Broadband-domain unified analysis method and system for multiple networking devices

Abstract

The invention relates to the technical field of power system stability analysis, in particular to a multi-networking equipment broadband domain unified analysis method and system, and the method comprises the steps: collecting the operation state data of multi-networking equipment, and generating a broadband domain track data set; constructing a deep Koopman neural network model, inputting a wide-frequency-domain trajectory data set to obtain dynamic state features, and mapping the dynamic state features to a high-dimensional lifting state space; modeling a broadband dynamic evolution relationship of the plurality of networking devices; generating a dynamic modulation coefficient based on the target working condition parameters, and performing linear transformation on the intermediate features of the deep Koopman neural network model; restoring a broadband dynamic track in a physical space; and calculating a channel participation index for the gradient of each characteristic mode, positioning a tracing device or a control channel of broadband oscillation, and generating a broadband stability evaluation report. According to the invention, the problems of modal aliasing and traceability failure caused by the wide-frequency-domain oscillation multi-band strong coupling characteristic of the new energy power system are effectively solved.

Description

Technical Field

[0001] This invention relates to the field of power system stability analysis technology, and in particular to a unified wideband analysis method and system for multi-network equipment. Background Technology

[0002] In the field of power system stability analysis, especially in power grid environments with a high proportion of renewable energy integration, the monitoring and tracing of broadband oscillations (0-2000Hz) has become a core challenge in ensuring the safe operation of the system. Its physical nature stems from the complex interaction between power electronic equipment and the traditional mechanical characteristics of the power grid. The rapid modulation response of wind power converters and the coupling of distributed parameters in long cables can excite subsynchronous oscillations, while the switching action of photovoltaic inverters can easily trigger high-frequency resonances. Furthermore, the attenuation of the grid foundation's rotating inertia further exacerbates the risk of low-frequency mechanical oscillations. Multi-band oscillation modes interpenetrate and intertwine during nonlinear energy transfer, forming a strong cross-band coupling effect. This effect not only causes abnormal transfer of dynamic energy between adjacent frequency bands but also leads to the mutual masking of key characteristic signals under complex operating conditions.

[0003] Existing technologies employ only global linear modeling methods, such as stability analysis based on dense matrices. Their fundamental flaw lies in solving the full-frequency system response using a single global matrix, which cannot avoid the dominance of high-energy low-frequency oscillations on high-frequency characteristic parameters. Furthermore, because the model parameter update mechanism always preferentially converges to the low-frequency modes that dominate system behavior, singular values ​​corresponding to the rapid transient processes of switching devices are continuously reduced during state propagation. Ultimately, oscillation components exceeding a specific frequency threshold are incorrectly treated as noise and filtered out during the calculation process. Dynamic gain adjustment methods, such as history-dependent control architectures, are also used to address nonlinear coupling problems through adaptive adjustment. However, single-channel parameter systems are forced to simultaneously handle slow mechanical dynamics and fast electromagnetic transients. When oscillation modes of different time scales are superimposed in a common channel, response frequency confusion occurs during gradient calculation. Especially in scenarios where subsynchronous and high-frequency harmonics coexist, the controller misjudges high-frequency components as noise derived from low-frequency oscillations and suppresses them, resulting in the failure to accurately locate the actual harmonic source. Furthermore, the lack of a frequency band isolation mechanism in the signal transmission path causes low-frequency energy to continuously infiltrate the high-frequency analysis subsystem. In complex events such as grounding faults in photovoltaic power station combiner boxes, the low-frequency components dynamically generated by the system's mechanical inertia can completely overwhelm the microsecond-level transient characteristics of the IGBT switching process, forming a blind spot in stability analysis.

[0004] The information disclosed in this background section is intended only to enhance the understanding of the general background of this disclosure and should not be construed as an admission or in any way implying that the information constitutes prior art known to those skilled in the art. Summary of the Invention

[0005] This invention provides a unified wideband analysis method and system for multi-structure network devices, which can effectively solve the problems in the background art.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A unified broadband analysis method for multi-structure network devices, the method comprising: Collect operating status data and dynamic parameters of multi-network equipment in the power system to generate a wideband trajectory dataset; A deep Koopman neural network model is constructed, comprising an encoder network, a block-diagonalized Koopman operator layer, and a decoder network connected in sequence. The wideband trajectory dataset is input into the deep Koopman neural network model. Dynamic state features are obtained from the encoder network and mapped to a high-dimensional enhanced state space as the input carrier of the block-diagonalized Koopman operator layer. In the high-dimensional improved state space, the broadband dynamic evolution relationship of the multi-structure network device is modeled according to the block diagonalized Koopman operator layer to deduce the temporal change trajectory of the high-dimensional dynamic state characteristics; Dynamic modulation coefficients are generated based on target operating condition parameters, and the intermediate features of the encoder network, the block diagonalized Koopman operator layer, and the decoder network are linearly transformed to perform parameter modulation; wherein the intermediate features include the hidden layer feature vector output by the encoder network, the state evolution output vector of the block diagonalized Koopman operator layer, and the intermediate representation vector of the decoder network. Based on the output of the block-diagonalized Koopman operator layer after parameter modulation, the broadband dynamic trajectory in the physical space of the system is restored according to the decoder network and used to quantize the damping ratio and harmonic distortion rate. The channel participation index is calculated using the gradient of the Jacobian matrix of the decoder network on the gradient of each characteristic mode, and the source device or control channel of the broadband oscillation is located, generating a broadband stability assessment report.

[0007] Furthermore, a deep Koopman neural network model is constructed, including: The multi-step prediction loss function and the state space constraint loss function are jointly optimized. The state space constraint loss function includes a damping marginal constraint term for limiting the magnitude of all eigenvalues ​​and a spectral energy balance regularization term for controlling the energy distribution balance of eigenvalues ​​in low-frequency electromechanical oscillations, subsynchronous oscillations and high-frequency switching harmonic bands. Based on the jointly optimized training results, the feature values ​​of the block-diagonalized Koopman operator layer are obtained to generate a multi-motion modal analytical spectrum. The physical mode parameters of the low-frequency electromechanical oscillation, the subsynchronous oscillation, and the high-frequency switching harmonic band are simultaneously analyzed based on the multi-motion mode analytical spectrum.

