Power grid electrical state determination method and device, storage medium and electronic equipment

Abstract

The application discloses a power grid electrical state determination method and device, a storage medium and an electronic device. The method comprises the following steps: acquiring a topological structure of a target power grid, measurement data of each bus in the target power grid, and a sensitivity matrix corresponding to a power flow equation of the target power grid; constructing a graph model of the target power grid based on the topological structure, wherein the graph model takes the buses as nodes and electrical branches as connecting edges; generating corresponding initial features for each node in the graph model based on the measurement data of each bus; determining the electrical coupling strength between different buses based on the sensitivity matrix, and performing weighted fusion processing on the initial features of all buses one by one according to the electrical coupling strength to obtain node fusion features corresponding to all buses; and performing multi-layer feature transformation processing on the node fusion features of all buses in combination with the node connection relationship of the graph model to obtain electrical state estimation values corresponding to all buses.

Description

Technical Field

[0001] This application relates to the field of power system technology, and in particular to a method, apparatus, storage medium and electronic equipment for determining the electrical state of a power grid. Background Technology

[0002] Accurate acquisition of the electrical status of the entire network, such as the voltage amplitude and phase angle of the bus, is the core foundation for power grid safety dispatch, fault identification, and new energy consumption control. The measurement data collected on site often suffers from problems such as communication interruption, sensor failure, and incomplete coverage. Relying solely on limited measurement data cannot directly obtain complete and reliable electrical information of the entire network. Therefore, it is necessary to determine the electrical status of the power grid.

[0003] In the field of power grid electrical state determination, the mainstream approach employs a weighted least squares method driven by a physical model. This method is based on the nonlinear power flow equations of the power grid, setting the voltage magnitude and phase angle of all network buses as the state variables to be estimated. It constructs measurement equations by combining measured data such as active power injection, reactive power injection, and node voltages collected by the system, with the optimization objective being to minimize the weighted sum of squared residuals between the measured values ​​and the calculated state variables. During the solution process, the method linearizes the power flow equations round by round, successively calculating the Jacobian matrix and gain matrix. It updates the incremental values ​​of the state variables by solving a set of linear correction equations, iterating repeatedly until the state increment meets a preset convergence threshold, and finally outputting the electrical state estimation results for all buses in the entire network. The core premise for this algorithm to obtain a stable and unique state estimation solution is that the system has sufficient measurement redundancy. It relies on multiple sets of mutually redundant measured data to jointly constrain the feasible region of the state variables, thereby offsetting the random errors of a single set of measurements.

[0004] However, the estimation accuracy of this method is highly dependent on the measurement redundancy. When the proportion of missing measurements exceeds the preset proportion, the effective observation information that can be used to constrain the state variables is greatly reduced, the observability of the system decreases, the interference of measurement noise on the solution results is amplified, the equation system is prone to underdetermined or ill-conditioned problems, and ultimately the accuracy of the state estimation results is greatly reduced. Summary of the Invention

[0005] In view of the above problems, this application provides a method, apparatus, storage medium and electronic device for determining the electrical state of a power grid.

[0006] To solve the above-mentioned technical problems, this application proposes the following solution: Firstly, this application provides a method for determining the electrical state of a power grid. The method includes: acquiring the topology of a target power grid, measurement data of each bus in the target power grid, and the sensitivity matrix corresponding to the power flow equation of the target power grid; constructing a graph model of the target power grid based on the topology, wherein the graph model has buses as nodes and electrical branches as connecting edges; generating corresponding initial features for each node in the graph model based on the measurement data of each bus; determining the electrical coupling strength between different buses based on the sensitivity matrix, and performing weighted fusion processing on the initial features of all buses one by one according to the electrical coupling strength to obtain the node fusion features corresponding to all buses; and performing multi-level feature transformation processing on the node fusion features of all buses in combination with the node connection relationship of the graph model to obtain the electrical state estimate corresponding to all buses.

[0007] This application no longer relies solely on sufficient redundant measurements to construct power flow optimization equations. Instead, it leverages sensitivity matrices to uncover inherent electrical coupling relationships between buses, completes the associated electrical information corresponding to missing measurement points through feature aggregation, and then deeply mines the implicit operating rules of the power grid through multi-layer feature transformation. Even under conditions of missing measurements and insufficient effective observation information, it can still supplement effective constraint information based on the coupling relationship of the entire network buses, avoiding underdetermined and ill-conditioned problems in the equation set. This solves the defect of a significant decrease in estimation accuracy when the weighted least squares method has missing measurements.

