such as isomap or locally linear embedding (LLE) can be used. These algorithms can effectively identify the low-dimensional essential geometric structure inherent within high-dimensional point clouds. By applying the manifold learning algorithm to the acquired multiple modal state vectors, a mathematical representation of the stable manifold can be obtained. This mathematical representation fully defines the geometry, range, and boundaries of the stable operating region, providing a foundation for calculating the distance between state points and the stable boundary in subsequent steps.

[0081] S4. Determine the future state point based on the preset operation of the power station, and assess the resonance risk by calculating the geometric relationship between the future state point and the stable manifold.

[0082] In this embodiment, the specific implementation method for determining the future state point based on the preset operation of the power station in step S4, and assessing the resonance risk by calculating the geometric relationship between the future state point and the stable manifold, is as follows:

[0083] After constructing the stable manifold characterizing the stable operating region in step S3, this step aims to conduct a forward-looking and quantitative risk assessment of any pre-planned operation of the power plant that may affect stability, thereby achieving the prediction of resonance risk.

[0084] In one possible implementation, it is first necessary to determine the future state point of the system in the modal coordinate system after a preset operation of the power plant is performed. The preset operation may include adjusting the active or reactive power output of one or more inverters, switching on / off a set of reactive power compensation devices, or changing controller parameters. To achieve this prediction, a prediction model needs to be established. This model describes the mapping relationship between the preset operation and the change in state point in the modal coordinate system. Through this prediction model, a specific physical domain operation can be transformed into an expected change in the modal state vector, thereby determining the future state point.

[0085] After determining the future state point, the next step is to calculate the geometric distance between the future state point and the boundary of the stable manifold constructed in step S3 in the modal coordinate system. Since the stable manifold is a nonlinear curved structure existing in a high-dimensional space, a simple Euclidean straight-line distance cannot accurately measure the true proximity of the state point to the stable boundary. Therefore, in this embodiment, the geodesic distance between the future state point and the boundary of the stable manifold is calculated.

[0086] Specifically, the geodesic distance is the shortest path length connecting two points on the surface of the manifold, accurately reflecting the distance the system state needs to traverse to move to the boundary within the stable region. This calculation requires first finding a point on the boundary of the stable manifold whose geodesic distance to the future state point is minimized.

[0087] After calculating the shortest geodesic distance, in order to obtain an intuitive and standardized risk metric, this embodiment uses the reciprocal of the geodesic distance as a risk index to quantify the resonance risk. The risk index can be calculated using the following resonance risk index quantification formula:

[0088]

[0089] In the formula, I risk p is the resonance risk index to be calculated. future The future state point determined by the prediction model; The boundary of the stable manifold constructed in step S3; Let p be the future state point future With the stable manifold boundary The shortest geodesic distance between them.

[0090] As can be seen from the formula, the smaller the geodesic distance between the future state point and the stable boundary, the larger the calculated risk index, indicating that the preset operation will cause the system to approach the unstable boundary, resulting in a higher risk of resonance. Conversely, if the geodesic distance is large, the risk index is small, indicating that the system remains in a deeply stable region after the operation, with a lower risk of resonance. This risk index provides a quantitative basis for the subsequent generation of preventative control strategies.

[0091] S5. Based on the assessed resonance risk, generate a preventive control strategy for the power plant;

[0092] In this embodiment, step S5, generating a preventative control strategy for the power plant based on the evaluated resonance risk, is implemented through a pre-trained multi-agent system (MAS) composed of multiple deep reinforcement learning sub-agents. This method deconstructs the original centralized control problem into a distributed collaborative decision-making problem, thereby obtaining an intelligent control strategy that can adapt to the complex nonlinear dynamics of the power plant while taking into account response speed, robustness, and scalability.

[0093] The specific implementation of this distributed control architecture is as follows:

[0094] Control Architecture: Controllable power electronic equipment (such as inverters) within the power plant is divided into multiple independent control units or clusters, and each unit is configured with a deep reinforcement learning sub-agent. These sub-agents together constitute a multi-agent network, realizing the transformation from a "central brain" to a "distributed brain cluster," laying the foundation for decentralized execution and distributed control.

[0095] Decision-making process: In the actual operation of the power plant, the various sub-agents make decisions in a collaborative manner. Specifically:

[0096] State Input: The state input of each sub-agent is composite, fusing global and local information. This composite state vector includes:

[0097] Local measurement information: such as real-time data of port voltage, current, power, etc. of the control unit controlled by the sub-smart agent.

