Simulation method and system for grid-connected operation of new energy station

Abstract

The invention provides an analogue simulation method and system for grid-connected operation of a new energy station, and relates to the technical field of new energy power system simulation, and the method comprises the steps: obtaining the operation configuration and power grid access information of the station, extracting the characteristics of a power generation unit, calculating the electrical similarity, and carrying out the clustering grouping and equivalent parameter construction; and identifying a node cluster boundary to construct a partition topology, executing asynchronous parallel simulation, and performing time alignment processing on boundary interaction data. According to the method, the simulation calculation complexity can be effectively reduced, the grid-connected simulation efficiency of the large-scale new energy station is improved, and the calculation precision and the real-time performance are ensured.

Description

Technical Field

[0001] This invention relates to the field of simulation technology for new energy power systems, and in particular to a simulation method and system for the grid-connected operation of new energy power plants. Background Technology

[0002] With the rapid development of renewable energy, the number of new energy power plants is increasing daily, and their grid-connected operation has a significant impact on the stability of the power system. New energy power plants consist of numerous distributed generation units, exhibiting intermittent, fluctuating, and uncertain characteristics, resulting in complex and varied impacts on the power grid. To ensure the safe and stable operation of the power system, it is necessary to conduct simulation analysis of the grid-connected operation of new energy power plants to predict their impact on the power grid and formulate corresponding control strategies.

[0003] Traditional grid connection simulation methods for new energy power plants mainly adopt the full-model direct simulation approach, which incorporates all power generation units and their control systems into the simulation model. This results in a large-scale simulation system with high computational complexity. As the installed capacity of new energy continues to expand, the full-model direct simulation approach can hardly meet the requirements of real-time simulation.

[0004] Currently, the simulation technology for grid-connected operation of new energy power plants still has problems such as the inability to effectively simplify large-scale new energy power plants, resulting in heavy computational burden, low simulation efficiency, lack of adaptive clustering methods based on physical characteristics, difficulty in achieving a good balance between accuracy and efficiency, and inability to perform targeted optimization based on the dynamic characteristics of different sub-network regions, resulting in low simulation speed and accuracy. Summary of the Invention

[0005] This invention provides a simulation method and system for the grid-connected operation of new energy power plants, which can at least solve some of the problems existing in the prior art.

[0006] A first aspect of this invention provides a simulation method for the grid-connected operation of a new energy power station, comprising: Obtain the operation configuration information of the new energy power station and the grid access point information. Based on the operation configuration information, extract the operation status features of the power generation unit and construct a time-series feature matrix. Extract the dominant feature components corresponding to the time-series feature matrix to construct a feature vector set and calculate the electrical similarity index between the power generation units. Based on the electrical similarity index, the power generation units are grouped into multiple power generation groups using an agglomerative hierarchical clustering algorithm. An adaptive weighting coefficient is calculated based on the electrical similarity index, and the electrical impedance characteristics of each power generation unit within the power generation group are weighted and aggregated to obtain equivalent electrical parameters. Based on the equivalent electrical parameters, equivalent nodes are constructed and a simplified topology is determined. Based on the access point information, the boundary constraints on the power grid side are determined and applied to the simplified topology. The Laplace matrix corresponding to the simplified topology is determined and a spectral embedding space is constructed. The equivalent nodes are mapped to the spectral embedding space to identify the node cluster boundaries and the simplified topology is divided into multiple sub-network regions to obtain a partitioned topology. Based on the partitioned topology, an independent simulation step size is configured for each sub-network region, and asynchronous parallel simulation is performed. Boundary node data is extracted at a preset synchronization time. After time alignment processing of the boundary interaction data using an extrapolation prediction algorithm, the data is transmitted to adjacent sub-network regions to obtain the global simulation results.

[0007] In one alternative implementation, The process involves acquiring the operation configuration information of the renewable energy power station and the grid connection point information. Based on the operation configuration information, the operation status features of the power generation units are extracted and a time-series feature matrix is ​​constructed. The dominant feature components corresponding to the time-series feature matrix are extracted to construct a feature vector set, and the electrical similarity index between the power generation units is calculated, including: The power output sequence and control response sequence of each power generation unit under multiple operating conditions are collected, and the timestamps are aligned and abnormal data points are removed to obtain the effective power sequence and effective control sequence. The time sequence feature matrix is ​​then constructed by splicing the sequences according to the time dimension. The time series feature matrix is ​​standardized to obtain a standardized time series matrix. A covariance matrix is ​​constructed based on the standardized time series matrix and eigenvalue decomposition is performed. The eigenvalues ​​are sorted in descending order according to their size and the eigenvectors corresponding to the top few eigenvalues ​​are selected as principal component directions based on the cumulative contribution rate. The standardized time series matrix is ​​then projected onto the principal component directions to obtain the dominant feature components. Extract the dominant feature components corresponding to each power generation unit and organize them according to the power generation unit number to construct the feature vector set. Perform covariance calculation on the feature vectors of any two power generation units in the feature vector set and construct a covariance matrix. Perform matrix inversion operation on the covariance matrix to obtain the inverse covariance matrix. Calculate the difference vector between the feature vectors of different power generation units and perform weighted distance calculation with the corresponding inverse covariance matrix to obtain the Mahalanobis distance. Perform normalization mapping on the Mahalanobis distance to obtain the electrical similarity index.

[0008] In one alternative implementation, Based on the electrical similarity index, the power generation units are grouped into multiple power generation groups using an agglomerative hierarchical clustering algorithm. Adaptive weighting coefficients are calculated based on the electrical similarity index, and the electrical impedance characteristics of each power generation unit within a power generation group are weighted and aggregated to obtain equivalent electrical parameters. Based on these equivalent electrical parameters, equivalent nodes are constructed, and a simplified topology is determined, including: Based on the electrical similarity index, a similarity matrix is ​​constructed and the power generation units are treated as independent clusters. The power generation unit pairs with the highest similarity in the similarity matrix are calculated and cluster merging operations are performed to obtain merged clusters and update the similarity matrix. The clustering quality is calculated by combining the silhouette coefficient evaluation algorithm. The cluster merging operation is repeated until the clustering quality converges to obtain multiple power generation groups. Calculate the electrical similarity index between each power generation unit in the power generation group and the center of the power generation group, and perform reverse mapping to obtain adaptive weighting coefficients. Extract the resistance parameters and reactance parameters of each power generation unit in the power generation group from the operation configuration information and construct an electrical impedance feature vector. An adjacency matrix is ​​constructed based on the electrical connection relationship between each power generation unit within the power generation group. A symmetric transformation is performed on the adjacency matrix using a graph convolution propagation algorithm to obtain a symmetric adjacency matrix. Multi-hop propagation is then performed on the electrical impedance feature vector to obtain multi-order neighborhood features. Layer-by-layer aggregation is performed on the multi-order neighborhood features to obtain neighborhood aggregation features. Weighted summation and nonlinear activation transformation operations are then performed on the adaptive weighting coefficients to obtain equivalent electrical parameters. The equivalent node is constructed for each power generation group based on the equivalent electrical parameters. The connection topology between the power generation unit and the combiner unit in the operation configuration information is extracted. The simplified topology is constructed based on the equivalent node and the connection topology.

[0009] In one alternative implementation, The adjacency matrix is ​​subjected to a symmetric transformation using a graph convolution propagation algorithm to obtain a symmetric adjacency matrix. This symmetric matrix is ​​then combined with the electrical impedance feature vector to perform a multi-hop propagation operation, resulting in multi-order neighborhood features, including: The connection relationship information of each power generation unit node in the adjacency matrix is ​​extracted and the node degree is counted to construct a degree matrix. The degree matrix is ​​subjected to an inverse transformation to obtain a degree inverse matrix, and the adjacency matrix is ​​subjected to a two-sided symmetric transformation to obtain a symmetric adjacency matrix. A topology-aware mask matrix is ​​constructed based on the symmetric adjacency matrix. The electrical impedance feature vector is used as a zero-order feature input. The capacity parameters and impedance parameters of each power generation unit are extracted from the operation configuration information and a node attribute embedding vector is constructed. The neighborhood propagation feature is obtained by performing neighborhood information aggregation on the current layer feature vector and the symmetric adjacency matrix using a graph convolution algorithm. The neighborhood propagation feature is then fused with the node attribute embedding vector using a feature fusion algorithm to obtain an enhanced propagation feature. The enhanced propagation feature is then subjected to a learnable weight transformation and combined with the topology-aware mask matrix to perform selective feature filtering to obtain the current layer output feature. The residual enhanced feature is then obtained by solving based on the current layer output feature and the current layer input feature. Determine whether the current propagation layer has reached the preset maximum propagation layer. If not, use the residual enhancement feature as the input feature of the next layer and repeat the neighborhood propagation operation. If it has reached the maximum, extract the residual enhancement features output by each propagation layer and calculate the hierarchical importance weight based on the propagation depth. Aggregate the residual enhancement features of each propagation layer with the corresponding hierarchical importance weight to obtain multi-level neighborhood features.

[0010] In one alternative implementation, Based on the access point information, boundary constraints on the power grid side are determined and applied to the simplified topology. The Laplace matrix corresponding to the simplified topology is determined, and a spectral embedding space is constructed. The equivalent nodes are mapped to the spectral embedding space to identify node cluster boundaries. The simplified topology is then divided into multiple sub-network regions to obtain a partitioned topology, including: Extract the access point location identifier and grid-side access capacity limitation information from the access point information; construct a power constraint set based on the grid-side access capacity limitation information and bind it to the equivalent node in the simplified topology corresponding to the access point location identifier; construct a boundary node label matrix based on the equivalent node that has been bound to the power constraint set. Extract the connection relationships between equivalent nodes in the simplified topology and construct a connection matrix. Calculate the number of connections for each equivalent node based on the connection matrix, construct a node degree vector, and convert it into a diagonal matrix to obtain a degree diagonal matrix. Perform a matrix difference operation between the degree diagonal matrix and the connection matrix to obtain the Laplace matrix. The eigenvalue decomposition of the Laplacian matrix is ​​performed to obtain the eigenvector matrix, and the spectral embedding space is constructed based on the eigenvector matrix; The equivalent nodes in the simplified topology are projected onto the spectral embedding space to obtain spectral coordinate representations. A spectral similarity matrix is ​​constructed using a spectral clustering algorithm. The spectral similarity matrix is ​​constrained and adjusted based on the boundary node label matrix, and the boundary positions between dense and sparse spectral coordinate regions are identified. The node cluster boundaries are determined based on the boundary positions, and the equivalent nodes in the simplified topology are divided into different sub-network regions. A sub-network region identifier mapping table is constructed, and the simplified topology is decomposed based on the sub-network region identifier mapping table to obtain a partitioned topology.