[0008] In conjunction with the first aspect, in one possible implementation, the sensitivity matrix is ​​truncated singular value decomposition, and low-dimensional feature vectors corresponding to each bus are constructed based on the decomposition results; the correlation weight between any two buses is calculated based on the low-dimensional feature vectors of each bus, and the correlation weight is used to characterize the electrical coupling strength between the buses; the initial features of all buses are weighted and fused one by one based on the correlation weight to obtain the node fusion features corresponding to all buses.

[0009] Large-scale power grids have extremely high-dimensional complete sensitivity matrices, resulting in enormous storage and computational costs. Furthermore, the original measurement features of a single busbar do not include the coupling effects of surrounding lines. This application optimizes node features through an integrated approach: matrix dimensionality reduction and compression, quantification of busbar coupling strength, and weighted fusion of neighborhood features. First, dimensionality reduction of the matrix filters out invalid noise, reducing the overall computational load. Then, the mutual influence weights between different buses are quantified, fusing the operational information of adjacent buses into the single-node features. This improves computational speed while simultaneously enhancing the accuracy of state estimation from the source.

[0010] In conjunction with the first aspect, in one possible implementation, the sensitivity matrix is ​​subjected to complete singular value decomposition to obtain an orthogonal singular vector matrix and a set of singular values. From the set of singular values, singular values ​​of a predetermined order and orthogonal singular vectors corresponding to the predetermined order are selected, and singular values ​​and orthogonal singular vectors corresponding to the remaining orders are discarded. For each row of the orthogonal singular vector matrix corresponding to a single busbar, the orthogonal singular vectors of the predetermined order corresponding to the row are extracted, and each orthogonal singular vector of each order is multiplied element-wise by the square root of its corresponding singular value to obtain the low-dimensional feature vector corresponding to the single busbar.

[0011] This application performs complete singular value decomposition on the sensitivity matrix and truncates and retains the core components of the preset order, eliminates high-order noise modes, and then combines singular value square root fusion to generate a low-dimensional feature vector of the bus. This not only compresses the high-dimensional matrix and significantly reduces the storage and computing overhead under large-scale power grids, but also completely preserves the real electrical coupling law of the power grid and avoids the problem of losing physical information due to random dimensionality reduction.

[0012] In conjunction with the first aspect, in one possible implementation, the low-dimensional feature vectors corresponding to any two buses are selected, and the corresponding values ​​of the two sets of low-dimensional feature vectors are multiplied item by item and all values ​​are added together to obtain the original interaction values ​​between the two buses. The original interaction values ​​are subjected to exponential transformation to obtain a single set of transformed values. Taking one of the two buses as the reference bus, the transformed values ​​corresponding to all buses connected to the reference bus by physical electrical branches are accumulated to obtain a summation result. Based on the summation result, the single set of transformed values ​​is normalized to obtain the correlation weight between any two buses.

[0013] This application obtains the original interaction values ​​of the busbars through the inner product of low-dimensional feature vectors, amplifies the numerical differences between strong and weak couplings through exponential transformation, and then performs local normalization to generate standardized correlation weights. The weight value range is unified and naturally satisfies the bidirectional symmetry characteristic, accurately quantifying the true electrical coupling strength between different buses. This allows the weighted aggregation process of the initial features of adjacent buses to conform to the physical law of bidirectional disturbance propagation in undirected branches of the power grid, strengthens the proportion of measurement information of strongly coupled buses in the node fusion features, and fills the measurement gaps by relying on the effective information of adjacent strongly coupled buses in the scenario of missing measurements, thereby improving the ability of node fusion features to restore the electrical operating status of the entire network.

[0014] In conjunction with the first aspect, in one possible implementation, a preset weight matrix is ​​used to perform a linear transformation on the node fusion features of each node to generate a linear transformation feature vector corresponding to each node; each element in the linear transformation feature vector is subjected to nonlinear mapping processing to obtain the output features of the single-layer feature transformation; the single-layer feature transformation is repeatedly executed and connected layer by layer, with each layer using the output features of the single-layer feature transformation of the previous layer as the input features of the current layer, thereby realizing multi-layer feature transformation.