[0008] Furthermore, the analytical spectrum of multiple motion modes is obtained, including: In the multi-motion mode analytical spectrum, a mode tracking identifier is configured for each motion mode. The generation rule of the mode tracking identifier is to set the corresponding frequency band category identifier and add a sequential number based on the low-frequency electromechanical oscillation frequency band, the subsynchronous oscillation frequency band and the high-frequency switching harmonic frequency band to which the motion mode belongs. The damping ratio index of each motion mode in the multi-motion modality analytical spectrum is calculated based on the eigenvalues ​​of the block-diagonalized Koopman operator layer. The modal tracking identifier and the corresponding damping ratio index are associated and mapped, and used in the generation process of the broadband stability assessment report.

[0009] Furthermore, the dynamic modulation coefficients are generated via a conditional generator, including: The target operating condition parameters and the dynamic state features output by the encoder network are concatenated as vectors to generate the input vector of the conditional generator. A high-dimensional feature mapping operation is performed on the input vector based on a hidden layer that includes nonlinear transformations to generate an intermediate latent feature representation; Based on the intermediate latent feature representation, three sets of independent coefficient tensors are generated and applied to the linear transformation process of the encoder network, the block diagonalized Koopman operator layer, and the decoder network, respectively, as the dynamic modulation coefficients.

[0010] Furthermore, linear transformation operations include: Based on the first set of coefficients corresponding to the encoder network in the dynamic modulation coefficients, the intermediate features of the encoder network are weighted and scaled to suppress noise and enhance key dynamics; Based on the second set of coefficients corresponding to the block-diagonalized Koopman operator layer in the dynamic modulation coefficients, adjust the representation gain of the broadband dynamic evolution relationship of the block-diagonalized Koopman operator layer in the high-dimensional lifting state space; Based on the third set of coefficients corresponding to the decoder network in the dynamic modulation coefficients, the intermediate features of the decoder network are nonlinearly compensated and modulated to ensure the integrity of the output of the block diagonalized Koopman operator layer after parameter modulation.

[0011] Furthermore, the calculation process for the channel participation index includes: The spatial contribution components corresponding to each feature mode are decomposed based on the Jacobian matrix gradient of the decoder network. Construct the transfer matrix of oscillation energy between the device port and the control channel in the multi-structure network device; Based on the weighted summation of the energy transfer path contribution from the device port to the control channel in the transfer matrix, the channel participation index is generated to locate the source device or control channel of the broadband oscillation.

[0012] Furthermore, the decoder network includes: The decoder network includes parallel-connected linear projection branches and nonlinear refinement branches; The linear projection branch directly maps the output of the block diagonalized Koopman operator layer after parameter modulation to the original physical space to maintain linear dynamic characteristics. The nonlinear refinement branch compensates for amplitude and phase deviations and high-frequency detail loss during the broadband dynamic trajectory restoration process using a multilayer perceptron. The output of the linear projection branch and the output of the nonlinear refinement branch are superimposed to generate the broadband dynamic trajectory in the physical space of the system.

[0013] Furthermore, in the process of modeling the broadband dynamic evolution relationship, an enhanced state-space constraint loss is implemented, including: A soft boundary penalty constraint is applied to the eigenvalue modulus of the block-diagonalized Koopman operator layer to preserve a preset damping margin in the broadband dynamic evolution relationship. Equalization regularization constraints are applied to the energy distribution of spectral blocks in the block-diagonalized Koopman operator layer to maintain full frequency domain coverage of low-frequency modes, subsynchronous modes, and mid-to-high frequency modes.

[0014] Furthermore, the process of mapping to a higher-dimensional lifted state space includes: The dynamic state features are transformed based on a nonlinear coding function to generate an initial boosted state vector. The initial boosted state vector is subjected to time delay embedding processing, and multiple sets of feature vectors in the current time and historical time window are constructed into a time delay embedding state space matrix; Based on a multi-scale feature selector, the low-frequency trend component and high-frequency fluctuation component in the time-delay embedded state space matrix are adaptively weighted and concatenated to generate the final high-dimensional improved state space representation.

[0015] A unified wideband analysis system for multi-structure network devices, the system comprising: The data acquisition module collects operating status data and dynamic parameters of multi-network equipment in the power system and generates a wideband trajectory dataset. The model building unit constructs a deep Koopman neural network model. It inputs a wide frequency domain trajectory dataset into the deep Koopman neural network model, obtains dynamic state features based on the encoder network, and maps the dynamic state features to a high-dimensional improved state space as the input carrier of the block diagonalized Koopman operator layer. The evolution relationship module, in a high-dimensional lifted state space, models the broadband dynamic evolution relationship of multi-structure network devices based on the block-diagonalized Koopman operator layer to deduce the temporal change trajectory of high-dimensional dynamic state characteristics. The linear transformation module generates dynamic modulation coefficients based on the target operating condition parameters and performs linear transformations on the intermediate features of the encoder network, the block diagonalized Koopman operator layer, and the decoder network to perform parameter modulation. The trajectory reconstruction module, based on the output of the block diagonalized Koopman operator layer after parameter modulation, reconstructs the broadband dynamic trajectory in the physical space of the system according to the decoder network, and is used to quantify the damping ratio and harmonic distortion rate. The evaluation results module calculates the channel participation index using the gradient of the Jacobian matrix of the decoder network against the gradient of each characteristic mode, locates the source device or control channel of broadband oscillation, and generates a broadband stability evaluation report.

[0016] The technical solution of this invention can achieve the following technical effects: By implementing cross-frequency band dynamic decoupling modeling through block-diagonalized Koopman operator layers, and combining multi-scale time delay embedded state space construction with a three-channel hierarchical modulation mechanism, the full-frequency domain coupling risks of low-frequency electromechanical oscillations, subsynchronous oscillations, and high-frequency harmonic oscillations are identified synchronously and accurately. This effectively solves the problem of mode mixing and source tracing failure caused by the strong coupling characteristics of wide-frequency domain oscillations in new energy power systems.