[0011] In one alternative implementation, Based on the partitioned topology, an independent simulation step size is configured for each sub-network region, and asynchronous parallel simulation is performed. Boundary node data is extracted at a preset synchronization time. After time alignment processing of the boundary interaction data using an extrapolation prediction algorithm, the data is transmitted to adjacent sub-network regions to obtain global simulation results, including: The topology complexity parameters and node quantity parameters of each sub-network region are extracted from the partition topology structure. The computational load of each sub-network region is calculated based on the topology complexity parameters and the node quantity parameters. Based on the computational load and the coupling strength between different sub-network regions, an independent simulation step size is allocated to each sub-network region through a load balancing optimization algorithm, and a step size configuration table is constructed based on the independent simulation step size. Identify the boundary connection relationships between each sub-network region from the partition topology and extract the boundary node identifiers, and construct a boundary node list based on the boundary node identifiers; According to the step size configuration table, each sub-network region is driven to perform asynchronous parallel simulation according to the corresponding independent simulation step size. When the simulation time of each sub-network region reaches the preset synchronization time, the corresponding boundary node data is extracted from the boundary node list as boundary interaction data and the corresponding actual extraction time is recorded. Based on the time deviation between the actual extraction time and the preset synchronization time, the boundary interaction data is time-aligned by an extrapolation prediction algorithm to obtain synchronized boundary data. According to the boundary connection relationship, the synchronized boundary data is passed to the adjacent sub-network region as boundary conditions. Asynchronous parallel simulation is repeated until the simulation ends. The simulation output data of each sub-network region is summarized to obtain the global simulation result.

[0012] In one alternative implementation, The topology complexity parameters and node count parameters of each sub-network region are extracted from the partitioned topology. The computational load of each sub-network region is calculated based on these parameters. Based on the computational load and the coupling strength between different sub-network regions, an independent simulation step size is allocated to each sub-network region using a load balancing optimization algorithm, including: The node count parameter is obtained by traversing each sub-network region in the partition topology and counting the total number of nodes in each sub-network region. The connection complexity between nodes in each sub-network region is calculated and combined with the topology level depth in the sub-network region to obtain the topology complexity parameter. The computational load of each sub-network region is calculated based on the node count parameter and the topology complexity parameter. Extract the number of boundary node connections between different sub-network regions in the partitioned topology, count the power exchange frequency of boundary nodes between each sub-network region and its adjacent sub-network regions, calculate the coupling strength between different sub-network regions based on the number of boundary node connections and the power exchange frequency of boundary nodes, and construct a coupling strength matrix. Based on the computational load and the coupling strength matrix, an optimization function is constructed with the goal of minimizing the total simulation time using a load balancing optimization algorithm. Electrical characteristic parameters of each sub-network region are extracted from the partitioned topology. Based on the electrical characteristic parameters and the preset synchronization time, a simulation numerical stability criterion is constructed. The simulation numerical stability criterion is added to the optimization function as a simulation stability constraint and the independent simulation step size of each sub-network region is obtained by solving the algorithm.

[0013] A second aspect of the present invention provides a simulation system for the grid-connected operation of a new energy power station, comprising: The feature clustering module is used to obtain the operation configuration information of the new energy power station and the access point information of the grid side. Based on the operation configuration information, the module extracts the operation status features of the power generation unit and constructs a time-series feature matrix. It extracts the dominant feature components corresponding to the time-series feature matrix to construct a feature vector set and calculates the electrical similarity index between the power generation units. The topology simplification module is used to group the power generation units into multiple power generation groups based on the electrical similarity index using an agglomerative hierarchical clustering algorithm, calculate adaptive weighting coefficients based on the electrical similarity index, and perform weighted aggregation of the electrical impedance characteristics of each power generation unit within the power generation group to obtain equivalent electrical parameters, construct equivalent nodes based on the equivalent electrical parameters, and determine the simplified topology structure. The spectral domain partitioning module is used to determine the boundary constraints on the power grid side based on the access point information and apply them to the simplified topology, determine the Laplace matrix corresponding to the simplified topology and construct the spectral embedding space, map the equivalent nodes to the spectral embedding space to identify the node cluster boundaries, and divide the simplified topology into multiple sub-network regions to obtain the partitioned topology. The asynchronous simulation module is used to configure independent simulation step sizes for sub-network regions according to the partition topology and perform asynchronous parallel simulation. It extracts boundary node data at a preset synchronization time, and after time alignment processing of the boundary interaction data by combining the extrapolation prediction algorithm, it transmits the data to adjacent sub-network regions to obtain global simulation results.

[0014] A third aspect of the present invention provides an electronic device, comprising: A processor and a memory for storing processor-executable instructions, wherein the processor is configured to invoke instructions stored in the memory to perform the aforementioned method.

[0015] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0016] In this invention, by extracting the temporal feature matrix and dominant feature components of the operating state of power generation units and calculating the electrical similarity index, scientific grouping of power generation units is achieved, effectively reducing the computational workload of modeling complex new energy power plants. Adaptive weighting coefficients are used to aggregate the electrical impedance characteristics of each power generation unit within the power generation group, construct equivalent nodes, and determine a simplified topology, significantly improving simulation modeling efficiency and reducing computational complexity. Based on spectral embedding technology, equivalent nodes are mapped to the spectral embedding space, realizing optimized partitioning of the network topology and enhancing the adaptability of the simulation model. Independent simulation step sizes are configured for sub-network regions and asynchronous parallel simulation is executed. Combined with extrapolation prediction algorithms, time alignment processing of boundary interaction data is performed, significantly improving simulation speed while ensuring simulation accuracy. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the simulation method for grid-connected operation of a new energy power station according to an embodiment of the present invention. Figure 2 This is a flowchart of the asynchronous power network simulation method for the grid-connected operation of new energy power plants, as described in an embodiment of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of 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.

[0019] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0020] Figure 1 This is a flowchart illustrating the simulation method for grid-connected operation of a new energy power station according to an embodiment of the present invention, as shown below. Figure 1 As shown, the method includes: Obtain the operation configuration information of the new energy power station and the grid access point information. Based on the operation configuration information, extract the operation status features of the power generation unit and construct a time-series feature matrix. Extract the dominant feature components corresponding to the time-series feature matrix to construct a feature vector set and calculate the electrical similarity index between the power generation units. Based on the electrical similarity index, the power generation units are grouped into multiple power generation groups using an agglomerative hierarchical clustering algorithm. An adaptive weighting coefficient is calculated based on the electrical similarity index, and the electrical impedance characteristics of each power generation unit within the power generation group are weighted and aggregated to obtain equivalent electrical parameters. Based on the equivalent electrical parameters, equivalent nodes are constructed and a simplified topology is determined. Based on the access point information, the boundary constraints on the power grid side are determined and applied to the simplified topology. The Laplace matrix corresponding to the simplified topology is determined and a spectral embedding space is constructed. The equivalent nodes are mapped to the spectral embedding space to identify the node cluster boundaries and the simplified topology is divided into multiple sub-network regions to obtain a partitioned topology. Based on the partitioned topology, an independent simulation step size is configured for each sub-network region, and asynchronous parallel simulation is performed. Boundary node data is extracted at a preset synchronization time. After time alignment processing of the boundary interaction data using an extrapolation prediction algorithm, the data is transmitted to adjacent sub-network regions to obtain the global simulation results.

[0021] In one alternative implementation, The process involves acquiring the operation configuration information of the renewable energy power station and the grid connection point information. Based on the operation configuration information, the operation status features of the power generation units are extracted and a time-series feature matrix is ​​constructed. The dominant feature components corresponding to the time-series feature matrix are extracted to construct a feature vector set, and the electrical similarity index between the power generation units is calculated, including: The power output sequence and control response sequence of each power generation unit under multiple operating conditions are collected, and the timestamps are aligned and abnormal data points are removed to obtain the effective power sequence and effective control sequence. The time sequence feature matrix is ​​then constructed by splicing the sequences according to the time dimension. The time series feature matrix is ​​standardized to obtain a standardized time series matrix. A covariance matrix is ​​constructed based on the standardized time series matrix and eigenvalue decomposition is performed. The eigenvalues ​​are sorted in descending order according to their size and the eigenvectors corresponding to the top few eigenvalues ​​are selected as principal component directions based on the cumulative contribution rate. The standardized time series matrix is ​​then projected onto the principal component directions to obtain the dominant feature components. Extract the dominant feature components corresponding to each power generation unit and organize them according to the power generation unit number to construct the feature vector set. Perform covariance calculation on the feature vectors of any two power generation units in the feature vector set and construct a covariance matrix. Perform matrix inversion operation on the covariance matrix to obtain the inverse covariance matrix. Calculate the difference vector between the feature vectors of different power generation units and perform weighted distance calculation with the corresponding inverse covariance matrix to obtain the Mahalanobis distance. Perform normalization mapping on the Mahalanobis distance to obtain the electrical similarity index.

[0022] The power output and control response data of each power generation unit are collected and preprocessed, and the power output sequence and control response sequence of each power generation unit in the new energy power station are collected under multiple typical operating conditions.

[0023] The collected raw data undergoes timestamp alignment. Because the data acquisition equipment of each power generation unit may have clock deviations, the acquisition times are not entirely consistent. Therefore, all data needs to be resampled according to a unified time base. In practical applications, a sampling interval of 100 milliseconds can be selected for time alignment. For example, for five wind turbine generators in a wind farm, active power output and pitch control response data can be collected under five typical operating conditions: normal operation, low wind speed, high wind speed, sudden wind speed increase, and sudden wind speed decrease. The collection time for each operating condition is 10 minutes.

[0024] After time alignment, outlier data points need to be removed. Outlier data may originate from sensor malfunctions, communication interruptions, or recording errors. A statistical anomaly detection strategy is employed, calculating the mean and standard deviation of each time series. Data points deviating from the mean by more than three times the standard deviation are marked as outliers and removed. In the aforementioned wind farm example, the effective power series obtained after removing outliers may be reduced from the original 6000 data points per operating condition to approximately 5900 effective data points.

[0025] The processed effective power sequence and effective control sequence are concatenated along the time dimension to construct a time-series feature matrix. For the aforementioned wind farm example, assuming each wind turbine collects 5900 effective data points under each operating condition, the time-series feature matrix for each turbine has a dimension of 5900×5×2, where 5 represents the 5 operating conditions and 2 represents the power and control data. Combining the data from the 5 turbines yields a comprehensive time-series feature matrix with a dimension of 5900×5×2×5.

[0026] The constructed time-series feature matrix is ​​standardized to transform data of different dimensions to the same scale range. The standardization process uses a standard normal distribution mapping with a mean of 0 and a standard deviation of 1. The mean of each column is subtracted, and then the result is divided by the standard deviation of that column to obtain the standardized time-series matrix. In the previous example, the standardized power and control data are distributed between -3 and 3, with most data points concentrated between -1 and 1.

[0027] A covariance matrix is ​​constructed based on the standardized time series matrix, and the correlations between different features are calculated. Eigenvalue decomposition is performed on the covariance matrix, decomposing it into a combination of eigenvalues ​​and eigenvectors. The eigenvalues ​​are then sorted in descending order of magnitude, and the cumulative contribution rate is calculated. The eigenvectors corresponding to the top few eigenvalues ​​with a cumulative contribution rate reaching 85% are selected as principal component directions. In the wind farm example, the first 3 to 5 principal components can be selected, as these can explain more than 85% of the variability of the original data.

[0028] The standardized time series matrix is ​​projected onto the principal components to obtain the dominant feature components. This maps the high-dimensional original feature space to a low-dimensional feature space defined by the principal components. Each dominant feature component represents the projection value of the original data onto the corresponding principal component direction. For the aforementioned wind farm example, if the first four principal components are selected, the dimension of the dominant feature component for each wind turbine is 5900×4.

[0029] The dominant feature components corresponding to each power generation unit are extracted and organized according to the power generation unit number to construct a feature vector set. For the aforementioned wind farm example, five feature vectors can be obtained, each with a dimension of 5900×4, corresponding to the main operating characteristics of the five wind turbine generators.