[0015] This application achieves feature dimension reorganization and information decoupling by performing linear transformation on node fusion features using a learnable weight matrix. Then, it performs nonlinear mapping on the transformed vector element by element to fully explore the complex nonlinear electrical coupling relationship behind the steady-state operation and dynamic disturbances of the power grid, enrich the feature expression capability, distinguish the changing patterns of the bus electrical state under different operating conditions, provide effective features with nonlinear expression capability for subsequent multi-layer deep feature extraction, further make up for the lack of information caused by measurement deficiencies, and improve the accuracy of the final electrical state estimation results.

[0016] In conjunction with the first aspect, in one possible implementation, when performing a single-layer feature transformation, the input features of the single-layer feature transformation are directly passed to the output end of the single-layer feature transformation and superimposed with the features output by the single-layer feature transformation; the feature transformations of a continuous preset number of layers are divided into a set of feature transformations, and the input features of the first layer feature transformation in the feature transformation set are directly passed to the output end of the last layer feature transformation in the feature transformation set and superimposed with the features output by the last layer feature transformation.

[0017] Deep network training often suffers from gradient vanishing, loss of original measurement information layer by layer, slow model convergence, and significant fluctuations in computational results under extreme scenarios such as faults and large disturbances. This application optimizes the training effect of deep networks by setting up a dual identity information transmission structure of intra-layer short paths and block long paths. Intra-layer paths ensure normal backpropagation of gradients between adjacent layers, while block long paths retain basic operational information across multiple layers, taking into account both global baseline conditions and local operational details, accelerating model convergence speed, and significantly improving the stability of state estimation results under extreme conditions.

[0018] In conjunction with the first aspect, in one possible implementation, the preset weight matrices corresponding to each layer of linear transformation are orthogonally constrained based on preset constraint rules to generate corresponding correction offsets. The preset constraint rule is that the difference magnitude of any two sets of node features before and after the linear transformation calculation of the same weight matrix remains equidistant. Based on the correction offsets and the deviation between the node features output by the multi-layer feature transformation and the actual electrical measurement data, the preset weight matrices corresponding to each layer of linear transformation are adjusted so that the linear transformation calculation of each layer meets the orthogonal equidistant transformation requirements.

[0019] Unconstrained training leads to weight matrix distortion, a continuous reduction in the relative distance between node features, decreased discrimination of deep features, and a decline in model generalization ability. This application introduces feature equidistance constraints to generate weight correction terms, and updates network parameters synchronously with measurement bias. The correction offset is generated with the constraint that the feature differences before and after transformation remain equidistance. The training process simultaneously considers estimation accuracy and geometric constraints, stabilizes the feature space structure, avoids convergence of deep features, and maintains high estimation accuracy and generalization ability even in noisy and incomplete measurement scenarios.

[0020] Secondly, this application provides a power grid electrical state determination device, which includes: The acquisition module is used to acquire the topology of the target power grid, the measurement data of each bus in the target power grid, and the sensitivity matrix corresponding to the power flow equation of the target power grid. The building module is used to construct a graph model of the target power grid based on the topology. The graph model uses buses as nodes and electrical branches as connecting edges. The initial feature module is used to generate corresponding initial features for each node in the graph model based on the measurement data of each bus. The feature fusion module is used to determine the electrical coupling strength between different buses based on the sensitivity matrix, and to perform weighted fusion processing on the initial features of all buses according to the electrical coupling strength to obtain the node fusion features corresponding to all buses. The determination module is used to perform multi-level feature transformation processing on the node fusion features of all buses in combination with the node connection relationship of the graph model to obtain the electrical state estimate of all buses.

[0021] To achieve the above objectives, according to a third aspect of this application, a storage medium is provided, the storage medium including a stored program, wherein, when the program is executed, the device where the storage medium is located is controlled to perform the power grid electrical state determination method of the first aspect.

[0022] To achieve the above objectives, according to a fourth aspect of this application, an electronic device is provided, the device including at least one processor, and at least one memory and bus connected to the processor; wherein the processor and memory communicate with each other through the bus; the processor is used to call program instructions in the memory to execute the power grid electrical state determination method of the first aspect described above.