[0098] Global context information: The global resonant mode information and risk assessment results obtained and issued by the global analysis calculations in steps S2-S4 above are used as its composite state input, such as the state projection of the unit in the global resonant mode coordinate system and the safe distance (or risk index) to the global stable manifold.

[0099] Collaborative decision-making: The sub-agents do not make decisions independently, but rather interact through a communication network, sharing their states or intentions. They collaboratively make decisions to maximize a common global long-term cumulative reward, outputting their respective optimal control adjustments. The set of actions output by all sub-agents collectively constitutes the distributed preventative control strategy for the entire power plant.

[0100] To achieve the aforementioned distributed cooperative control, this preventative control problem is modeled as a multi-agent Markov decision process (Multi-Agent-MDP). Its core elements lie in the state space, the joint action space, and the shared reward function.

[0101] The shared reward function is jointly optimized by all sub-agents, integrating global resonance risk, total control cost, and overall power plant operating performance indicators. In a specific embodiment, this reward function can be defined as the following reward function model:

[0102]

[0103] In the formula, R t s represents the global instantaneous reward at time t; t The global state of the system can be composed of the local states of all sub-agents and the global modal information; R is the control adjustment amount applied by the i-th sub-agent at time t; N is the total number of sub-agents; risk The first term is the risk reward, directly related to the global resonance risk; the second term is the total control cost of all sub-agent actions; P(s) t ) represents the boundary penalty term; λ c These are the weighting coefficients.

[0104] Offline training: To enable offline training of the aforementioned multi-agent system, the advanced training paradigm of "Centralized-Training-with-Decentralized-Execution" (CTDE) is preferred.

[0105] Construct a high-fidelity simulation environment that can accurately reproduce the dynamic characteristics of a power plant.

[0106] During training, a centralized evaluation module is employed. This module has access to global information during training, including observations and actions of all sub-agents. It is responsible for evaluating the quality of joint actions and providing accurate guidance signals for policy updates of each sub-agent.

[0107] Each sub-agent (acting as an actuator) uses the collected experience samples and a pre-defined multi-agent reinforcement learning algorithm to iteratively update its internal network parameters based on local observations and guidance from the centralized evaluation module.

[0108] In some embodiments, advanced algorithms such as Multi-Agent Deep Deterministic Policy Gradient (MADDPG) may be preferred, as these algorithms are effective methods for handling multi-agent continuous control problems.

[0109] The above iterative process continues until the policies of all sub-agents converge to the optimal cooperative policy. After training is complete, the centralized evaluation module can be discarded, and the trained sub-agents are solidified and deployed in their respective control units. In actual operation, they only need local information and necessary neighbor communication to perform efficient, decentralized decision-making and execution, thereby achieving intelligent, fast, robust, and distributed preventive control of resonance risks.

[0110] Please see Figure 2 The present invention also provides a multimodal monitoring and distributed control system for power plant resonance based on path aggregation impedance, the system comprising the following modules:

[0111] The data acquisition and construction module is used to acquire time-varying dynamic impedance data of preset nodes in the power plant, and construct a data tensor containing frequency, spatial and time dimensions based on the time-varying dynamic impedance data.

[0112] This module is responsible for interacting with the physical equipment of the power plant. It acquires dynamic impedance data that changes in real time with the power plant's operating conditions by performing online measurements at multiple pre-set key nodes (such as inverter grid connection points and common coupling points). Subsequently, the module systematically organizes the collected discrete data, distributed across different time points, physical locations, and frequencies, into a structured third-order data tensor. The three dimensions of this tensor correspond to frequency, space, and time, respectively, providing a unified data foundation for subsequent analysis.

[0113] The resonant mode decoupling module is used to decompose the data tensor to obtain the factor matrix characterizing the resonant mode;

[0114] The system receives the data tensor output by the data acquisition and construction module and performs a tensor decomposition algorithm that considers the physical topology constraints of the power station on it. This decomposition process decouples the linearly superimposed and complex dynamic information in the original data into a series of mutually independent fundamental resonant modes with clear physical meaning. The decomposition results in several factor matrices, which finely characterize the complete features of each independent resonant mode from three orthogonal dimensions: frequency characteristics, spatial distribution across the entire station, and temporal evolution.

[0115] A stable manifold construction module is used to define a modal coordinate system using the time dimension features of the factor matrix, and to construct a stable manifold in the modal coordinate system that represents the set of stable operating states of the power plant.