[0030] Covariance calculation is performed on the eigenvectors of any two power generation units in the eigenvector set to construct a covariance matrix. The covariance matrix describes the correlation structure between the dominant features of different power generation units. The inverse covariance matrix is ​​then obtained by performing a matrix inversion operation on this covariance matrix. In the aforementioned wind farm example, the covariance matrix between the two wind turbines has a dimension of 4×4, corresponding to the covariance relationship of the four principal components.

[0031] The difference vector between the eigenvectors of different power generation units is calculated, and a weighted distance is calculated by comparing it with the corresponding inverse covariance matrix to obtain the Mahalanobis distance. The Mahalanobis distance considers the distribution characteristics and correlations of each dimension in the feature space, and can more accurately measure the differences in the operating characteristics of different power generation units. In the aforementioned wind farm example, a Mahalanobis distance value can be calculated between any two wind turbines; for example, the Mahalanobis distance between wind turbine 1 and wind turbine 2 is 8.76, and the Mahalanobis distance between wind turbine 1 and wind turbine 3 is 12.34.

[0032] The calculated Mahalanobis distance is normalized to obtain an electrical similarity index. The normalization mapping uses an exponential function transformation, resulting in a similarity index ranging from 0 to 1. The closer the value is to 1, the more similar the electrical characteristics of the two power generation units are. In the aforementioned wind farm example, the electrical similarity between wind turbine 1 and wind turbine 2 is 0.87, and the electrical similarity between wind turbine 1 and wind turbine 3 is 0.65.

[0033] In this embodiment, by aligning the power sequence and control sequence with timestamps and removing outliers, the impact of time drift and noise interference between different operating conditions and sampling sources on the analysis results is avoided. This ensures the consistency and reliability of the input data in the time dimension, improving the effectiveness of subsequent feature analysis. Based on a multidimensional time-series feature matrix and combined with standardization and covariance analysis, biases caused by differences in dimensions and inconsistent amplitude scales can be eliminated. Principal component analysis is used to extract dominant feature components, effectively compressing redundant information in high-dimensional time-series data. This highlights the main power change patterns and control response characteristics of the power generation unit under different operating conditions, significantly reducing noise sensitivity and computational complexity. In one alternative implementation, Based on the electrical similarity index, the power generation units are grouped into multiple power generation groups using an agglomerative hierarchical clustering algorithm. Adaptive weighting coefficients are calculated based on the electrical similarity index, and the electrical impedance characteristics of each power generation unit within a power generation group are weighted and aggregated to obtain equivalent electrical parameters. Based on these equivalent electrical parameters, equivalent nodes are constructed, and a simplified topology is determined, including: Based on the electrical similarity index, a similarity matrix is ​​constructed and the power generation units are treated as independent clusters. The power generation unit pairs with the highest similarity in the similarity matrix are calculated and cluster merging operations are performed to obtain merged clusters and update the similarity matrix. The clustering quality is calculated by combining the silhouette coefficient evaluation algorithm. The cluster merging operation is repeated until the clustering quality converges to obtain multiple power generation groups. Calculate the electrical similarity index between each power generation unit in the power generation group and the center of the power generation group, and perform reverse mapping to obtain adaptive weighting coefficients. Extract the resistance parameters and reactance parameters of each power generation unit in the power generation group from the operation configuration information and construct an electrical impedance feature vector. An adjacency matrix is ​​constructed based on the electrical connection relationship between each power generation unit within the power generation group. A symmetric transformation is performed on the adjacency matrix using a graph convolution propagation algorithm to obtain a symmetric adjacency matrix. Multi-hop propagation is then performed on the electrical impedance feature vector to obtain multi-order neighborhood features. Layer-by-layer aggregation is performed on the multi-order neighborhood features to obtain neighborhood aggregation features. Weighted summation and nonlinear activation transformation operations are then performed on the adaptive weighting coefficients to obtain equivalent electrical parameters. The equivalent node is constructed for each power generation group based on the equivalent electrical parameters. The connection topology between the power generation unit and the combiner unit in the operation configuration information is extracted. The simplified topology is constructed based on the equivalent node and the connection topology.

[0034] Based on the aforementioned calculated electrical similarity index, a similarity matrix is ​​constructed for grouping power generation units. This similarity matrix is ​​symmetric, with each element representing the electrical similarity between two corresponding power generation units. For a large photovoltaic power plant containing 30 photovoltaic power generation units, the initially constructed similarity matrix has a dimension of 30×30, with all diagonal elements being 1, indicating that each power generation unit is completely similar to itself. Off-diagonal elements have values ​​between 0 and 1; larger values ​​indicate greater similarity in the electrical characteristics of the two power generation units.

[0035] After constructing the similarity matrix, each power generation unit is treated as an independent cluster and grouped using a hierarchical clustering method. The power generation unit pair with the highest similarity in the similarity matrix is ​​calculated; for example, the similarity between power generation unit 7 and power generation unit 12 is 0.96, which is the maximum value among all off-diagonal elements. A cluster merging operation is performed, merging power generation unit 7 and power generation unit 12 into a new cluster. This new cluster contains the characteristic information of both power generation units. After merging the clusters, the similarity matrix is ​​updated using an average linking strategy, meaning the similarity between the new cluster and other power generation units is the average of the similarities between the two original clusters and that power generation unit. Specifically, the similarity between the new cluster and power generation unit 3 is equal to the average of the similarity values ​​of power generation unit 7 and power generation unit 3 (0.85) and power generation unit 12 and power generation unit 3 (0.81), resulting in 0.83. This method updates the entire similarity matrix, reducing its dimension to 29×29.

[0036] Each time a cluster merge operation is performed, the clustering quality of the current clustering result is calculated. The clustering quality is evaluated using the silhouette coefficient algorithm. The calculation process is as follows: for each power generation unit, the average similarity value between the current power generation unit and other power generation units in its cluster is calculated, denoted as cohesion; the average similarity value between the current power generation unit and all power generation units not in its current cluster is calculated, denoted as separation; the silhouette coefficient is equal to the difference between separation and cohesion divided by the maximum of the two. The average silhouette coefficient of all power generation units is taken as the overall clustering quality. In the example above, the initial clustering quality was 0.57, and after the first cluster merge, the clustering quality improved to 0.59, indicating an improvement in the clustering effect.

[0037] The cluster merging operation is repeated, and the clustering quality is calculated until the clustering quality no longer improves significantly. Specifically, the criterion is that the clustering quality improvement after three consecutive cluster merging operations is less than 0.01, indicating that the clustering results have converged. In the example above, after 11 cluster merging operations, the clustering quality reached 0.76 and stabilized, ultimately forming five power generation groups, each containing a different number of power generation units. Group 1 contains 8 units, Group 2 contains 6 units, Group 3 contains 7 units, Group 4 contains 5 units, and Group 5 contains 4 units. These groups represent sets of power generation units with similar electrical characteristics.

[0038] For each power generation group, the electrical similarity index between each power generation unit and the group center is calculated. The group center is defined as the average of the dominant characteristic components of all power generation units within that group. In the previous example, the electrical similarities between the eight power generation units in group 1 and the group center are 0.94, 0.92, 0.89, 0.87, 0.86, 0.85, 0.84, and 0.82, respectively. These similarity indices are then reverse-mapped to obtain adaptive weighted coefficients. The mapping method involves dividing the similarity value by the sum of all similarity values ​​within the group. Specifically, the sum of the similarities of the eight units in group 1 is 6.99. The adaptive weighted coefficient for the first unit is 0.94 divided by 6.99, which equals 0.141. The weighted coefficients for the remaining units are calculated similarly to 0.138, 0.134, 0.131, 0.129, 0.128, 0.126, and 0.123, respectively. The sum of these weighted coefficients is 1, which will be used for subsequent parameter equivalence calculations.

[0039] The resistance and reactance parameters of each generating unit within each generating group are extracted from the operational configuration information to construct an electrical impedance feature vector. This feature vector contains the resistance, reactance, and rated capacity information of the generating unit. In the aforementioned example, the resistances of the eight generating units in group 1 are 0.040, 0.042, 0.038, 0.041, 0.037, 0.043, 0.039, and 0.044 ohms, respectively; the reactances are 0.120, 0.125, 0.118, 0.123, 0.116, 0.127, 0.119, and 0.129 ohms, respectively; and the rated capacity is 2500 kilowatts for each unit. These parameters constitute the set of electrical impedance feature vectors for this group.

[0040] An adjacency matrix is ​​constructed based on the electrical connections between the power generation units within a power generation group. The construction method is as follows: the number of rows and columns of the matrix is ​​the same as the number of power generation units within the group. Elements in the matrix indicate whether there is a direct connection between two power generation units; directly connected units have a matrix element value of 1, otherwise, it is 0. In the example above, the adjacency matrix of group 1 has a dimension of 8×8. The matrix elements are filled according to the actual connection relationships between the 8 power generation units. Assuming unit 1 is connected to units 2 and 3, then the first row and second column, the first row and third column, and the corresponding second row and first column, and third row and first column of the matrix are all 1, and so on.

[0041] A symmetric transformation is performed on the adjacency matrix using the graph convolutional propagation algorithm to ensure the stability of the feature propagation process. The symmetric transformation includes: adding self-joins, i.e., setting all diagonal elements of the adjacency matrix to 1; and normalizing the adjacency matrix so that the sum of elements in each row is 1. In the previous example, after adding self-joins, the original adjacency matrix of group 1 has all diagonal elements becoming 1, indicating that each power generation unit considers both the influence of its neighbors and its own characteristics. Subsequently, each row of the matrix is ​​normalized; for example, the first row has three elements with a value of 1, which becomes 0.333 after normalization.

[0042] Multi-hop propagation is performed by combining the symmetric adjacency matrix and the electrical impedance eigenvector to capture multi-order neighborhood information. Multi-hop propagation is achieved by multiplying the symmetric adjacency matrix with the eigenvector; each matrix multiplication considers the influence of first-order neighbors. In the previous example, two propagation operations are performed on group 1: multiplying the symmetric adjacency matrix with the original electrical impedance eigenvector yields first-order neighborhood features; multiplying the symmetric adjacency matrix with the first-order neighborhood features yields second-order neighborhood features.

[0043] Layer-by-layer aggregation is performed on multi-order neighborhood features, combining neighborhood features of different orders with weights to obtain aggregated neighborhood features. The aggregation method assigns different weight coefficients to zero-order, first-order, and second-order features, and then sums them with weights. In the previous example, the zero-order feature (the original feature) is assigned a weight of 0.5; the first-order feature is assigned a weight of 0.3; and the second-order feature is assigned a weight of 0.2. For each power generation unit in group 1, its zero-order, first-order, and second-order features are summed with the above weights to obtain the aggregated neighborhood features. It should be noted that the weight settings can be determined according to task requirements or actual conditions.

[0044] Combining the aforementioned adaptive weighting coefficients, a weighted summation operation is performed on the neighborhood aggregation features to achieve weighted fusion of the characteristics of each power generation unit within the power generation group. The neighborhood aggregation features of each power generation unit within the group are multiplied by the corresponding adaptive weighting coefficient and then summed. In the aforementioned example, the neighborhood aggregation features of the eight power generation units in group 1 are multiplied by their corresponding weighting coefficients of 0.141, 0.138, 0.134, 0.131, 0.129, 0.128, 0.126, and 0.123, respectively, and then summed to obtain the preliminary fusion result.