[0023] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0024] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of this application. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 A flowchart illustrating a method for determining the electrical state of a power grid according to an embodiment of this application is shown. Figure 2 A flowchart illustrating a node fusion feature determination method provided in an embodiment of this application is shown. Figure 3 A flowchart illustrating a single-layer feature transformation method provided in an embodiment of this application is shown; Figure 4 This illustration shows a flowchart of a fault handling method provided in an embodiment of this application. Figure 5 This paper shows a schematic diagram of the structure of a power grid electrical state determination device provided in an embodiment of this application; Figure 6 A schematic diagram of the structure of an electronic device provided in an embodiment of this application is shown. Detailed Implementation

[0025] Exemplary embodiments of the present application will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present application are shown in the drawings, it should be understood that the present application may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this application will be thorough and complete, and will fully convey the scope of the present application to those skilled in the art.

[0026] In the embodiments of this application, the terms "first," "second," etc., do not have a logical or temporal dependency, nor do they limit the quantity or execution order. It should also be understood that although the following description uses the terms "first," "second," etc., to describe various elements, these elements should not be limited by the terms. These terms are merely used to distinguish one element from another.

[0027] In this application, the term "at least one" means one or more, and the term "multiple" means two or more.

[0028] It should also be understood that the term “if” can be interpreted as “when” or “upon”, or “in response to determination” or “in response to detection”. Similarly, depending on the context, the phrase “if determination…” or “if detection [the stated condition or event]” can be interpreted as “when determination…” or “in response to determination…” or “when detection [the stated condition or event]” or “in response to detection [the stated condition or event]”.

[0029] Power grid safety and stability dispatch, rapid fault diagnosis, and distributed new energy coordinated control all rely on complete and accurate information on the voltage amplitude and phase angle of the entire network bus. However, communication failures and sensor damage are common in the field SCADA and PMU acquisition channels, resulting in incomplete coverage of measurement points across the entire network. It is difficult to reconstruct the operating conditions of the entire power grid based solely on scattered measured data.

[0030] The currently common state estimation method in the industry is the weighted least squares algorithm, which is a purely physical model-driven solution. This algorithm builds an optimization model based on the nonlinear power flow equations of the power grid, treating all bus voltage amplitudes and phase angles as state variables to be solved. It incorporates measured active power, reactive power, and node voltage data collected from the field to construct measurement constraint equations. The optimization objective is to minimize the weighted sum of squared residuals between the measured values ​​and the model-calculated values. In the iterative solution phase, each round makes a linear approximation of the power flow equations, simultaneously calculating the Jacobian matrix and gain matrix. The state variable deviation is iteratively updated by solving the linear correction equations until the state difference between two consecutive rounds is less than a set threshold, at which point the estimated electrical state values ​​of all buses are output.

[0031] The estimation accuracy of this method relies entirely on sufficient redundant measurement points. This method compresses random measurement errors and locks in a unique feasible solution using multiple sets of mutually redundant measured data. However, if the proportion of missing power grid measurement points is too high, the effective observation constraints are significantly reduced, the observability of the power grid decreases, the interference effect of measurement noise is amplified, and the power flow equations are prone to underdetermined and ill-conditioned conditions. Ultimately, the deviations in the final output voltage and phase angle estimates increase sharply, failing to meet the requirements of power grid operation and management.

[0032] The method for determining the electrical state of the power grid will be explained in detail below with reference to the accompanying drawings. Figure 1 This application provides a flowchart illustrating a method for determining the electrical state of a power grid. Specifically, it includes the following steps: Step 110: Obtain the topology of the target power grid, the measurement data of each bus in the target power grid, and the sensitivity matrix corresponding to the power flow equation of the target power grid.

[0033] First, basic data collection and preprocessing are performed. The basic data in this application includes three categories: the topology of the target power grid, measurement data of each bus in the target power grid, and the sensitivity matrix corresponding to the power flow equations of the target power grid. The topology clearly defines the actual connectivity between all buses and electrical branches. The measurement data of each bus carries the measured information of the current operating state of the power grid. The sensitivity matrix characterizes the transmission and impact patterns of electrical disturbances between different buses. This step sequentially completes the acquisition, calculation, and consistency verification of the three types of data, as detailed below.