[0017] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1This is a flowchart illustrating a unified wideband analysis method for multi-structure network devices. Figure 2 A flowchart illustrating the optimization process of a deep Koopman neural network model; Figure 3 This is a flowchart illustrating a linear transformation operation. Figure 4 This is a flowchart illustrating the process of calculating the channel participation metric. Detailed Implementation

[0020] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0021] Unless otherwise defined, 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. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0022] Example 1; like Figure 1 As shown, this application provides a unified wideband analysis method for multi-structure network devices, the method including: S10: Collect operating status data and dynamic parameters of multi-network equipment in the power system, and generate a wideband trajectory dataset; S20: Construct a deep Koopman neural network model, including an encoder network, a block-diagonalized Koopman operator layer, and a decoder network connected in sequence; input the wide frequency domain trajectory dataset into the deep Koopman neural network model, obtain dynamic state features according to the encoder network, and map the dynamic state features to a high-dimensional enhanced state space as the input carrier of the block-diagonalized Koopman operator layer. S30: In the high-dimensional lifting state space, the broadband dynamic evolution relationship of multi-structure network devices is modeled according to the block diagonalized Koopman operator layer to deduce the temporal change trajectory of the high-dimensional dynamic state characteristics. S40: Generate dynamic modulation coefficients based on target operating condition parameters, and perform linear transformation on the intermediate features of the encoder network, the block diagonalized Koopman operator layer and the decoder network to perform parameter modulation; wherein the intermediate features include the hidden layer feature vector output by the encoder network, the state evolution output vector of the block diagonalized Koopman operator layer and the intermediate representation vector of the decoder network. S50: Based on the output of the block diagonalized Koopman operator layer after parameter modulation, the wideband dynamic trajectory in the physical space of the system is restored according to the decoder network and used to quantize the damping ratio and harmonic distortion rate. S60: Calculate the channel participation index using the gradient of the Jacobian matrix of the decoder network on the gradient of each characteristic mode, locate the source device or control channel of broadband oscillation, and generate a broadband stability assessment report.

[0023] Specifically, firstly, to construct a wideband trajectory dataset, operational status data of multi-network equipment, including doubly-fed induction generator (DFIG) wind turbine converters, photovoltaic inverters, and SVG reactive power compensation devices, across the 0-2000Hz frequency band, were obtained from the power grid dispatch system. This included installing high-frequency measurement units at key nodes in the wind farm to synchronously capture three-phase voltage and current waveforms, recording voltage sags and phase transition events at a sampling rate of 10kHz. Furthermore, millisecond-level electromagnetic dynamic parameters, such as the DC bus voltage ripple characteristics of wind turbine converters and the transient trajectory of IGBT switches in photovoltaic inverters, were extracted. Second-level electromechanical process data, such as the power angle swing curve of synchronous generators, were obtained through a PMU device. A system was then established... The deep Koopman neural network model employs a three-layer cascaded structure. First, there's the encoder network, composed of three layers of bidirectional LSTM units. Its input is a time-domain voltage and current sequence, and its output is mapped to a 512-dimensional high-dimensional state space. The final layer incorporates a sliding time window mechanism, dividing the data into 200ms windows to allow low-frequency components below 150Hz and high-frequency components above 1kHz to enter independent feature channels. The block-diagonalized Koopman operator layer further subdivides the 512-dimensional state space into four 128-dimensional orthogonal subspaces, corresponding to different frequency bands: 0-5Hz, 5-150Hz, 150-1kHz, and 1-2kHz. In each subspace, an independent state evolution matrix is ​​constructed, and the off-diagonal block elements are constrained to zero to prevent cross-frequency energy transfer. For example, when a 50Hz subsynchronous component is detected, only the parameters of the second subspace are adjusted to avoid affecting adjacent high-frequency subspaces. The dynamic modulation mechanism enables parameters to be adjusted in real time according to changes in operating conditions to adapt to different grid conditions. This mechanism receives grid topology parameters, such as renewable energy penetration rate and line impedance, and generates modulation coefficients for the four frequency bands. When wind power output changes rapidly to 30% of the rated value, a weight of 0.6 is applied to the electromechanical dynamic subspace to suppress the low-frequency dominance effect. When a flexible DC commutation failure is detected, the high-frequency frequency is boosted. The response sensitivity of the subspace is weighted to 1.5. Next, the decoder network maps the modulated state space back to the physical quantity space through a deconvolution structure, thereby outputting the broadband trajectory reconstructed by each node. In the oscillation source localization stage, the Jacobian matrix at the decoder output is calculated to quantify the gradient contribution of each device parameter change to the characteristic mode. For example, when a photovoltaic power station experiences 1150Hz resonance, the gradient calculation shows that the participation rate of the inverter DC capacitor parameter channel reaches 82.7%, while that of the SVG control channel is only 5.3%, achieving accurate location of the fault source. Finally, a stability assessment report is generated, marking the energy distribution of oscillations in each frequency band and listing the critical risk devices.

[0024] The technical solution of this invention achieves cross-frequency band dynamic decoupling modeling by dividing the Koopman operator layer into blocks and diagonalizing them. Combined with multi-scale time delay embedded state space construction and three-channel hierarchical modulation mechanism, it synchronously and accurately identifies the full-frequency domain coupling risks of low-frequency electromechanical oscillations, subsynchronous oscillations and high-frequency harmonic oscillations. This effectively solves the problem of mode mixing and source tracing failure caused by the strong coupling characteristics of wide-frequency domain oscillations in new energy power systems.

[0025] Furthermore, such as Figure 2 As shown, the deep Koopman neural network model is constructed, including: The multi-step prediction loss function and the state-space constraint loss function are jointly optimized. The state-space constraint loss function includes a damping marginal constraint term for limiting the magnitude of all eigenvalues ​​and a spectral energy balance regularization term for controlling the energy distribution balance of eigenvalues ​​in low-frequency electromechanical oscillations, subsynchronous oscillations and high-frequency switching harmonic bands. Based on the jointly optimized training results, the feature values ​​of the block-diagonalized Koopman operator layer are obtained to generate a multi-motion modal analytical spectrum. The physical mode parameters of low-frequency electromechanical oscillation, subsynchronous oscillation, and high-frequency switching harmonic bands are simultaneously analyzed based on the multi-motion mode analytical spectrum.