[0045] A nonlinear activation transformation was applied to the preliminary fusion results to introduce nonlinear characteristics and improve the model's expressive power. The ReLU function was used as the activation function, taking the maximum value of 0 or the input value itself. After the activation transformation, the equivalent electrical parameters of group 1 were obtained: equivalent resistance of 0.035 ohms, equivalent reactance of 0.108 ohms, and equivalent capacity of 19600 kW, that is, the sum of the capacities of the 8 units is 8 × 2500 kW.

[0046] Based on the calculated equivalent electrical parameters, an equivalent node is constructed for each power generation group, simplifying the originally dispersed multiple power generation units into a single node with equivalent characteristics. In the aforementioned example, five equivalent nodes are constructed for each of the five power generation groups. Each equivalent node has the calculated equivalent resistance, equivalent reactance, and equivalent capacity characteristics. The equivalent node will replace the 30 power generation units in the original topology in subsequent simulation calculations.

[0047] The connection topology between generation units and combiner units is extracted from the operational configuration information to determine how each equivalent node connects to other parts of the power grid. In the previous example, 30 generation units are ultimately connected to 3 combiner units through a multi-level connection method. In simplification, it is necessary to determine the connection relationships between 5 equivalent nodes and 3 combiner units. The determination of these connection relationships is based on the connection status of each generation unit and combiner unit in the original topology. If a generation unit in a generation group is connected to a combiner unit, then the corresponding equivalent node is also connected to that combiner unit.

[0048] In this embodiment, by constructing a similarity matrix based on electrical similarity indices and adopting a stepwise clustering method, the grouping process of power generation units is directly constrained by the similarity of their actual electrical behavior. This can more accurately reflect the dynamic consistency of power generation units during operation. The silhouette coefficient is used to evaluate the clustering quality and determine the number of groups under convergence conditions, avoiding the subjectivity caused by manually setting the number of clusters. This improves the stability and interpretability of the grouping results. By back-mapping the electrical similarity between power generation units and group centers to adaptive weighting coefficients, the calculation of equivalent parameters can highlight power generation units with strong representativeness and high typicality, reducing the interference of edge units or abnormal operating conditions on the equivalent results, thereby improving the accuracy of equivalent modeling. Through the layer-by-layer aggregation and nonlinear mapping of multi-order neighborhood features, the effective expression of complex electrical relationships is achieved, which can better preserve the electrical influence paths and interaction characteristics within the groups.

[0049] In one alternative implementation, The adjacency matrix is ​​subjected to a symmetric transformation using a graph convolution propagation algorithm to obtain a symmetric adjacency matrix. This symmetric matrix is ​​then combined with the electrical impedance feature vector to perform a multi-hop propagation operation, resulting in multi-order neighborhood features, including: The connection relationship information of each power generation unit node in the adjacency matrix is ​​extracted and the node degree is counted to construct a degree matrix. The degree matrix is ​​subjected to an inverse transformation to obtain a degree inverse matrix, and the adjacency matrix is ​​subjected to a two-sided symmetric transformation to obtain a symmetric adjacency matrix. A topology-aware mask matrix is ​​constructed based on the symmetric adjacency matrix. The electrical impedance feature vector is used as a zero-order feature input. The capacity parameters and impedance parameters of each power generation unit are extracted from the operation configuration information and a node attribute embedding vector is constructed. The neighborhood propagation feature is obtained by performing neighborhood information aggregation on the current layer feature vector and the symmetric adjacency matrix using a graph convolution algorithm. The neighborhood propagation feature is then fused with the node attribute embedding vector using a feature fusion algorithm to obtain an enhanced propagation feature. The enhanced propagation feature is then subjected to a learnable weight transformation and combined with the topology-aware mask matrix to perform selective feature filtering to obtain the current layer output feature. The residual enhanced feature is then obtained by solving based on the current layer output feature and the current layer input feature. Determine whether the current propagation layer has reached the preset maximum propagation layer. If not, use the residual enhancement feature as the input feature of the next layer and repeat the neighborhood propagation operation. If it has reached the maximum, extract the residual enhancement features output by each propagation layer and calculate the hierarchical importance weight based on the propagation depth. Aggregate the residual enhancement features of each propagation layer with the corresponding hierarchical importance weight to obtain multi-level neighborhood features.

[0050] The connection information of each power generation unit node is extracted based on the constructed adjacency matrix, and the degree of each node is counted to construct a degree matrix. The degree refers to the number of other nodes directly connected to a given node, representing the connection density of a node in the topology network. In the example above, group 1 contains 8 power generation unit nodes, and the constructed degree matrix is ​​a diagonal matrix with diagonal elements of 3, 2, 3, 4, 2, 3, 2, 1, indicating that the first power generation unit is connected to 3 other units, the second power generation unit is connected to 2 other units, and so on.

[0051] Performing an inverse transformation on the degree matrix yields the degree inverse matrix. The inverse transformation is calculated by taking the reciprocal of each non-zero diagonal element of the degree matrix. In the previous example, the diagonal elements of the degree inverse matrix are 0.333, 0.500, 0.333, 0.250, 0.500, 0.333, 0.500, and 1.000. The degree inverse matrix plays a crucial role in the subsequent adjacency matrix normalization, balancing the influence of nodes with different connectivity degrees.

[0052] A two-sided symmetric transformation is performed on the adjacency matrix to obtain a symmetric adjacency matrix. This transformation involves three steps: adding self-joins, left multiplication of the degree inverse matrix, and right multiplication of the degree inverse matrix. Adding self-joins involves adding 1 to the diagonal elements of the original adjacency matrix, indicating that nodes are connected to themselves. Left and right multiplication of the degree inverse matrix involves multiplying the degree inverse matrix from the left and right sides respectively with the adjacency matrix after adding self-joins, thus normalizing the node connection information. In the previous example, performing this transformation on the adjacency matrix of group 1 yields an 8×8 symmetric adjacency matrix. Each element in this matrix has a value between 0 and 1 and exhibits symmetry, ensuring that feature propagation in the graph structure is bidirectionally balanced.

[0053] A topology-aware mask matrix is ​​constructed based on a symmetric adjacency matrix. This mask matrix controls the information flow path during feature propagation. The mask matrix is ​​constructed by setting a threshold for each element in the symmetric adjacency matrix: when the element value is greater than 0.1, the mask value at the corresponding position is 1; otherwise, it is 0. In the previous example, the resulting mask matrix has a dimension of 8×8, the same as the symmetric adjacency matrix, but the elements only have two values: 0 and 1, indicating whether the connection relationship between corresponding nodes is considered during feature propagation.

[0054] The electrical impedance eigenvector is used as the zero-order feature input. The zero-order feature represents the initial characteristics of the node and has not been propagated through any graph structure. In the example above, the eigenvector composed of the resistance, reactance, and rated capacity parameters of the 8 generating units in group 1 is used as the zero-order feature, with a dimension of 8×3, and each row represents the feature of one generating unit.

[0055] Capacity and impedance parameters of each power generation unit are extracted from the operational configuration information to construct a node attribute embedding vector. This vector contains richer feature information about the power generation units, such as active power, reactive power, voltage amplitude, and voltage phase angle. In the aforementioned example, the operational parameters of the eight power generation units in group 1 are extracted, including active power of 2000, 1950, 2100, 2050, 1900, 2150, 2000, and 1850 kW, reactive power of 300, 290, 310, 305, 285, 320, 300, and 275 kW, voltage amplitude of 35 kV for all units, and phase angles of 0.02, 0.01, 0.03, 0.02, 0.01, 0.03, 0.02, and 0.01 radians. The power generation unit feature information, together with the electrical impedance features, constitutes a more complete node attribute embedding vector with a dimension of 8×7.

[0056] The graph convolution algorithm aggregates neighborhood information by combining the current layer's feature vector with a symmetric adjacency matrix, resulting in neighborhood propagation features. Graph convolution calculates by multiplying the symmetric adjacency matrix with the feature vector, enabling information transfer and aggregation of features between connected nodes. In the previous example, in the first propagation layer, an 8×8 dimensional symmetric adjacency matrix is ​​multiplied with an 8×3 dimensional zero-order feature vector to obtain an 8×3 dimensional neighborhood propagation feature, representing the new representation of each node after incorporating features from its neighbors.

[0057] The neighborhood propagation features and node attribute embedding vectors are fused using a feature fusion algorithm to obtain enhanced propagation features. Feature fusion employs an attention mechanism, dynamically assigning fusion weights based on the importance of different features. Specifically, it calculates the correlation score between the neighborhood propagation features and the node attribute embedding vectors, and then performs weighted fusion based on the score. In the aforementioned example, the 8×3 dimensional neighborhood propagation features are fused with the 8×7 dimensional node attribute embedding vectors to obtain 8×10 dimensional enhanced propagation features, which contain comprehensive information on electrical parameters and operating status.

[0058] A learnable weight transformation is performed on the enhancement propagation features, and selective feature filtering is then performed in conjunction with a topology-aware mask matrix to obtain the output features of the current layer. The learnable weight transformation applies a linear transformation to the enhancement propagation features, changing the feature's dimension or representation; selective feature filtering controls which connections participate in feature propagation through the mask matrix. For example, applying the learnable weight transformation to an 8×10 dimensional enhancement propagation feature transforms the feature dimension to 8×5, and combining this with the mask matrix retains only the features corresponding to valid connections, resulting in the output features of the current layer.

[0059] The residual enhancement features are calculated based on the output features and input features of the current layer. The residual connection is implemented by directly adding the input features to the output features to form the residual enhancement features. In the example above, the input features and output features of the first propagation layer are added to obtain the residual enhancement features, which have the same dimension as the output features, 8×5.

[0060] The system determines whether the current propagation layer has reached the preset maximum propagation layer number, which is set to 3 in the previous example. Since the first propagation layer has just finished executing and the maximum number of layers has not been reached, the residual enhancement feature is used as the input feature for the second propagation layer, and the neighborhood propagation operation is repeated. In the second propagation layer, the symmetric adjacency matrix is ​​multiplied with the first layer residual enhancement feature, and feature fusion, learnable weight transformation, and selective feature filtering are performed to calculate the second layer residual enhancement feature. Similarly, the above operations are repeated in the third propagation layer to obtain the third layer residual enhancement feature.

[0061] When the number of propagation layers reaches the preset maximum of 3, the residual enhancement features of each propagation layer's output are extracted, and the layer importance weights are calculated based on the propagation depth. The layer importance weights reflect the contribution of features from different propagation layers and generally decrease with increasing propagation depth. In the aforementioned example, the layer importance weights for the three propagation layers are set to 0.5, 0.3, and 0.2 respectively, with a sum of 1, indicating that shallower features account for a higher proportion.

[0062] The residual enhancement features of each propagation layer are aggregated with their corresponding layer importance weights to obtain multi-level neighborhood features. The aggregation method involves multiplying the residual enhancement features of each layer by their corresponding layer importance weights and then summing the results. In the example above, the residual enhancement features of the first layer are multiplied by 0.5, the second layer by 0.3, and the third layer by 0.2, and then summed to obtain the final multi-level neighborhood features, which have a dimension of 8×5 and contain node connection information at different scales from local to global.