[0116] The focus is on utilizing the temporal dimension feature information output by the resonant mode decoupling module. First, a novel modal coordinate system is defined based on the temporal evolution characteristics of each resonant mode. Then, all moments in the power plant's historical operation records that are in a stable state are mapped to a series of state points in this coordinate system, forming a high-dimensional point cloud. Finally, the module applies a manifold learning algorithm to learn and construct a low-dimensional, nonlinear stable manifold from this point cloud data. This manifold geometrically accurately depicts the set of all safe and stable operating states of the power plant and its boundaries.

[0117] The resonance risk assessment module is used to determine the future state point based on the preset operation of the power plant, and to assess the resonance risk by calculating the geometric relationship between the future state point and the stable manifold.

[0118] This module is used for proactive risk assessment. Upon receiving a pre-set power plant operation command, it first uses a predictive model to calculate the future state point the system will reach in the modal coordinate system after the operation. Next, it calculates the shortest geometric distance (i.e., geodesic distance) between this future state point and the boundary of the stable manifold. This distance accurately reflects the proximity of the system state to the unstable region. The module converts the reciprocal of this distance into an intuitive resonance risk index, thereby achieving a quantitative assessment of operational risk.

[0119] The distributed cooperative control module is used to deploy a multi-agent control system. Based on the assessed resonance risk and by integrating global state information and local measurement data, the system generates and executes a distributed preventive control strategy through collaborative decision-making among the agents.

[0120] The core of this module is a distributed control network composed of multiple deep reinforcement learning sub-agents, replacing the traditional single centralized controller. These sub-agents have learned and internalized a set of optimal cooperative control laws capable of coping with complex and ever-changing working conditions during the offline training phase, which adopts a "centralized training, distributed execution" paradigm.

[0121] During the power plant's online operation phase, this module receives global analysis results (such as resonant mode information and risk index) provided by the aforementioned modules and distributes them as macroscopic context information to each sub-agent in the network. Each sub-agent combines the received global information with its own control unit's local real-time measurement data (such as port voltage and current) to form the composite state input required for its decision-making. Based on this composite state and through information interaction between sub-agents, the system can quickly formulate a set of optimal control adjustments in a distributed manner. This set of control adjustments collectively constitutes the final distributed preventive control strategy, which is then distributed to the corresponding power electronic devices, achieving rapid, robust, and highly scalable intelligent intervention against resonant risks.

[0122] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A method for multimodal monitoring and distributed control of power plant resonance based on path aggregation impedance, characterized in that, The method includes the following steps: S1. Obtain time-varying dynamic impedance data of preset nodes in the power plant, and construct a data tensor containing frequency, spatial and time dimensions based on the time-varying dynamic impedance data. S2. Decompose the data tensor to obtain the factor matrix characterizing the resonance mode; S3. Define a modal coordinate system using the time dimension characteristics of the factor matrix, and construct a stable manifold in the modal coordinate system to characterize the set of stable operating states of the power station. S4. Determine the future state point based on the preset operation of the power station, and assess the resonance risk by calculating the geometric relationship between the future state point and the stable manifold. S5. Based on the assessed resonance risk, a preventive control strategy for the power plant is generated through a distributed collaborative control architecture.

2. The method for multi-mode monitoring and distributed control of power plant resonance based on path aggregation impedance according to claim 1, characterized in that, Step S1, which involves acquiring time-varying dynamic impedance data of preset nodes within the power plant and constructing a data tensor containing frequency, spatial, and time dimensions based on the time-varying dynamic impedance data, includes the following steps: A weighted directed graph model of the power plant electrical network is constructed, wherein the nodes of the weighted directed graph model represent electrical equipment, the edges represent physical connections, and the weight of the edges is the line complex impedance; time-varying dynamic impedance data of multiple nodes are collected; and based on the collected time-varying dynamic impedance data, a third-order impedance data tensor is constructed as the data tensor.

3. The method for multimodal monitoring and distributed control of power plant resonance based on path aggregation impedance according to claim 2, characterized in that, Based on the weighted directed graph model, a path transfer matrix is ​​constructed to describe any specified path within the power station, and the path aggregate impedance of the specified path is calculated based on the path transfer matrix.

4. The method for multimodal monitoring and distributed control of power plant resonance based on path aggregation impedance according to claim 2, characterized in that, Step S2, the step of decomposing the data tensor to obtain the factor matrix characterizing the resonance mode, includes: The decomposition of the data tensor is performed by solving an optimization objective function, which includes topological information calculated based on the weighted directed graph model, in order to separate the physically meaningful resonant modes from the data tensor.