[0026] As a preferred embodiment of the above, firstly, during model training, the multi-step prediction loss function and the improved state-space constraint loss function are jointly optimized. The multi-step prediction loss function focuses on ensuring the accuracy of the model in broadband trajectory prediction. It uses voltage or current time-series data collected from the grid-connected terminal as input to train the model's ability to predict the trajectory of physical quantities at multiple consecutive time steps. By calculating the mean square error between the predicted and actual values, it ensures the fidelity of the observation reconstruction and enables the prediction of the dynamic behavior of the equipment at a series of future moments based on the current state. The improved state-space constraint loss function optimizes the structural stability of the model, including damping marginal constraint terms. The spectral energy balance regularization term and the damping marginal constraint term impose bidirectional soft boundary constraints on the discrete magnitude of the eigenvalues ​​of the Koopman operator, ensuring the stability of the system characteristics and avoiding unstable amplitude growth. Simultaneously, the spectral energy balance regularization term controls the balance of eigenvalue energy distribution among low-frequency electromechanical oscillations, subsynchronous oscillations, and high-frequency switching harmonic bands. By regulating the energy distribution of these bands, the characteristics of each band are fully analyzed without shifting. A preferred embodiment can demonstrate how the regularization term provides a balanced energy distribution across different frequency bands using specific power grid data, thereby enhancing the model's predictive power and applicability. The joint optimization strategy effectively obtains the eigenvalues ​​of the block-diagonalized Koopman operator layer from the optimization results during the training phase, thereby generating detailed multi-motion mode analytical spectra. These spectra provide a refined interpretation of the dynamic characteristics of different frequency bands, offering reliable data support for subsequent analysis. Based on the multi-motion mode analytical spectra, the physical modal parameters of low-frequency electromechanical oscillations, subsynchronous oscillations, and high-frequency switching harmonics in the power system can be simultaneously analyzed. These parameters include vibration frequency, damping ratio, and oscillation intensity, providing accurate physical indicators for the stability assessment of the entire power system. Specifically, it can demonstrate the stability of a wind turbine converter in... In a multi-band coexisting oscillation state, how can the information obtained from the analytical spectrum guide equipment adjustments to prevent the impact of oscillations? Finally, based on the multi-motion mode analytical spectrum, the physical mode parameters of low-frequency electromechanical oscillations, subsynchronous oscillations, and high-frequency switching harmonic bands are further analyzed. For example, during the training process, if the parameters of the subsynchronous oscillation in a wind farm system change, the joint optimization mechanism can adjust the spectral energy distribution in real time and ensure that the analytical spectrum for that specific frequency band accurately reflects the changes in oscillation characteristics. In this process, the adjusted parameter information can be accurately reflected in the final stability analysis report, thereby providing users with effective guidance for equipment adjustment and maintenance.

[0027] Furthermore, obtaining the analytical spectrum of multiple motion modes includes: In the multi-motion mode analytical spectrum, a mode tracking identifier is configured for each motion mode. The generation rule for the mode tracking identifier is to set the corresponding frequency band category identifier and add a sequential number based on the low-frequency electromechanical oscillation frequency band, subsynchronous oscillation frequency band and high-frequency switching harmonic frequency band to which the motion mode belongs. The damping ratio index of each motion mode in the multi-motion modal analytical spectrum is calculated based on the eigenvalues ​​of the block-diagonalized Koopman operator layer. A correlation mapping relationship is established between modal tracking identifiers and their corresponding damping ratio indices, and this relationship is used in the generation process of broadband stability assessment reports.

[0028] As a preferred embodiment of the above, it is first emphasized that each motion mode is effectively tracked and classified by configuring modal tracking identifiers. This step includes assigning a corresponding frequency band category identifier to each mode based on the different frequency bands to which the motion mode belongs, such as the low-frequency electromechanical oscillation band, the subsynchronous oscillation band, and the high-frequency switching harmonic band, and further adding a sequential number. This method provides clear and orderly identifiers for different motion modes, making subsequent analysis and data tracking more intuitive and systematic. On this basis, the eigenvalues ​​of each motion mode are calculated by a block-diagonalized Koopman operator layer, thereby determining the damping ratio index of each motion mode in the multi-motion modal analytical spectrum. The calculation process here includes a detailed analysis of the real and imaginary parts of the eigenvalues, the results of which directly affect the determination of the damping ratio. This index is an important parameter for measuring modal stability and is crucial in the evaluation of system dynamic performance. This process plays a crucial role. In a preferred embodiment, it demonstrates how to assess the vibration resistance of wind turbines or photovoltaic equipment in a power grid by extracting feature values ​​and calculating damping ratios. Next, a mapping relationship is established between the modal tracking identifier and the corresponding damping ratio index. This step ensures that the characteristics and performance indicators of each mode can be accurately traced and verified during system analysis. This mapping relationship is the basis for generating a broadband stability assessment report. By combining the tracking identifier and the damping ratio index, potential stability problems in the system can be accurately identified. For example, in a specific application, a wind farm may experience instability caused by low-frequency electromechanical oscillations. In this case, by analyzing the modal tracking identifier in the analytical spectrum and the detailed analysis of the damping ratio, the root cause of the problem can be quickly and accurately identified, and detailed suggested solutions can be provided in the assessment report to help maintenance personnel make data-driven decisions.

[0029] Furthermore, the dynamic modulation coefficients are generated through a conditional generator, including: The target operating condition parameters and the dynamic state features output by the encoder network are concatenated as vectors to generate the input vector of the conditional generator. The input vector is subjected to a high-dimensional feature mapping operation based on a hidden layer that includes nonlinear transformations, generating an intermediate latent feature representation. Three sets of independent coefficient tensors are generated based on the intermediate hidden feature representation, and are respectively applied to the linear transformation process of the encoder network, the block diagonalized Koopman operator layer, and the decoder network as dynamic modulation coefficients.