[0034] The topology of the target power grid is read from the dispatch system model library or standard power flow calculation files in BPA and PSASP formats. These files contain the unique bus numbers, rated voltage levels, component access attributes, and the starting and ending bus correspondences, resistance, reactance, ground susceptance, and rated capacity parameters of transmission lines, transformers, and other electrical branches. After reading the basic topology parameters, real-time switch remote signaling data from the dispatch system is accessed. Branch connectivity is adjusted based on the on / off status of circuit breakers and disconnectors. Branches with both ends of the switch closed retain valid electrical connections, while branches in open states are temporarily removed from the valid connection set. During the adjustment process, all bus nodes are retained, and only the connectivity attributes of branches are updated, ultimately forming the topology adjacency information for the corresponding operating mode. Taking a 1354-node area transmission network as an example, initially, 1354 bus parameters and 1872 electrical branch parameters are read. After real-time remote signaling status updates, 11 branches in open states are removed, completing the dynamic topology update.

[0035] Measurement data for each busbar are acquired via synchronous phasor measurement devices and a SCADA telemetry system. The data types include four categories of electrical quantities: active power injection, reactive power injection, voltage amplitude, and voltage phase angle. For buses with measurement coverage, the 3σ criterion is first used to filter out out-of-limit abrupt changes, eliminating invalid measurements exceeding power quality standard limits. Then, Z-score standardization is used to eliminate dimensional differences between different electrical quantities. The standardization calculation formula is as follows: In the formula, x is the original measurement value. This represents the statistical average of the corresponding electrical quantities across the entire network. This represents the statistical standard deviation of the electrical quantities across the entire network. The processed measurements are then concatenated one by one along each bus to form the initial measurement vector corresponding to a single bus. For buses experiencing communication interruptions, sensor failures, or lacking measurement devices, all measurement fields are set to zero, without removing the corresponding bus nodes, ensuring the total scale of the measurement data remains consistent with the total number of buses in the topology.

[0036] The sensitivity matrix is ​​derived from the power balance relationship of the power grid. Based on Kirchhoff's current law and voltage law, and combined with the impedance and ground susceptance parameters of transmission lines and transformers, the active and reactive power balance equations for all buses in the entire network are written, with the voltage amplitude and phase angle of each bus as the variables to be solved. This set of power balance equations constitutes the power flow equations. The power flow equations are a nonlinear set of equations. During the calculation, the bus measurement data after cleaning at the current moment is used as the initial operating condition, and the Newton-Raphson iterative power flow solution is performed. After the iteration converges, the reference point for stable operation of the power grid is obtained. At this reference point, a first-order Taylor linearization expansion is performed on the entire power balance equation set. For each set of balance equations, the partial derivatives corresponding to all bus state variables are calculated. All partial derivatives are then regularized and combined according to the bus number to form a sensitivity matrix, the expression of which is as follows: In the formula, The power flow equations are used to determine the power balance of the entire network. The bus voltage state vector corresponding to the steady-state operating reference point, and the elements inside the matrix. This characterizes the impact of small voltage fluctuations on the power balance equation of the i-th bus. From the perspective of power grid power transmission mechanisms, the electrical coupling between any two buses exhibits bidirectional equivalence, thus the derived sensitivity matrix synchronously satisfies symmetry. Furthermore, electrical coupling within the power grid is concentrated only between spatially adjacent buses, resulting in a matrix with a significantly low-rank characteristic. Higher-order singular values ​​approach zero, with only the first few singular values ​​carrying the core electrical coupling information of the power grid. After matrix calculation, a sparse matrix format is used for storage, retaining only the positions and values ​​of non-zero elements to reduce hardware storage requirements.

[0037] After all three types of data have been processed, a consistency check is performed to verify that the total number of topology buses, the number of rows of measurement data, and the order of the row and column of the sensitivity matrix are completely equal. The bus numbering mapping rules within all data are unified to ensure that the same number corresponds to the same physical bus.

[0038] Step 120: Construct a graphical model of the target power grid based on the topology.

[0039] The power transmission network formed by buses, transmission lines, and transformers naturally possesses graph structure characteristics. Voltage and power disturbances are transmitted between buses solely through physical electrical branches; buses without physical connections cannot generate direct electrical coupling. Transforming the power grid topology into a standard graph model transforms the complex power grid entities into a unified and standardized mathematical expression, clarifies the connectivity relationships between buses within the network, delineates the interaction range of nodes based on physical branches, ensures that all calculations conform to the actual physical connection patterns of the power grid, and eliminates invalid correlation calculations between buses without electrical connections. This step utilizes the complete topology parameters output from step 110, corrected for switch states, to complete the construction of the undirected graph model. The complete operation process is as follows.