[0063] In this embodiment, by normalizing the degree of adjacency relationships and constructing a symmetric adjacency matrix, the neighborhood information maintains numerical stability and scale consistency during propagation, avoiding excessive amplification of feature propagation by highly connected nodes, thereby improving the robustness of feature learning. Electrical impedance is used as the basic input feature, and node attributes such as the capacity and impedance of the power generation unit are embedded, so that the propagated features simultaneously include network topology information and the physical attributes of the nodes themselves. This overcomes the problem of incomplete information caused by relying solely on topology or attribute modeling. A residual enhancement mechanism is introduced, so that each propagation layer retains the original features while superimposing neighborhood information, effectively alleviating the problems of feature oversmoothing and information attenuation that may occur during multi-layer propagation. By performing hierarchical importance weighted aggregation of the output features of different propagation layers, adaptive integration of multi-level neighborhood information is achieved, significantly improving the completeness, discriminability, and stability of electrical feature expression, and providing a more accurate and reliable multi-level neighborhood feature foundation for subsequent equivalent parameter calculation and topology simplification.

[0064] In one alternative implementation, Based on the access point information, boundary constraints on the power grid side are determined and applied to the simplified topology. The Laplace matrix corresponding to the simplified topology is determined, and a spectral embedding space is constructed. The equivalent nodes are mapped to the spectral embedding space to identify node cluster boundaries. The simplified topology is then divided into multiple sub-network regions to obtain a partitioned topology, including: Extract the access point location identifier and grid-side access capacity limitation information from the access point information; construct a power constraint set based on the grid-side access capacity limitation information and bind it to the equivalent node in the simplified topology corresponding to the access point location identifier; construct a boundary node label matrix based on the equivalent node that has been bound to the power constraint set. Extract the connection relationships between equivalent nodes in the simplified topology and construct a connection matrix. Calculate the number of connections for each equivalent node based on the connection matrix, construct a node degree vector, and convert it into a diagonal matrix to obtain a degree diagonal matrix. Perform a matrix difference operation between the degree diagonal matrix and the connection matrix to obtain the Laplace matrix. The eigenvalue decomposition of the Laplacian matrix is ​​performed to obtain the eigenvector matrix, and the spectral embedding space is constructed based on the eigenvector matrix; The equivalent nodes in the simplified topology are projected onto the spectral embedding space to obtain spectral coordinate representations. A spectral similarity matrix is ​​constructed using a spectral clustering algorithm. The spectral similarity matrix is ​​constrained and adjusted based on the boundary node label matrix, and the boundary positions between dense and sparse spectral coordinate regions are identified. The node cluster boundaries are determined based on the boundary positions, and the equivalent nodes in the simplified topology are divided into different sub-network regions. A sub-network region identifier mapping table is constructed, and the simplified topology is decomposed based on the sub-network region identifier mapping table to obtain a partitioned topology.

[0065] The access point location identifier and grid-side access capacity limitation information are extracted from the access point information. The access point location identifier includes the grid-side connection node number and geographical location information, while the grid-side access capacity limitation information includes the active power limit, reactive power range, and allowable voltage fluctuation range. In the aforementioned example, the renewable energy power station has two access points: Access point 1 is located at substation bus node 7, with an active power limit of 80 MW, a reactive power range of -25 to 25 MW, and a voltage fluctuation range of ±5% of the rated value; Access point 2 is located at substation bus node 12, with an active power limit of 60 MW, a reactive power range of -20 to 20 MW, and a voltage fluctuation range of ±5% of the rated value.

[0066] A power constraint set is constructed based on grid-side access capacity limitation information and bound to the equivalent node in the simplified topology corresponding to the access point location identifier. The power constraint set includes active power upper limit constraints, reactive power range constraints, and voltage constraints. In the aforementioned example, the power constraint set of access point 1 is bound to the equivalent node connected to substation bus node 7 in the simplified topology, and the power constraint set of access point 2 is bound to the equivalent node connected to substation bus node 12.

[0067] A boundary node labeling matrix is ​​constructed based on the equivalent nodes with bound power constraint sets. The boundary node labeling matrix is ​​a vector with the same number of equivalent nodes, where nodes bound to the power constraint set are labeled as 1, and other nodes are labeled as 0. In the previous example, the simplified topology contains 5 equivalent nodes. Assuming that equivalent node 1 and equivalent node 3 are connected to access point 1 and access point 2 respectively, the boundary node labeling matrix is ​​1, 0, 1, 0, 0, indicating that equivalent node 1 and equivalent node 3 are boundary nodes and are subject to power constraints on the grid side.

[0068] Extract the connections between equivalent nodes in the simplified topology and construct a connection matrix. The connection matrix is ​​a square matrix with dimensions equal to the number of equivalent nodes. Matrix elements indicate whether nodes are directly connected. In the previous example, the connections between 5 equivalent nodes form a 5×5 connection matrix. If equivalent node 1 is connected to equivalent node 2, the corresponding element in the matrix has a value of 1; otherwise, it has a value of 0. The diagonal elements of the connection matrix are set to 0, indicating that a node is not connected to itself.

[0069] The connection count of each equivalent node is calculated based on the connection matrix. A node degree vector is constructed and converted into a diagonal matrix, resulting in a degree diagonal matrix. Node degree refers to the number of other nodes directly connected to that node, calculated by summing the elements of each row of the connection matrix. In the previous example, the connection counts of equivalent nodes 1, 2, 3, 4, and 5 are 2, 3, 2, 2, and 1 respectively, forming a node degree vector. This vector is then converted into a diagonal matrix, resulting in a 5×5 degree diagonal matrix with diagonal elements of 2, 3, 2, 2, and 1, and other elements of 0.

[0070] The Laplacian matrix is ​​obtained by performing a matrix difference operation between the degree diagonal matrix and the connectivity matrix. The matrix difference operation subtracts the connectivity matrix from the degree diagonal matrix. In the previous example, the Laplacian matrix has a dimension of 5×5, with diagonal elements equal to the degree of the corresponding node, and off-diagonal elements having a value of -1 if two nodes are connected, and 0 otherwise. The Laplacian matrix characterizes the global properties of the topology, containing information about the connections between nodes and the network's connectivity.

[0071] Eigenvalue decomposition is performed on the Laplacian matrix to obtain an eigenvector matrix, and a spectral embedding space is constructed based on this matrix. Eigenvalue decomposition breaks down the Laplacian matrix into eigenvalues ​​and their corresponding eigenvectors; the eigenvector matrix contains all eigenvectors. In the previous example, the 5×5 Laplacian matrix was decomposed into 5 eigenvalues ​​and their corresponding eigenvectors. After sorting the eigenvalues ​​in ascending order, the second and third eigenvectors were selected to construct the spectral embedding space. These two eigenvectors were chosen because the first eigenvector corresponds to an eigenvalue of 0, containing less information, while subsequent eigenvectors correspond to larger eigenvalues ​​and typically contain noise.

[0072] The equivalent nodes in the simplified topology are projected into the spectral embedding space to obtain spectral coordinate representations, and a spectral similarity matrix is ​​constructed using a spectral clustering algorithm. Spectral coordinates represent the coordinates of nodes in the spectral embedding space, calculated by associating node numbers with the element values ​​of the corresponding rows in the feature vector matrix. In the previous example, the coordinates of the five equivalent nodes in the spectral embedding space are (0.35, 0.15), (0.40, 0.20), (-0.30, 0.25), (-0.35, -0.20), and (-0.10, -0.40), respectively. The Euclidean distance between nodes is calculated based on these coordinates, and the spectral similarity matrix is ​​constructed. A Gaussian kernel function is used for spectral similarity calculation; the closer the nodes are, the higher the similarity.

[0073] The spectral similarity matrix is ​​constrained and adjusted based on the boundary node label matrix, and the boundary positions between dense and sparse spectral coordinate regions are identified. Constraint adjustment refers to reducing the weight of the similarity values ​​of boundary nodes, making them more likely to serve as separators between different regions. In the aforementioned example, the similarity values ​​of equivalent node 1 and equivalent node 3 are multiplied by a coefficient of 0.8 for weight reduction. Dense spectral coordinate regions refer to areas where nodes are clustered in the spectral space, while sparse regions are areas where nodes are sparsely distributed. The boundary positions are generally located where there are significant intervals in the coordinate distribution.

[0074] The node cluster boundary is determined based on the dividing location, and the equivalent nodes in the simplified topology are divided into different sub-network regions. The node cluster boundary is determined based on the distribution of nodes in the spectral space, usually selecting locations with significant changes in node density. Spectral clustering is used to divide the 5 equivalent nodes into 3 sub-network regions: Region 1 contains equivalent node 1 and equivalent node 2, Region 2 contains equivalent node 3, and Region 3 contains equivalent node 4 and equivalent node 5. Note that the boundary nodes, equivalent node 1 and equivalent node 3, belong to different regions, which meets the requirements for grid connection point zoning.

[0075] A sub-network region identifier mapping table is constructed, and the simplified topology is decomposed based on this table to obtain a partitioned topology. The sub-network region identifier mapping table records the region identifier to which each equivalent node belongs. In the previous example, the mapping table content is: equivalent node 1 corresponds to region 1, equivalent node 2 corresponds to region 1, equivalent node 3 corresponds to region 2, equivalent node 4 corresponds to region 3, and equivalent node 5 corresponds to region 3. Based on this mapping table, the simplified topology is decomposed into three sub-network regions while preserving the connectivity between regions.

[0076] In this embodiment, by directly binding grid-side access capacity limitation information to the corresponding equivalent nodes and constructing boundary node markers, the topology partitioning process explicitly perceives the power constraint characteristics of the access points. This avoids the shortcomings of partitioning based solely on structural connectivity while ignoring operational constraints, thus improving the feasibility and security of the partitioning results in engineering applications. Based on the connection relationships between equivalent nodes, a Laplace matrix is ​​constructed and spectral embedding is performed, enabling the global topological relationships between nodes to be continuously expressed in a low-dimensional space, which is more conducive to identifying potential weakly coupled boundaries. By identifying the boundary positions between dense and sparse spectral coordinate regions, adaptive determination of sub-network boundaries is achieved, avoiding the subjectivity and instability caused by manually setting the number of partitions or boundary positions. While maintaining the original simplified overall topological connectivity, the partitioned topology structure achieves weak coupling and clear boundaries between sub-networks. It significantly improves the rationality of partitioning, constraint consistency, and the computational efficiency of subsequent parallel analysis and partition optimization, providing a more reliable topological foundation for partitioned scheduling and coordinated control of large-scale power generation systems.

[0077] In one alternative implementation, Based on the partitioned topology, an independent simulation step size is configured for each sub-network region, and asynchronous parallel simulation is performed. Boundary node data is extracted at a preset synchronization time. After time alignment processing of the boundary interaction data using an extrapolation prediction algorithm, the data is transmitted to adjacent sub-network regions to obtain global simulation results, including: The topology complexity parameters and node quantity parameters of each sub-network region are extracted from the partition topology structure. The computational load of each sub-network region is calculated based on the topology complexity parameters and the node quantity parameters. Based on the computational load and the coupling strength between different sub-network regions, an independent simulation step size is allocated to each sub-network region through a load balancing optimization algorithm, and a step size configuration table is constructed based on the independent simulation step size. Identify the boundary connection relationships between each sub-network region from the partition topology and extract the boundary node identifiers, and construct a boundary node list based on the boundary node identifiers; According to the step size configuration table, each sub-network region is driven to perform asynchronous parallel simulation according to the corresponding independent simulation step size. When the simulation time of each sub-network region reaches the preset synchronization time, the corresponding boundary node data is extracted from the boundary node list as boundary interaction data and the corresponding actual extraction time is recorded. Based on the time deviation between the actual extraction time and the preset synchronization time, the boundary interaction data is time-aligned by an extrapolation prediction algorithm to obtain synchronized boundary data. According to the boundary connection relationship, the synchronized boundary data is passed to the adjacent sub-network region as boundary conditions. Asynchronous parallel simulation is repeated until the simulation ends. The simulation output data of each sub-network region is summarized to obtain the global simulation result.