[0030] As a preferred embodiment of the above, the target operating condition parameters, including the penetration rate of new energy, the equivalent impedance of the line, the load mutation indicator, and the 512-dimensional dynamic state features output by the encoder network, are first subjected to vector concatenation to form a conditional input vector with a dimension of 2560. For example, when the wind farm output increases sharply, the operating condition label of the power change rate > 0.5 pu / min and the real-time LSTM features are injected. Then, the input vector is fed into a nonlinear transformation network including two hidden layers. The first hidden layer consists of 1024 linear rectifier units with a slope of 0.25, which perform nonlinear sparse mapping on the input features to generate intermediate feature representations. Specifically, the vector elements are subjected to a linear transformation operation with a maximum lower limit of zero. The second hidden layer adopts a compressed excitation structure to adaptively allocate the feature channel weights. The output dimension of this layer is compressed to 256. Finally, Based on this intermediate implicit feature, three sets of independent coefficient tensors are generated: the first tensor is a 384×512-dimensional scaling factor acting on the row vector weights of the encoder network feature transformation matrix, which generates a 1.25-fold high-frequency feature scaling factor when a photovoltaic cluster start-up disturbance is detected; the second tensor is a 128×128-dimensional diagonal weighted coefficient acting on the state evolution matrix of the block-diagonalized Koopman operator subspace, which triggers a 1.8-fold gain in the third high-frequency subspace, such as when a commutation failure event in a flexible DC system triggers the third high-frequency subspace; and the third tensor is a 256×384-dimensional feature modulation vector mapped to the deconvolution kernel of the decoder network. Each set of tensors is generated through an independent linear output layer. For example, in a certain low-frequency oscillation event in the power grid, the three sets of tensors output reference adjustment coefficients of 0.65 / 1.12 / 0.92 respectively, ultimately realizing multi-granularity correlation modulation between changes in grid operating conditions and dynamic responses of equipment.

[0031] Furthermore, such as Figure 3 As shown, the linear transformation operation includes: Based on the first set of coefficients in the dynamic modulation coefficients corresponding to the encoder network, weighted scaling is performed on the intermediate features of the encoder network to suppress noise and enhance key dynamics. Based on the second set of coefficients corresponding to the block-diagonalized Koopman operator layer in the dynamic modulation coefficients, adjust the representation gain of the wideband dynamic evolution relationship of the block-diagonalized Koopman operator layer in the high-dimensional lifting state space. Based on the third set of coefficients corresponding to the decoder network in the dynamic modulation coefficients, nonlinear compensatory modulation is applied to the intermediate features of the decoder network to ensure the integrity of the output results of the block diagonalized Koopman operator layer after parameter modulation.

[0032] As a preferred embodiment of the above, the model layers are optimized in a targeted manner using the dynamic modulation coefficients. Specifically, for the intermediate features of the encoder network, namely the dynamic state features extracted from the output of the bidirectional LSTM network layer, a weighted scaling operation is performed based on the first set of coefficients. This first set of coefficients is a 384×512-dimensional scaling matrix, whose elements correspond to the weight values ​​of each dimension of the feature vector. The intermediate feature vector is scaled element by element through matrix multiplication. This process achieves noise suppression and key dynamic enhancement in the feature space. Specifically, the weight of high-frequency background noise components is reduced to below 0.1, and the weight of low-frequency dynamic components, including phase abrupt changes or transient events, is increased to more than 1.2 times. In the example, when the power grid experiences a 20% load drop, this scaling operation can enhance the 0-5Hz electromechanical oscillation features, increasing the weight to 1.15 times, while suppressing switching noise above 1kHz, reducing the weight to 0.05, significantly improving the feature signal-to-noise ratio. Subsequently, a second set of coefficients is applied to the block-diagonalized Koopman operator layer. The coefficients are used to adjust the gain of the broadband dynamic evolution relationship: The second set of coefficients is a 128×128-dimensional diagonal weighted matrix, corresponding to the diagonal elements of the state evolution matrix of each of the four subspaces in this layer. The subspaces are associated with independent frequency bands such as 5-150Hz subsynchronous oscillations. The enhancement amplitude of the state update step size is changed through matrix cascading operations to optimize the evolution accuracy of broadband dynamics. In actual operation, when a 1160Hz resonance event is detected in the flexible DC converter station, the coefficients are assigned to increase the weighting value of the diagonal elements of the 1-2kHz frequency band in the high-frequency subspace to 1.8 times the gain coefficient, which accelerates the dynamic response of this frequency band and avoids conflicts with other subspaces. Finally, the intermediate features of the decoder network are embedded with nonlinear compensatory modulation based on the third set of coefficients: The third set of coefficients is a 256×384-dimensional feature modulation vector, which is applied to the feature channels before the deconvolution operation. Through the combination of vector superposition and specific activation functions, including piecewise linear operation to suppress zeros, the frequency and phase distortion caused by parameter modulation is compensated.

[0033] Furthermore, such as Figure 4 As shown, the calculation process for the channel participation index includes: The spatial contribution components of each feature mode are decomposed by the Jacobian matrix gradient decomposition based on the decoder network. Construct a transfer matrix for oscillation energy between device ports and control channels in multi-structure network devices; Based on the energy transfer path contribution from the device port to the control channel in the weighted cumulative transfer matrix, a channel participation index is generated to locate the source device or control channel of broadband oscillation.