5. The method for multimodal monitoring and distributed control of power plant resonance based on path aggregation impedance according to claim 4, characterized in that, To perform the decomposition, a specific mathematical expression for the optimization objective function is constructed. This mathematical expression includes a data fidelity term and a physical topology regularization term. The data fidelity term measures the error between the tensor reconstructed from the factor matrix obtained from the decomposition and the original data tensor constructed. The physical topology regularization term uses the topological information of the weighted directed graph model as a constraint to ensure that the decomposed resonant modes conform to physical connectivity characteristics in their spatial distribution. The mathematical expression for the optimization objective function is: In the formula, Let A be the data tensor; A, B, and C be the frequency, spatial, and temporal factor matrices to be solved, respectively; a r ,b r ,c r These are the column vectors corresponding to the r-th resonant mode in the factor matrix; R is the total number of resonant modes; For vector outer product; ||·|| F γ is the Frobenius norm; γ is the regularization coefficient; tr(·) is the trace of the matrix; B H L is the conjugate transpose of the spatial factor matrix B; L is the Laplace matrix calculated by the weighted directed graphical model.

6. The method for multi-mode monitoring and distributed control of power plant resonance based on path aggregation impedance according to claim 1, characterized in that, Step S3, which involves defining a modal coordinate system using the time dimension characteristics of the factor matrix and constructing a stable manifold representing the set of stable operating states of the power plant within the modal coordinate system, includes: The column vectors of the time factor matrix in the factor matrix obtained by decomposition are defined as the basis of the modal coordinate system; multiple modal state vectors corresponding to the historical stable operating conditions of the power station are collected, which are composed of the row vectors of the time factor matrix; and a manifold learning algorithm is applied to the collected multiple modal state vectors to construct the stable manifold.

7. The method for multimodal monitoring and distributed control of power plant resonance based on path aggregation impedance according to claim 1, characterized in that, In step S4, the step of determining the future state point based on the preset operation of the power plant and assessing the resonance risk by calculating the geometric relationship between the future state point and the stable manifold includes: A prediction model is established based on the preset operation of the power station to determine the future state point in the modal coordinate system; the geodesic distance between the future state point and the boundary of the stable manifold is calculated; and the reciprocal of the geodesic distance is used as a risk index to quantify the resonance risk.

8. The method for multi-mode monitoring and distributed control of power plant resonance based on path aggregation impedance according to claim 1, characterized in that, Step S5, which involves generating a preventative control strategy for the power plant based on the assessed resonance risk, includes: The controllable equipment in the power station is defined as multiple control units, and each control unit is configured with a deep reinforcement learning sub-agent to form a multi-agent control system. The local measurement information corresponding to each sub-agent, as well as the global resonant mode information and risk assessment results obtained from the analysis, are used as the composite state input; Through information interaction and collaborative decision-making within the multi-agent system, the optimal control adjustment is output, which together constitutes the preventive control strategy, thereby realizing distributed control of the power plant.

9. The method for multimodal monitoring and distributed control of power plant resonance based on path aggregation impedance according to claim 8, characterized in that, The offline training steps for the multi-agent system include: Construct a simulation environment consistent with the dynamic characteristics of the power station; In the simulation environment, a training paradigm of centralized training and decentralized execution is adopted. That is, during the training process, a centralized evaluation module with access to global information is used to evaluate the behavior of each sub-agent and generate guidance signals to optimize the strategy of each sub-agent. Based on the acquired reward signal, and using a reinforcement learning algorithm to iteratively update the internal parameters of each sub-agent until they converge to the optimal cooperative strategy, after training, each sub-agent can be deployed and executed in a distributed manner.

10. A multimodal monitoring and distributed control system for power plant resonance based on path aggregation impedance, applied to the method described in any one of claims 1-9, characterized in that, The system includes the following modules: The data acquisition and construction module is used to acquire time-varying dynamic impedance data of preset nodes in the power plant, and construct a data tensor containing frequency, spatial and time dimensions based on the time-varying dynamic impedance data. The resonant mode decoupling module is used to decompose the data tensor to obtain the factor matrix characterizing the resonant mode; A stable manifold construction module is used to define a modal coordinate system using the time dimension features of the factor matrix, and to construct a stable manifold in the modal coordinate system that represents the set of stable operating states of the power plant. The resonance risk assessment module is used to determine the future state point based on the preset operation of the power plant, and to assess the resonance risk by calculating the geometric relationship between the future state point and the stable manifold. The distributed collaborative control module is used to deploy a multi-agent control system. Based on the assessed resonance risk and by integrating global state information and local measurement data, the system generates and executes a distributed preventive control strategy through collaborative decision-making among the agents.

Images