[0078] The topology complexity parameters and node count parameters of each sub-network region are extracted from the partitioned topology. Topology complexity parameters include metrics such as network density, average path length, and connectivity distribution, used to measure the complexity of connections within a sub-network region. In the example above, the partitioned topology contains three sub-network regions: Region 1 has a topology complexity parameter of 0.8 and 2 nodes; Region 2 has a topology complexity parameter of 0.3 and 1 node; and Region 3 has a topology complexity parameter of 0.5 and 2 nodes. Higher topology complexity parameters indicate more complex network connections and greater simulation computation difficulty.

[0079] The computational load of each sub-network region is calculated based on the topology complexity parameter and the number of nodes. The computational load assessment considers the product of the square of the number of nodes and the topology complexity parameter, reflecting the computational resources required to solve the power flow equations. In the previous example, the computational load of region 1 is 2 × 2 × 0.8 = 3.2; the computational load of region 2 is 1 × 1 × 0.3 = 0.3; and the computational load of region 3 is 2 × 2 × 0.5 = 2.0. The sum of the computational loads of all regions is 5.5, which serves as the baseline value for load allocation.

[0080] Based on the computational load and the coupling strength between different sub-network regions, an independent simulation step size is assigned to each sub-network region using a load balancing optimization algorithm. The coupling strength is determined by the number of connections between regions and the amount of interactive data. Higher coupling strength requires smaller differences in simulation step sizes between adjacent regions to reduce interaction errors. In the example above, there is one connection between region 1 and region 2 with a coupling strength of 0.4; one connection between region 1 and region 3 with a coupling strength of 0.3; and one connection between region 2 and region 3 with a coupling strength of 0.5. The load balancing optimization algorithm allocates simulation step sizes based on the inverse relationship between computational load and coupling strength constraints, ensuring full utilization of computational resources while maintaining time synchronization accuracy between interactive regions.

[0081] A step size configuration table is constructed based on independent simulation step sizes. This table records the simulation step size and corresponding synchronization period for each sub-network region, guiding the asynchronous parallel simulation process. In the aforementioned example, the step size configuration table contains the following: simulation step size for region 1 is 5 milliseconds; simulation step size for region 2 is 20 milliseconds; and simulation step size for region 3 is 8 milliseconds. The synchronization period is set to 40 milliseconds, meaning boundary data exchange and synchronization occur every 40 milliseconds. The allocation of simulation step sizes follows the principle of smaller step sizes for regions with heavier computational loads and larger step sizes for regions with lighter computational loads, to achieve a balance in overall computation time.

[0082] The boundary connections between sub-network regions are identified from the partitioned topology, and boundary node identifiers are extracted. Boundary connections describe the physical connections between different sub-network regions, and boundary nodes are the nodes connecting different regions. In the previous example, the boundary connection between region 1 and region 2 is as follows: equivalent node 2 of region 1 is connected to equivalent node 3 of region 2; the boundary connection between region 2 and region 3 is as follows: equivalent node 3 of region 2 is connected to equivalent node 4 of region 3; the boundary connection between region 1 and region 3 is as follows: equivalent node 1 of region 1 is connected to equivalent node 5 of region 3. The boundary node identifiers are equivalent node 1, equivalent node 2, equivalent node 3, equivalent node 4, and equivalent node 5, respectively.

[0083] A boundary node list is constructed based on the boundary node identifiers. The boundary node list contains information such as all boundary nodes, their respective regions, connected regions, and interaction data types. In the example above, the boundary node list contains: equivalent node 1 belongs to region 1, connects to region 3, and its interaction data type is voltage and power; equivalent node 2 belongs to region 1, connects to region 2, and its interaction data type is voltage and power; equivalent node 3 belongs to region 2, connects to regions 1 and 3, and its interaction data type is voltage and power; equivalent node 4 belongs to region 3, connects to region 2, and its interaction data type is voltage and power; equivalent node 5 belongs to region 3, connects to region 1, and its interaction data type is voltage and power.

[0084] The simulation is driven by a step size configuration table, which drives each sub-network region to perform asynchronous parallel simulations according to its corresponding independent simulation step size. Asynchronous parallel simulation allows different sub-network regions to advance independently with different time steps, reducing waiting time and improving computational efficiency. In the example above, region 1 updates its state every 5 milliseconds, region 2 updates its state every 20 milliseconds, and region 3 updates its state every 8 milliseconds. The three regions perform simulation calculations in parallel, each advancing the simulation time at a different rate.

[0085] When the simulation time of each sub-network region reaches the preset synchronization time, the corresponding boundary node data is extracted from the boundary node list as boundary interaction data, and the corresponding actual extraction time is recorded. The preset synchronization time is the specified time point for data interaction between regions, such as 40 milliseconds or 80 milliseconds in the previous example. Due to the different step sizes of each region, the actual simulation time may not accurately reach the synchronization time, so it is necessary to record the actual data extraction time for subsequent time alignment processing. In the previous example, at the first synchronization time of 40 milliseconds, the actual extraction time of region 1 is 40 milliseconds, the actual extraction time of region 2 is 40 milliseconds, and the actual extraction time of region 3 is 40 milliseconds, and the time of the three regions is exactly synchronized; at the second synchronization time of 80 milliseconds, the actual extraction time of region 1 is 80 milliseconds, the actual extraction time of region 2 is 80 milliseconds, and the actual extraction time of region 3 is 80 milliseconds, and the time of the three regions is synchronized again.

[0086] Based on the time deviation between the actual extraction time and the preset synchronization time, an extrapolation prediction algorithm is used to perform time alignment processing on the boundary interaction data to obtain synchronized boundary data. The extrapolation prediction algorithm predicts the state value of the boundary nodes at the synchronization time based on the changing trends of historical data. In the aforementioned example, at the third synchronization time of 120 milliseconds, the actual extraction time of region 1, region 2, and region 3 is 120 milliseconds, and the times of the three regions are completely synchronized again, requiring no extrapolation prediction. At the fourth synchronization time of 160 milliseconds, the actual extraction time of region 1, region 2, and region 3 is still 160 milliseconds, and the times of the three regions remain synchronized.

[0087] It should be noted that although the actual extraction time of each region in the aforementioned examples coincides with the preset synchronization time, this is due to the selection of a specific combination of simulation step size and synchronization period, which ensures that the calculation time point falls precisely at the synchronization time. In practical engineering applications, time asynchrony between regions is common, requiring extrapolation prediction algorithms to ensure the accuracy of data interaction. For example, if the simulation step size of region 2 is modified to 18 milliseconds, then at the second synchronization time of 80 milliseconds, the actual calculation points for region 2 are 72 milliseconds and 90 milliseconds, which cannot be accurately reached at 80 milliseconds. At this time, the extrapolation prediction algorithm comes into play, based on the voltage value of 225.6 volts at the equivalent node 3 at 72 milliseconds and the voltage value of 226.8 volts at 90 milliseconds, and calculates the estimated voltage value of 226.1 volts at 80 milliseconds through linear interpolation. For power data with high volatility, quadratic extrapolation can be used to improve accuracy. For example, if the power is 35.4 MW at 72 ms and 37.2 MW at 90 ms, and considering the power value of 33.8 MW at the previous moment of 54 ms, the power at 80 ms can be predicted to be approximately 36.1 MW through quadratic curve fitting. This time-aligned data is passed as boundary conditions to Region 1 and Region 3 to ensure the overall simulation accuracy.

[0088] Based on the boundary connectivity, synchronous boundary data is transmitted to adjacent sub-network regions as boundary conditions. Boundary conditions refer to the state information received by a region from its neighboring regions, serving as known conditions for the next simulation calculation of that region. In the aforementioned example, region 1 transmits the voltage and power data of equivalent node 1 and equivalent node 2 to region 3 and region 2; region 2 transmits the voltage and power data of equivalent node 3 to region 1 and region 3; and region 3 transmits the voltage and power data of equivalent node 4 and equivalent node 5 to region 2 and region 1.

[0089] The asynchronous parallel simulation is repeated until the simulation ends. The simulation end time is specified by the user and represents the total duration of the simulation process. In the example above, the simulation end time is set to 1000 milliseconds. Each sub-network region continues to perform simulation calculations based on boundary data interaction. Region 1 needs to perform 200 calculations, Region 2 needs to perform 50 calculations, and Region 3 needs to perform 125 calculations. The simulation process in the three regions proceeds simultaneously until the termination condition is met.

[0090] The simulation output data from each sub-network region are aggregated to obtain the global simulation results. The simulation output data includes electrical quantities such as voltage, current, and power of nodes in each region, as well as system stability evaluation indicators. In the example above, the simulation output data of the three regions are organized according to region affiliation and time sequence to form a complete simulation result of the grid-connected operation of the new energy power station, including the state changes of the five equivalent nodes throughout the simulation process, as well as the overall system stability and grid-connected performance evaluation.

[0091] In this embodiment, the computational load is evaluated by comprehensively considering the topological complexity and node size of the sub-network regions. Based on this, differentiated simulation step sizes are allocated to different sub-network regions, so that the configuration of computing resources can match the actual computing needs of each region. This avoids the problem of local complex regions slowing down the overall progress and significantly improves the overall simulation efficiency. By identifying the boundary connection relationship between sub-networks and building a unified boundary node management mechanism, cross-regional information exchange has clear structural constraints, which helps to reduce unnecessary data synchronization and communication overhead. In the asynchronous parallel simulation process, boundary nodes are interacted only at preset synchronization times. Compared with the traditional parallel method of synchronization at each step, the synchronization frequency and communication cost are significantly reduced, and the parallel execution efficiency is improved. For the time deviation caused by the independent step size operation of different sub-network regions, the boundary interaction data is time extrapolated and aligned to keep the data transmitted across regions consistent in time scale. This avoids the problem of simulation error accumulation caused by directly using unaligned data. While ensuring the advantages of asynchronous execution, the consistency and credibility of the global simulation results are maintained.

[0092] Figure 2 This is a flowchart of the asynchronous power network simulation method for the grid-connected operation of new energy power plants, as described in an embodiment of the present invention.

[0093] In one alternative implementation, The topology complexity parameters and node count parameters of each sub-network region are extracted from the partitioned topology. The computational load of each sub-network region is calculated based on these parameters. Based on the computational load and the coupling strength between different sub-network regions, an independent simulation step size is allocated to each sub-network region using a load balancing optimization algorithm, including: The node count parameter is obtained by traversing each sub-network region in the partition topology and counting the total number of nodes in each sub-network region. The connection complexity between nodes in each sub-network region is calculated and combined with the topology level depth in the sub-network region to obtain the topology complexity parameter. The computational load of each sub-network region is calculated based on the node count parameter and the topology complexity parameter. Extract the number of boundary node connections between different sub-network regions in the partitioned topology, count the power exchange frequency of boundary nodes between each sub-network region and its adjacent sub-network regions, calculate the coupling strength between different sub-network regions based on the number of boundary node connections and the power exchange frequency of boundary nodes, and construct a coupling strength matrix. Based on the computational load and the coupling strength matrix, an optimization function is constructed with the goal of minimizing the total simulation time using a load balancing optimization algorithm. Electrical characteristic parameters of each sub-network region are extracted from the partitioned topology. Based on the electrical characteristic parameters and the preset synchronization time, a simulation numerical stability criterion is constructed. The simulation numerical stability criterion is added to the optimization function as a simulation stability constraint and the independent simulation step size of each sub-network region is obtained by solving the algorithm.