[0034] As a preferred embodiment of the above, the channel participation index calculation process is used to accurately locate the source device or control channel of broadband oscillation. First, a gradient decomposition operation is performed based on the Jacobian matrix of the decoder network. The Jacobian matrix is ​​generated in the partial derivative mapping of the input features at the model output. By parsing the gradient vector row by row, the matrix elements represent the correlation sensitivity between data points, and eigenvalue decomposition is applied to separate key feature modes. These modes correspond to the core dynamic components in the broadband domain. Specifically, the gradient matrix is ​​projected column by column onto the orthogonal basis vector group to decompose multiple spatial contribution components of different frequency bands. For example, the contribution component vector of the 25Hz subsynchronous mode comes from 85% of the variance explained in the first column. In the example, when the wind farm experiences a 50Hz forced oscillation, this method can separate the mode component corresponding to the wind speed change, and its spatial weight of 72% is significantly higher than others. Next, the transfer matrix of oscillation energy between the device port and the control channel in the multi-network device is constructed. This transfer matrix is ​​in the form of a square matrix, and its dimension is equal to the total number of device ports and control channels to be analyzed. The matrix row index is... The physical connection ports of the equipment are represented, such as the DC terminal ports of the converter of a doubly fed wind turbine. The column index represents the control logic channel, such as the voltage regulation loop channel of an SVG device. The value of each element is determined by calculating the coupling strength in the energy transfer path. Specifically, the signal transmission data streams between the port and the channel are filtered and the frequency domain correlation coefficient is calculated. Then, the data is normalized to the 0-1 interval to characterize the relative energy intensity. In the application, for a 950Hz switching harmonic event in a coastal wind farm, a transfer matrix is ​​constructed, including elements from the wind turbine port to the flexible DC control channel and elements from the photovoltaic series resonant channel, reflecting the interleaved influence of multi-source signals. Finally, the channel participation index is generated by weighted summing of the energy transfer path contribution of each equipment port to the control channel in the transfer matrix. This index is a numerical scalar and is calculated independently for a specific equipment or channel. First, a dynamic weighting coefficient is applied to the target path elements. The coefficient comes from the normalized weight of the aforementioned spatial contribution components. Then, all port-to-channel path elements are summed, and a sorted list is output to locate the dominant source. A contribution greater than 50% is considered a major source point.

[0035] Furthermore, the decoder network includes: The decoder network consists of parallel-connected linear projection branches and nonlinear refinement branches; The linear projection branch directly maps the output of the parameter-modulated block diagonalized Koopman operator layer to the original physical space to maintain linear dynamic characteristics. The nonlinear refinement branch compensates for amplitude and phase deviations and high-frequency detail loss during broadband dynamic trajectory reconstruction using a multilayer perceptron. The output of the linear projection branch is superimposed with the output of the nonlinear refinement branch to generate a wideband dynamic trajectory in the physical space of the system.

[0036] As a preferred embodiment of the above, the decoder network adopts a parallel cooperative structure of linear projection branch and nonlinear refinement branch. The outputs of the two are superimposed to form the final dynamic trajectory. The linear projection branch is responsible for maintaining the linear essential property of broadband dynamics: this branch is implemented by a single-layer mapping operation, which directly maps the output of the modulated block diagonalized Koopman operator layer, that is, the discretized state evolution result, to the original physical space. Specifically, the subspace state vector is reconstructed by linear combination through orthogonal basis transformation matrix. For example, when analyzing the 500Hz switching harmonics of a doubly fed wind turbine, this branch maps the high-frequency subspace state quantity to three-phase current physical quantities. The waveform phase error is less than 1.5 degrees, but there is high-frequency ripple distortion. This process completely preserves the linear dynamic characteristics in the state space, such as the phase continuity of characteristic frequencies; while the nonlinear refinement branch... The network employs a multilayer perceptron structure with three layers. The first hidden layer contains 512 linear rectifier units with a slope of 0.3, while the second hidden layer uses 328 neurons with a random inactivation rate of 10% to compensate for amplitude and phase deviations and high-frequency losses in trajectory reconstruction. This branch takes the same state vector as input and first performs a feature dimension upscaling transformation from 128 to 256 dimensions. Then, it identifies complex coupling features through two nonlinear transformations, particularly targeting high-frequency attenuation and local nonlinear distortions. Finally, it outputs a compensation vector with dimensions consistent with the original physical space. Finally, it performs a dual-branch output superposition operation: the basic trajectory output from the linear projection branch and the compensation vector output from the nonlinear refinement branch are added element-wise. Both have dimensions of 512 original physical variables, and the superposition weight is set to 1:1 by default. The superposition result is constrained by an anti-saturation operator to limit the amplitude range.

[0037] Furthermore, in the process of modeling broadband dynamic evolution relationships, the state-space constraint loss is improved by implementing measures including: A soft boundary penalty constraint is applied to the eigenvalue modulus of the block-diagonalized Koopman operator layer to preserve the preset damping margin in the broadband dynamic evolution relationship. Equalization regularization constraints are applied to the energy distribution of spectral blocks in the block-diagonalized Koopman operator layer to maintain full-frequency domain coverage of low-frequency modes, subsynchronous modes, and mid-to-high-frequency modes.

[0038] As a preferred embodiment of the above, when modeling broadband dynamic evolution relationships, the physical rationality of the model is enhanced by increasing the state space constraint loss, and the constraint is applied to the training process of the block-diagonalized Koopman operator layer; a soft boundary penalty constraint is applied to the eigenvalue magnitude: the complex plane magnitude value of each subspace eigenvalue is calculated, and the magnitude of the eigenvalue of the state evolution matrix reflects the dynamic stability. A bidirectional tolerance interval is set based on a preset damping margin threshold, and an asymptotic quadratic penalty term is applied to eigenvalues ​​that exceed the interval. The deviation value is amplified by a quadratic function to increase the loss weight, and linear decay is used near the boundary of the tolerance interval to smooth the gradient; a balanced regularization constraint is applied to the spectral block energy distribution: The four spectral blocks of the block-diagonalized operator layer are independently statistically analyzed and associated with the output energies of low-frequency (0-15Hz) electromechanical dynamics, subsynchronous (15-150Hz) oscillations, mid-frequency (150-1000Hz) harmonics, and high-frequency (1-3kHz) switching transients. The sum of squares and averages of the output vectors of each block are calculated. A ternary loss function is used to force the energy ratio of the low-frequency and high-frequency bands to remain within a set range, such as requiring the energy difference between the subsynchronous band and the switching transient band to not exceed 20 dB. The specific loss function is composed of the standard deviation of the logarithmic difference of the energy of each block. The dispersion of the energy distribution of all samples is statistically analyzed. These two constraints are jointly optimized, and the operator layer parameters are adjusted through backpropagation.

[0039] Furthermore, the process of mapping to a higher-dimensional improved state space includes: An initial boosted state vector is generated by transforming dynamic state features based on a nonlinear coding function. The initial boosted state vector is subjected to time-delay embedding processing, and multiple sets of feature vectors within the current time and historical time window are constructed into a time-delay embedding state space matrix. Based on a multi-scale feature selector, the low-frequency trend component and high-frequency fluctuation component in the time-delay embedded state space matrix are adaptively weighted and concatenated to generate the final high-dimensional improved state space representation.