[0040] Extract the unique IDs of all busbars from the topology parameter records. Each busbar corresponds to an independent node in the graph model. All busbar nodes are integrated to form the node set V of the graph model. In step 110, only buses without measurement points (those with zeroed measurement data and no deletion process) are retained in the node set. The total number of nodes is consistent with the dimension of the measurement matrix and sensitivity matrix output in step 110, ensuring one-to-one matching of busbar IDs for all types of data and eliminating dimension mismatch issues.

[0041] Traversing all electrical branches recorded in the topology parameters, only transmission lines and transformers with both ends of the switch kept closed are mapped as undirected connecting edges within the graph model. The undirected edges corresponding to all valid branches form the edge set E of the graph model. Power grid transmission branches exhibit bidirectional power transmission characteristics, and disturbances at both ends of the same branch have equivalence. The undirected graph structure can match this bidirectional symmetrical electrical coupling law. Branches in the open state do not generate corresponding connecting edges; only the direct connection between nodes is disconnected, and the node set is not modified in any way.

[0042] After completing the one-to-one mapping of nodes and edges, a corresponding adjacency index table is generated. The index table records the numbers of all neighboring nodes directly connected to each node in the order of bus numbering. The adjacency index table is bound to the graph model for storage. The index table limits the interaction objects between nodes, and only buses connected by physical branches can participate in the same group of operations, so that all calculation logic conforms to the actual connection state of the power grid.

[0043] Taking the 1354-node regional power transmission network as an example, the modeling is completed based on the topology parameters corrected in step 110, resulting in an undirected graph model containing 1354 nodes and 1861 undirected edges, and a 1354-order adjacency index table is generated simultaneously.

[0044] Step 130: Generate corresponding initial features for each node in the graph model based on the measurement data of each bus.

[0045] The bus measurement data output in step 110 includes the measured electrical information of all buses in the entire network. The data corresponds one-to-one with the bus number. All nodes in the graphical model are mapped to buses, and their numbering systems are completely unified. Based on the measurement data, unique initial features are generated for each node, which can transform the actual operating conditions of the power grid into standardized vectors that can be used for numerical calculations. This completes the conversion of physical quantities into model input vectors and assigns calculable operating status information to each node in the graphical model. The complete processing procedure is as follows.

[0046] Retrieve the bus measurement dataset after cleaning and standardization in step 110. Each data set in this dataset corresponds to one bus, and the data includes four types of values: normalized active power injection, reactive power injection, voltage amplitude, and voltage phase angle. The node numbers in the graphical model are perfectly aligned with the bus numbers in the measurement data, and the corresponding measurement information is matched one by one in numerical order.

[0047] For any node in the graph model, extract all standardized measurement values ​​of the corresponding bus. Concatenate the values ​​of various electrical quantities into a one-dimensional vector in a fixed order. This one-dimensional vector is the initial feature corresponding to the node. For nodes corresponding to buses with no valid measurements or all measurement fields set to zero, all elements in the concatenated initial feature vector are 0, and the nodes are not removed to ensure that the number of nodes in the graph model and the total number of initial feature vector sets remain consistent.

[0048] The initial feature vectors of all nodes are integrated to form a feature matrix. The number of rows in the matrix is ​​equal to the total number of nodes in the graph model, and each row of the matrix corresponds to the initial feature vector of one node. This feature matrix is ​​stored in conjunction with the graph model and the adjacency index table, and the row indices within the matrix correspond one-to-one with the node numbers.

[0049] Taking the 1354-node regional power transmission network as an example, the graph model has a total of 1354 nodes. It matches the 1354 sets of standardized measurement data output in step 110. Each set of measurement contains 4 types of electrical quantities, and finally generates an initial feature matrix with a dimension of 1354×4.

[0050] Step 140: Determine the electrical coupling strength between different buses based on the sensitivity matrix, and perform weighted fusion processing on the initial features of all buses one by one according to the electrical coupling strength to obtain the node fusion features corresponding to all buses.

[0051] The initial node features generated in step 130 can only carry the measured operating information of a single busbar, and there is no correlation information between electrical couplings between buses within the feature vector. During the actual operation of the power grid, when the power and voltage of any busbar fluctuate, the disturbance will be propagated to other buses along the electrical branches. A single node feature cannot reflect this cross-node interaction relationship. Directly using the initial features for calculation will ignore the inherent coupling law of the power grid, causing the state calculation results to deviate from the actual operating conditions.