[0094] The node count parameter is obtained by traversing each sub-network region in the partitioned topology and counting the total number of nodes in each sub-network region. The node count parameter directly reflects the region size and affects the computational complexity. In the example above, the partitioned topology contains 3 sub-network regions: Region 1 contains 2 nodes (equivalent node 1 and equivalent node 2); Region 2 contains 1 node (equivalent node 3); and Region 3 contains 2 nodes (equivalent node 4 and equivalent node 5).

[0095] The topology complexity parameter is obtained by calculating the connection complexity between nodes within each sub-network region and combining it with the topology depth within the sub-network region. Connection complexity is calculated as the ratio of the number of connections within a region to the maximum possible number of connections. The topology depth is the number of levels in the longest path between nodes within a region. In the previous example, region 1 has 1 connection, a maximum possible number of connections of 1, a connection complexity of 1.0, and a topology depth of 1, resulting in a combined topology complexity parameter of 0.8. Region 2 has only 1 node, 0 internal connections, and a topology complexity parameter of 0.3. Region 3 has 1 connection, a maximum possible number of connections of 1, a connection complexity of 1.0, and a topology depth of 1, resulting in a combined topology complexity parameter of 0.5. Higher topology complexity indicates more complex power flow calculations within the region.

[0096] The computational load of each sub-network region is calculated based on the node count parameter and the topology complexity parameter. The computational load assessment considers the product of the square of the node count and the topology complexity parameter, reflecting the computational resources required for power flow calculation within the region. In the example above, the computational load of region 1 is 2 × 2 × 0.8 = 3.2; the computational load of region 2 is 1 × 1 × 0.3 = 0.3; and the computational load of region 3 is 2 × 2 × 0.5 = 2.0.

[0097] Extract the number of boundary node connections between different sub-network regions in the partitioned topology. A boundary node connection refers to a direct connection between different sub-network regions, and each connection involves one boundary node in each of the two regions. In the example above, there is one boundary node connection between region 1 and region 2, and the equivalent node 2 of region 1 is connected to the equivalent node 3 of region 2; there is one boundary node connection between region 2 and region 3, and the equivalent node 3 of region 2 is connected to the equivalent node 4 of region 3; there is one boundary node connection between region 1 and region 3, and the equivalent node 1 of region 1 is connected to the equivalent node 5 of region 3.

[0098] The power exchange frequency between boundary nodes of each sub-network region and its adjacent sub-network regions is statistically analyzed. Power exchange frequency refers to the number of times the power value between boundary nodes exceeds a preset threshold per unit time, reflecting the dynamic characteristics of boundary interactions. In the aforementioned example, based on historical simulation data analysis, the power exchange frequency between boundary nodes in region 1 and region 2 is 8 times per second; the power exchange frequency between boundary nodes in region 2 and region 3 is 12 times per second; and the power exchange frequency between boundary nodes in region 1 and region 3 is 6 times per second. A higher power exchange frequency indicates more frequent interactions between regions and a higher requirement for synchronization accuracy.

[0099] The coupling strength between different sub-network regions is calculated based on the number of boundary node connections and the power exchange frequency of the boundary nodes, and a coupling strength matrix is ​​constructed. The coupling strength is calculated by a weighted combination of the number of boundary node connections and the power exchange frequency, reflecting the tightness of the interaction between regions. In the example above, the coupling strength between region 1 and region 2 is 0.4; the coupling strength between region 2 and region 3 is 0.5; and the coupling strength between region 1 and region 3 is 0.3. A 3×3 coupling strength matrix is ​​constructed based on the coupling strength, with diagonal elements of 0, indicating that a region is not coupled to itself. The coupling strength matrix is ​​a key constraint for simulation step size optimization.

[0100] Based on the computational load and coupling strength matrix, an optimization function is constructed using a load balancing optimization algorithm to minimize the total simulation time. The total simulation time is determined by the sub-network region with the longest computation time, which is related to the computational load of the region and the simulation step size. The optimization objective is to minimize the overall simulation completion time by reasonably allocating the simulation step size while ensuring simulation accuracy. In the aforementioned example, the optimization function considers the computational loads of each region (3.2, 0.3, 2.0) and the coupling strength matrix between them to construct the objective function for solving the simulation step size.

[0101] Electrical characteristic parameters are extracted from each sub-network region within the partitioned topology. These parameters include the inertial time constant of the generator units, the excitation system time constant, and the speed control system time constant. These parameters influence the characteristic timescale of the electrical transient processes within the region. In the aforementioned example, the main electrical characteristic parameters for region 1 are: an inertial time constant of 4 seconds and an excitation system time constant of 0.05 seconds; for region 2, they are: an inertial time constant of 3.5 seconds and an excitation system time constant of 0.04 seconds; and for region 3, they are: an inertial time constant of 4.2 seconds and an excitation system time constant of 0.06 seconds. These electrical characteristic parameters determine the maximum allowable step size for the region simulation.

[0102] A numerical stability criterion for simulation is constructed based on electrical characteristic parameters and a preset synchronization time. This criterion is used to evaluate the numerical stability of the computation process under a given simulation step size. It typically determines the maximum allowable step size based on electrical characteristic parameters and assesses the coordination of step size combinations in conjunction with the preset synchronization time. In the aforementioned example, based on the electrical characteristic parameters of region 1, its maximum allowable step size is 10 milliseconds; the maximum allowable step size for region 2 is 25 milliseconds; and the maximum allowable step size for region 3 is 12 milliseconds. Simultaneously, considering the constraint of a preset synchronization time of 40 milliseconds, the step size for each region must be divisible by the synchronization period or produce a controllable time deviation.

[0103] The simulation numerical stability criterion is added as a simulation stability constraint to the optimization function, and the independent simulation step size for each sub-network region is obtained by solving the problem. The simulation stability constraint ensures that the allocated step size does not exceed the maximum allowable step size for each region and satisfies the coupling strength requirements between adjacent regions. In the previous example, by solving the optimization problem considering computational load balancing and simulation stability, the independent simulation step size for region 1 is 5 milliseconds, the independent simulation step size for region 2 is 20 milliseconds, and the independent simulation step size for region 3 is 8 milliseconds. The computational load per millisecond for region 1 is 3.2 / 5 = 0.64, the computational load per millisecond for region 2 is 0.3 / 20 = 0.015, and the computational load per millisecond for region 3 is 2.0 / 8 = 0.25. Meanwhile, considering the high coupling strength of 0.4 between region 1 and region 2, the step size ratio is 5:20=1:4, which satisfies the coupling strength constraint; the higher coupling strength of 0.5 between region 2 and region 3, the step size ratio is 20:8=2.5:1, which also satisfies the coupling degree constraint; the lower coupling strength of 0.3 between region 1 and region 3, the step size ratio is 5:8=0.625:1, which also satisfies the constraint conditions.

[0104] In this embodiment, by simultaneously considering the node size, internal connection complexity, and topology depth of the sub-network region, the computational load is comprehensively quantified, enabling the load evaluation results to truly reflect the actual overhead of different sub-networks in the simulation calculation, significantly improving the accuracy of load characterization. The number of boundary node connections and power exchange frequency are introduced to measure the coupling strength between sub-networks, allowing for a quantitative expression of the degree of mutual influence between regions. This provides a dual constraint basis of structure and operation for step size allocation. By constructing a load balancing optimization model with the goal of minimizing the total simulation time, and incorporating the electrical characteristics of the sub-network and the preset synchronization time into the numerical stability criterion, the step size allocation process is explicitly controlled by the numerical stability requirements of the system while pursuing computational efficiency. This avoids the simulation divergence risk caused by excessively large step sizes in local regions and also prevents overall efficiency loss caused by overly conservative step size settings.

[0105] A second aspect of the present invention provides a simulation system for the grid-connected operation of a new energy power station, comprising: The feature clustering module is used to obtain the operation configuration information of the new energy power station and the access point information of the grid side. Based on the operation configuration information, the module extracts the operation status features of the power generation unit and constructs a time-series feature matrix. It extracts the dominant feature components corresponding to the time-series feature matrix to construct a feature vector set and calculates the electrical similarity index between the power generation units. The topology simplification module is used to group the power generation units into multiple power generation groups based on the electrical similarity index using an agglomerative hierarchical clustering algorithm, calculate adaptive weighting coefficients based on the electrical similarity index, and perform weighted aggregation of the electrical impedance characteristics of each power generation unit within the power generation group to obtain equivalent electrical parameters, construct equivalent nodes based on the equivalent electrical parameters, and determine the simplified topology structure. The spectral domain partitioning module is used to determine the boundary constraints on the power grid side based on the access point information and apply them to the simplified topology, determine the Laplace matrix corresponding to the simplified topology and construct the spectral embedding space, map the equivalent nodes to the spectral embedding space to identify the node cluster boundaries, and divide the simplified topology into multiple sub-network regions to obtain the partitioned topology. The asynchronous simulation module is used to configure independent simulation step sizes for sub-network regions according to the partition topology and perform asynchronous parallel simulation. It extracts boundary node data at a preset synchronization time, and after time alignment processing of the boundary interaction data by combining the extrapolation prediction algorithm, it transmits the data to adjacent sub-network regions to obtain global simulation results.

[0106] A third aspect of the present invention provides an electronic device, comprising: A processor and a memory for storing processor-executable instructions, wherein the processor is configured to invoke instructions stored in the memory to perform the aforementioned method.

[0107] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0108] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0109] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A simulation method for grid-connected operation of new energy power plants, characterized in that, include: Obtain the operation configuration information of the new energy power station and the grid access point information. Based on the operation configuration information, extract the operation status features of the power generation unit and construct a time-series feature matrix. Extract the dominant feature components corresponding to the time-series feature matrix to construct a feature vector set and calculate the electrical similarity index between the power generation units. Based on the electrical similarity index, the power generation units are grouped into multiple power generation groups using an agglomerative hierarchical clustering algorithm. An adaptive weighting coefficient is calculated based on the electrical similarity index, and the electrical impedance characteristics of each power generation unit within the power generation group are weighted and aggregated to obtain equivalent electrical parameters. Based on the equivalent electrical parameters, equivalent nodes are constructed and a simplified topology is determined. Based on the access point information, the boundary constraints on the power grid side are determined and applied to the simplified topology. The Laplace matrix corresponding to the simplified topology is determined and a spectral embedding space is constructed. The equivalent nodes are mapped to the spectral embedding space to identify the node cluster boundaries and the simplified topology is divided into multiple sub-network regions to obtain a partitioned topology. Based on the partitioned topology, an independent simulation step size is configured for each sub-network region, and asynchronous parallel simulation is performed. Boundary node data is extracted at a preset synchronization time. After time alignment processing of the boundary interaction data using an extrapolation prediction algorithm, the data is transmitted to adjacent sub-network regions to obtain the global simulation results.