[0040] As a preferred embodiment of the above, the original features are first transformed by a nonlinear coding function. This function consists of a two-layer transformation structure. The first layer is an affine transformation containing 256 neural units, with a modified linear unit activation at a slope of 0.2. The second layer is a fully connected operation that compresses the dimension to 128 and uses hyperbolic tangent output normalization. Next, time-delay embedding processing is performed: the initial boost vector within a continuous preset time window is obtained. The window length is adaptively set according to the target frequency band; for example, analyzing a 1kHz switching ripple requires no less than 20 sampling points. The current time vector and the historical vectors of a specific step length are arranged in reverse chronological order and concatenated into a time-delay embedding state space matrix. The number of rows is the feature dimension of 128, and the number of columns is the time series length. For example, in the 750Hz circulating current tracing of a certain energy storage power station, 32 sets of historical data within a 50ms window are used. The historical vector and the current vector are concatenated into a 128×33 dimensional matrix, effectively preserving the phase evolution information of the transient process. Finally, an adaptive weighted concatenation is achieved through a multi-scale feature selector: this selector includes a parallel dual-path filtering structure. The first path uses a low-pass filter with a time constant of 0.5 seconds to extract low-frequency trend components and obtain the slow dynamics from the fundamental frequency to 50Hz. The second path uses a band-stop filter with a stopband attenuation of 60 dB to separate high-frequency fluctuation components and focus on fast-changing disturbances above 150Hz. The outputs of the two are weighted by independent weight coefficients, and the weights are determined by historical data statistics: the low-frequency weight value is positively correlated with the reciprocal of the mean of the most recent data, and the high-frequency weight is proportional to the logarithm of the local signal variance. Then, channel-level concatenation is performed to finally generate a 256-dimensional high-dimensional enhanced state vector, thereby realizing the global optimized expression of wide-frequency domain dynamic features.

[0041] Example 2; Based on the same inventive concept as the broadband unified analysis method for multi-network devices in the foregoing embodiments, the present invention also provides a broadband unified analysis system for multi-network devices, the system comprising: The data acquisition module collects operating status data and dynamic parameters of multi-network equipment in the power system and generates a wideband trajectory dataset. The model building unit constructs a deep Koopman neural network model. It inputs a wide frequency domain trajectory dataset into the deep Koopman neural network model, obtains dynamic state features based on the encoder network, and maps the dynamic state features to a high-dimensional improved state space as the input carrier of the block diagonalized Koopman operator layer. The evolution relationship module, in a high-dimensional lifted state space, models the broadband dynamic evolution relationship of multi-structure network devices based on the block-diagonalized Koopman operator layer to deduce the temporal change trajectory of high-dimensional dynamic state characteristics. The linear transformation module generates dynamic modulation coefficients based on the target operating condition parameters and performs linear transformations on the intermediate features of the encoder network, the block diagonalized Koopman operator layer, and the decoder network to perform parameter modulation. The trajectory reconstruction module, based on the output of the block diagonalized Koopman operator layer after parameter modulation, reconstructs the broadband dynamic trajectory in the physical space of the system according to the decoder network, and is used to quantify the damping ratio and harmonic distortion rate. The evaluation results module calculates the channel participation index using the gradient of the Jacobian matrix of the decoder network against the gradient of each characteristic mode, locates the source device or control channel of broadband oscillation, and generates a broadband stability evaluation report.

[0042] The adjustment system described above in this invention can effectively realize a unified analysis method for wideband domain of multi-structure network devices, and the technical effects it can achieve are as described in the above embodiments, which will not be repeated here.

[0043] Although this application has been described in conjunction with specific features and embodiments, it is obvious that various modifications and combinations can be made thereto without departing from the spirit and scope of this application. Accordingly, this specification and drawings are merely exemplary illustrations of the application as defined herein, and are to be considered as covering any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Thus, if such modifications and modifications fall within the scope of this application and its equivalents, this application intends to include such modifications and modifications.

Claims

1. A unified wideband analysis method for multi-structure network devices, characterized in that, The method includes: Collect operating status data and dynamic parameters of multi-network equipment in the power system to generate a wideband trajectory dataset; A deep Koopman neural network model is constructed, comprising an encoder network, a block-diagonalized Koopman operator layer, and a decoder network connected in sequence. The wideband trajectory dataset is input into the deep Koopman neural network model. Dynamic state features are obtained from the encoder network and mapped to a high-dimensional enhanced state space as the input carrier of the block-diagonalized Koopman operator layer. In the high-dimensional improved state space, the broadband dynamic evolution relationship of the multi-structure network device is modeled according to the block diagonalized Koopman operator layer to deduce the temporal change trajectory of the high-dimensional dynamic state characteristics; Dynamic modulation coefficients are generated based on target operating condition parameters, and the intermediate features of the encoder network, the block diagonalized Koopman operator layer, and the decoder network are linearly transformed to perform parameter modulation; wherein the intermediate features include the hidden layer feature vector output by the encoder network, the state evolution output vector of the block diagonalized Koopman operator layer, and the intermediate representation vector of the decoder network. Based on the output of the block-diagonalized Koopman operator layer after parameter modulation, the broadband dynamic trajectory in the physical space of the system is restored according to the decoder network and used to quantize the damping ratio and harmonic distortion rate. The channel participation index is calculated using the gradient of the Jacobian matrix of the decoder network on the gradient of each characteristic mode, and the source device or control channel of the broadband oscillation is located, generating a broadband stability assessment report.

2. The broadband unified analysis method for multi-structure network devices according to claim 1, characterized in that, Constructing a deep Koopman neural network model includes: The multi-step prediction loss function and the state space constraint loss function are jointly optimized. The state space constraint loss function includes a damping marginal constraint term for limiting the magnitude of all eigenvalues ​​and a spectral energy balance regularization term for controlling the energy distribution balance of eigenvalues ​​in low-frequency electromechanical oscillations, subsynchronous oscillations and high-frequency switching harmonic bands. Based on the jointly optimized training results, the feature values ​​of the block-diagonalized Koopman operator layer are obtained to generate a multi-motion modal analytical spectrum. The physical mode parameters of the low-frequency electromechanical oscillation, the subsynchronous oscillation, and the high-frequency switching harmonic band are simultaneously analyzed based on the multi-motion mode analytical spectrum.