2. The method according to claim 1, characterized in that, The process involves acquiring the operation configuration information of the renewable energy power station and the grid connection point information. Based on the operation configuration information, the operation status features of the power generation units are extracted and a time-series feature matrix is ​​constructed. The dominant feature components corresponding to the time-series feature matrix are extracted to construct a feature vector set, and the electrical similarity index between the power generation units is calculated, including: The power output sequence and control response sequence of each power generation unit under multiple operating conditions are collected, and the timestamps are aligned and abnormal data points are removed to obtain the effective power sequence and effective control sequence. The time sequence feature matrix is ​​then constructed by splicing the sequences according to the time dimension. The time series feature matrix is ​​standardized to obtain a standardized time series matrix. A covariance matrix is ​​constructed based on the standardized time series matrix and eigenvalue decomposition is performed. The eigenvalues ​​are sorted in descending order according to their size and the eigenvectors corresponding to the top few eigenvalues ​​are selected as principal component directions based on the cumulative contribution rate. The standardized time series matrix is ​​then projected onto the principal component directions to obtain the dominant feature components. Extract the dominant feature components corresponding to each power generation unit and organize them according to the power generation unit number to construct the feature vector set. Perform covariance calculation on the feature vectors of any two power generation units in the feature vector set and construct a covariance matrix. Perform matrix inversion operation on the covariance matrix to obtain the inverse covariance matrix. Calculate the difference vector between the feature vectors of different power generation units and perform weighted distance calculation with the corresponding inverse covariance matrix to obtain the Mahalanobis distance. Perform normalization mapping on the Mahalanobis distance to obtain the electrical similarity index.

3. The method according to claim 1, characterized in that, Based on the electrical similarity index, the power generation units are grouped into multiple power generation groups using an agglomerative hierarchical clustering algorithm. Adaptive weighting coefficients are calculated based on the electrical similarity index, and the electrical impedance characteristics of each power generation unit within a power generation group are weighted and aggregated to obtain equivalent electrical parameters. Based on these equivalent electrical parameters, equivalent nodes are constructed, and a simplified topology is determined, including: Based on the electrical similarity index, a similarity matrix is ​​constructed and the power generation units are treated as independent clusters. The power generation unit pairs with the highest similarity in the similarity matrix are calculated and cluster merging operations are performed to obtain merged clusters and update the similarity matrix. The clustering quality is calculated by combining the silhouette coefficient evaluation algorithm. The cluster merging operation is repeated until the clustering quality converges to obtain multiple power generation groups. Calculate the electrical similarity index between each power generation unit in the power generation group and the center of the power generation group, and perform reverse mapping to obtain adaptive weighting coefficients. Extract the resistance parameters and reactance parameters of each power generation unit in the power generation group from the operation configuration information and construct an electrical impedance feature vector. An adjacency matrix is ​​constructed based on the electrical connection relationship between each power generation unit within the power generation group. A symmetric transformation is performed on the adjacency matrix using a graph convolution propagation algorithm to obtain a symmetric adjacency matrix. Multi-hop propagation is then performed on the electrical impedance feature vector to obtain multi-order neighborhood features. Layer-by-layer aggregation is performed on the multi-order neighborhood features to obtain neighborhood aggregation features. Weighted summation and nonlinear activation transformation operations are then performed on the adaptive weighting coefficients to obtain equivalent electrical parameters. The equivalent node is constructed for each power generation group based on the equivalent electrical parameters. The connection topology between the power generation unit and the combiner unit in the operation configuration information is extracted. The simplified topology is constructed based on the equivalent node and the connection topology.

4. The method according to claim 3, characterized in that, The adjacency matrix is ​​subjected to a symmetric transformation using a graph convolution propagation algorithm to obtain a symmetric adjacency matrix. This symmetric matrix is ​​then combined with the electrical impedance feature vector to perform a multi-hop propagation operation, resulting in multi-order neighborhood features, including: The connection relationship information of each power generation unit node in the adjacency matrix is ​​extracted and the node degree is counted to construct a degree matrix. The degree matrix is ​​subjected to an inverse transformation to obtain a degree inverse matrix, and the adjacency matrix is ​​subjected to a two-sided symmetric transformation to obtain a symmetric adjacency matrix. A topology-aware mask matrix is ​​constructed based on the symmetric adjacency matrix. The electrical impedance feature vector is used as a zero-order feature input. The capacity parameters and impedance parameters of each power generation unit are extracted from the operation configuration information and a node attribute embedding vector is constructed. The neighborhood propagation feature is obtained by performing neighborhood information aggregation on the current layer feature vector and the symmetric adjacency matrix using a graph convolution algorithm. The neighborhood propagation feature is then fused with the node attribute embedding vector using a feature fusion algorithm to obtain an enhanced propagation feature. The enhanced propagation feature is then subjected to a learnable weight transformation and combined with the topology-aware mask matrix to perform selective feature filtering to obtain the current layer output feature. The residual enhanced feature is then obtained by solving based on the current layer output feature and the current layer input feature. Determine whether the current propagation layer has reached the preset maximum propagation layer. If not, use the residual enhancement feature as the input feature of the next layer and repeat the neighborhood propagation operation. If it has reached the maximum, extract the residual enhancement features output by each propagation layer and calculate the hierarchical importance weight based on the propagation depth. Aggregate the residual enhancement features of each propagation layer with the corresponding hierarchical importance weight to obtain multi-level neighborhood features.

5. The method according to claim 1, characterized in that, Based on the access point information, boundary constraints on the power grid side are determined and applied to the simplified topology. The Laplace matrix corresponding to the simplified topology is determined, and a spectral embedding space is constructed. The equivalent nodes are mapped to the spectral embedding space to identify node cluster boundaries. The simplified topology is then divided into multiple sub-network regions to obtain a partitioned topology, including: Extract the access point location identifier and grid-side access capacity limitation information from the access point information; construct a power constraint set based on the grid-side access capacity limitation information and bind it to the equivalent node in the simplified topology corresponding to the access point location identifier; construct a boundary node label matrix based on the equivalent node that has been bound to the power constraint set. Extract the connection relationships between equivalent nodes in the simplified topology and construct a connection matrix. Calculate the number of connections for each equivalent node based on the connection matrix, construct a node degree vector, and convert it into a diagonal matrix to obtain a degree diagonal matrix. Perform a matrix difference operation between the degree diagonal matrix and the connection matrix to obtain the Laplace matrix. The eigenvalue decomposition of the Laplacian matrix is ​​performed to obtain the eigenvector matrix, and the spectral embedding space is constructed based on the eigenvector matrix; The equivalent nodes in the simplified topology are projected onto the spectral embedding space to obtain spectral coordinate representations. A spectral similarity matrix is ​​constructed using a spectral clustering algorithm. The spectral similarity matrix is ​​constrained and adjusted based on the boundary node label matrix, and the boundary positions between dense and sparse spectral coordinate regions are identified. The node cluster boundaries are determined based on the boundary positions, and the equivalent nodes in the simplified topology are divided into different sub-network regions. A sub-network region identifier mapping table is constructed, and the simplified topology is decomposed based on the sub-network region identifier mapping table to obtain a partitioned topology.

6. The method according to claim 1, characterized in that, Based on the partitioned topology, an independent simulation step size is configured for each sub-network region, and asynchronous parallel simulation is performed. Boundary node data is extracted at a preset synchronization time. After time alignment processing of the boundary interaction data using an extrapolation prediction algorithm, the data is transmitted to adjacent sub-network regions to obtain global simulation results, including: The topology complexity parameters and node quantity parameters of each sub-network region are extracted from the partition topology structure. The computational load of each sub-network region is calculated based on the topology complexity parameters and the node quantity parameters. Based on the computational load and the coupling strength between different sub-network regions, an independent simulation step size is allocated to each sub-network region through a load balancing optimization algorithm, and a step size configuration table is constructed based on the independent simulation step size. Identify the boundary connection relationships between each sub-network region from the partition topology and extract the boundary node identifiers, and construct a boundary node list based on the boundary node identifiers; According to the step size configuration table, each sub-network region is driven to perform asynchronous parallel simulation according to the corresponding independent simulation step size. When the simulation time of each sub-network region reaches the preset synchronization time, the corresponding boundary node data is extracted from the boundary node list as boundary interaction data and the corresponding actual extraction time is recorded. Based on the time deviation between the actual extraction time and the preset synchronization time, the boundary interaction data is time-aligned by an extrapolation prediction algorithm to obtain synchronized boundary data. According to the boundary connection relationship, the synchronized boundary data is passed to the adjacent sub-network region as boundary conditions. Asynchronous parallel simulation is repeated until the simulation ends. The simulation output data of each sub-network region is summarized to obtain the global simulation result.

7. The method according to claim 6, characterized in that, The topology complexity parameters and node count parameters of each sub-network region are extracted from the partitioned topology. The computational load of each sub-network region is calculated based on these parameters. Based on the computational load and the coupling strength between different sub-network regions, an independent simulation step size is allocated to each sub-network region using a load balancing optimization algorithm, including: The node count parameter is obtained by traversing each sub-network region in the partition topology and counting the total number of nodes in each sub-network region. The connection complexity between nodes in each sub-network region is calculated and combined with the topology level depth in the sub-network region to obtain the topology complexity parameter. The computational load of each sub-network region is calculated based on the node count parameter and the topology complexity parameter. Extract the number of boundary node connections between different sub-network regions in the partitioned topology, count the power exchange frequency of boundary nodes between each sub-network region and its adjacent sub-network regions, calculate the coupling strength between different sub-network regions based on the number of boundary node connections and the power exchange frequency of boundary nodes, and construct a coupling strength matrix. Based on the computational load and the coupling strength matrix, an optimization function is constructed with the goal of minimizing the total simulation time using a load balancing optimization algorithm. Electrical characteristic parameters of each sub-network region are extracted from the partitioned topology. Based on the electrical characteristic parameters and the preset synchronization time, a simulation numerical stability criterion is constructed. The simulation numerical stability criterion is added to the optimization function as a simulation stability constraint and the independent simulation step size of each sub-network region is obtained by solving the algorithm.

8. A simulation system for grid-connected operation of new energy power plants, used to implement the method described in any one of claims 1-7, characterized in that, include: The feature clustering module is used to obtain the operation configuration information of the new energy power station and the access point information of the grid side. Based on the operation configuration information, the module extracts the operation status features of the power generation unit and constructs a time-series feature matrix. It extracts the dominant feature components corresponding to the time-series feature matrix to construct a feature vector set and calculates the electrical similarity index between the power generation units. The topology simplification module is used to group the power generation units into multiple power generation groups based on the electrical similarity index using an agglomerative hierarchical clustering algorithm, calculate adaptive weighting coefficients based on the electrical similarity index, and perform weighted aggregation of the electrical impedance characteristics of each power generation unit within the power generation group to obtain equivalent electrical parameters, construct equivalent nodes based on the equivalent electrical parameters, and determine the simplified topology structure. The spectral domain partitioning module is used to determine the boundary constraints on the power grid side based on the access point information and apply them to the simplified topology, determine the Laplace matrix corresponding to the simplified topology and construct the spectral embedding space, map the equivalent nodes to the spectral embedding space to identify the node cluster boundaries, and divide the simplified topology into multiple sub-network regions to obtain the partitioned topology. The asynchronous simulation module is used to configure independent simulation step sizes for sub-network regions according to the partition topology and perform asynchronous parallel simulation. It extracts boundary node data at a preset synchronization time, and after time alignment processing of the boundary interaction data by combining the extrapolation prediction algorithm, it transmits the data to adjacent sub-network regions to obtain global simulation results.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images