3. The broadband unified analysis method for multi-structure network devices according to claim 2, characterized in that, Obtaining the analytical spectrum of multiple motion modes, including: In the multi-motion mode analytical spectrum, a mode tracking identifier is configured for each motion mode. The generation rule of the mode tracking identifier is to set the corresponding frequency band category identifier and add a sequential number based on the low-frequency electromechanical oscillation frequency band, the subsynchronous oscillation frequency band and the high-frequency switching harmonic frequency band to which the motion mode belongs. The damping ratio index of each motion mode in the multi-motion modality analytical spectrum is calculated based on the eigenvalues ​​of the block-diagonalized Koopman operator layer. The modal tracking identifier and the corresponding damping ratio index are associated and mapped, and used in the generation process of the broadband stability assessment report.

4. The broadband unified analysis method for multi-structure network devices according to claim 1, characterized in that, Dynamic modulation coefficients are generated based on target operating condition parameters, including: The target operating condition parameters and the dynamic state features output by the encoder network are concatenated as vectors to generate the input vector of the conditional generator. A high-dimensional feature mapping operation is performed on the input vector based on a hidden layer that includes nonlinear transformations to generate an intermediate latent feature representation; Based on the intermediate latent feature representation, three sets of independent coefficient tensors are generated and applied to the linear transformation process of the encoder network, the block diagonalized Koopman operator layer, and the decoder network, respectively, as the dynamic modulation coefficients.

5. The broadband unified analysis method for multi-structure network devices according to claim 1, characterized in that, Linear transformations, including: Based on the first set of coefficients corresponding to the encoder network in the dynamic modulation coefficients, the intermediate features of the encoder network are weighted and scaled to suppress noise and enhance key dynamics; Based on the second set of coefficients corresponding to the block-diagonalized Koopman operator layer in the dynamic modulation coefficients, adjust the representation gain of the broadband dynamic evolution relationship of the block-diagonalized Koopman operator layer in the high-dimensional lifting state space; Based on the third set of coefficients corresponding to the decoder network in the dynamic modulation coefficients, the intermediate features of the decoder network are nonlinearly compensated and modulated to ensure the integrity of the output of the block diagonalized Koopman operator layer after parameter modulation.

6. The broadband unified analysis method for multi-structure network devices according to claim 1, characterized in that, The channel participation index is calculated using the gradient of the Jacobian matrix of the decoder network with respect to the gradient of each feature mode, including: The spatial contribution components corresponding to each feature mode are decomposed based on the Jacobian matrix gradient of the decoder network. Construct the transfer matrix of oscillation energy between the device port and the control channel in the multi-structure network device; Based on the weighted summation of the energy transfer path contribution from the device port to the control channel in the transfer matrix, the channel participation index is generated to locate the source device or control channel of the broadband oscillation.

7. The broadband unified analysis method for multi-network devices according to claim 1, characterized in that, The decoder network includes: The decoder network includes parallel-connected linear projection branches and nonlinear refinement branches; The linear projection branch directly maps the output of the block diagonalized Koopman operator layer after parameter modulation to the original physical space to maintain linear dynamic characteristics. The nonlinear refinement branch compensates for amplitude and phase deviations and high-frequency detail loss during the broadband dynamic trajectory restoration process using a multilayer perceptron. The output of the linear projection branch and the output of the nonlinear refinement branch are superimposed to generate the broadband dynamic trajectory in the physical space of the system.

8. The broadband unified analysis method for multi-structure network devices according to claim 1, characterized in that, In the process of modeling the broadband dynamic evolution relationship, the state space constraint loss is improved, including: A soft boundary penalty constraint is applied to the eigenvalue modulus of the block-diagonalized Koopman operator layer to preserve a preset damping margin in the broadband dynamic evolution relationship. Equalization regularization constraints are applied to the energy distribution of spectral blocks in the block-diagonalized Koopman operator layer to maintain full frequency domain coverage of low-frequency modes, subsynchronous modes, and mid-to-high frequency modes.

9. The broadband unified analysis method for multi-structure network devices according to claim 1, characterized in that, The process of mapping to a higher-dimensional improved state space includes: The dynamic state features are transformed based on a nonlinear coding function to generate an initial boosted state vector. The initial boosted state vector is subjected to time delay embedding processing, and multiple sets of feature vectors in the current time and historical time window are constructed into a time delay embedding state space matrix; Based on a multi-scale feature selector, the low-frequency trend component and high-frequency fluctuation component in the time-delay embedded state space matrix are adaptively weighted and concatenated to generate the final high-dimensional improved state space representation.

10. A unified wideband analysis system for multi-structure network devices, characterized in that, The system includes: The data acquisition module collects operating status data and dynamic parameters of multi-network equipment in the power system and generates a wideband trajectory dataset. The model building unit constructs a deep Koopman neural network model. It inputs a wide frequency domain trajectory dataset into the deep Koopman neural network model, obtains dynamic state features based on the encoder network, and maps the dynamic state features to a high-dimensional improved state space as the input carrier of the block diagonalized Koopman operator layer. The evolution relationship module, in a high-dimensional enhanced state space, models the broadband dynamic evolution relationship of multi-structure network devices based on the block-diagonalized Koopman operator layer to deduce the temporal change trajectory of high-dimensional dynamic state characteristics. The linear transformation module generates dynamic modulation coefficients based on the target operating condition parameters and performs linear transformations on the intermediate features of the encoder network, the block diagonalized Koopman operator layer, and the decoder network to perform parameter modulation. The trajectory reconstruction module, based on the output of the block diagonalized Koopman operator layer after parameter modulation, reconstructs the broadband dynamic trajectory in the physical space of the system according to the decoder network, and is used to quantify the damping ratio and harmonic distortion rate. The evaluation results module calculates the channel participation index using the gradient of the Jacobian matrix of the decoder network against the gradient of each characteristic mode, locates the source device or control channel of broadband oscillation, and generates a broadband stability evaluation report.

Images