[0052] The sensitivity matrix is ​​obtained by linearizing the power flow equations of the entire network. The values ​​within the matrix can quantify the strength of the disturbance interaction between any two buses. Using this as a quantification benchmark to calculate the correlation weights between buses allows the weighting rules for feature fusion to align with the actual physical mechanisms of the power system. Based on this, this application performs a weighted aggregation operation on the initial features of all nodes using the correlation weights. This integrates the operating information of electrically coupled buses across the entire network into the representation of a single node. The fused features obtained after aggregation for each node retain the operating information carried by its own original bus measurements while incorporating the electrical disturbance effects from other buses in the entire network. This compensates for the insufficiency of the initial features, which only possess single-point local information, thereby obtaining node fusion features that integrate global electrical coupling relationships. Figure 2 This paper illustrates a flowchart of a node fusion feature determination method provided in an embodiment of this application. The specific implementation process is as follows: Figure 2 As shown.

[0053] Step 210: Perform truncated singular value decomposition on the sensitivity matrix, and construct low-dimensional feature vectors corresponding to each bus based on the decomposition results.

[0054] The sensitivity matrix fully records the electrical coupling relationships between all pairs of buses in the entire power grid, with the matrix dimension consistent with the total number of buses in the grid. In large-scale power grid scenarios, the matrix data volume is enormous, and directly using the complete matrix for correlation weight calculation would consume a large amount of storage resources and prolong computation time. At the same time, the matrix has an approximately low-rank characteristic, with most electrical coupling information concentrated in only a small number of singular value components with high amplitudes, and the components corresponding to higher-order singular values ​​containing only weak and negligible coupling noise. Therefore, this application adopts a truncated singular value decomposition method to reduce the dimension of the matrix, discarding higher-order components carrying invalid noise, and generating low-dimensional feature vectors corresponding to each bus based on the retained core components, thereby reducing the data computation scale while preserving the key electrical coupling patterns of the power grid.

[0055] In one implementation, the sensitivity matrix is ​​subjected to complete singular value decomposition to obtain an orthogonal singular vector matrix and a set of singular values. Singular values ​​of the first preset order and orthogonal singular vectors corresponding to the preset order are selected from the set of singular values, and singular values ​​and orthogonal singular vectors corresponding to the remaining orders are removed. For each row of the orthogonal singular vector matrix corresponding to a single busbar, the orthogonal singular vectors of the first preset order corresponding to the row are extracted, and each orthogonal singular vector of each order is multiplied element-wise by the square root of the corresponding singular value to obtain the low-dimensional feature vector corresponding to the single busbar.

[0056] Specifically, retrieve the complete sensitivity matrix obtained in step 110. ,matrix for A square matrix, where n is the total number of power grid buses. For the matrix... Perform a complete singular value decomposition, the operation process is as follows.

[0057] The first step is to calculate the square matrix. With its own transpose matrix product matrix The first step is to find all the eigenvalues ​​of the product matrix. Taking the square root of each eigenvalue yields a set of values, which are the singular values. The second step is to sort all the singular values ​​in descending order of value. Then, fill the diagonal positions of the diagonal matrix with these sorted singular values, filling the remaining positions with 0, resulting in a diagonal matrix of singular values. The third step is to solve for the product matrix. The eigenvectors corresponding to all eigenvalues ​​are arranged as column vectors of a matrix in the order of their singular values, resulting in an orthogonal singular vector matrix U, where any two column vectors are mutually orthogonal. The entire decomposition operation satisfies the identity: The decomposition operation outputs two sets of results, namely, orthogonal singular vector matrices. A diagonal matrix composed of singular value sets The dimension of the orthogonal singular vector matrix is... All column vectors are pairwise orthogonal. Each row of the matrix corresponds to a bus, and the row vector records the response characteristics of that bus under various electrical coupling modes across the entire network. (Singular value diagonal matrix) The dimension is Only the diagonal has non-zero singular values, while all other positions are 0. The larger the singular value, the stronger the influence of the corresponding electrical coupling mode on the power grid and voltage disturbance. The extremely small high-order singular values ​​only correspond to measurement noise and have no actual physical meaning. Representative matrix The transpose of .

[0058] After completing the full singular value decomposition and obtaining the sorted singular value diagonal matrix After the orthogonal singular vector matrix U, to eliminate high-order meaningless noise components and reduce subsequent computational and storage overhead, a fixed truncation order k is pre-configured. In this embodiment, k is set to 64 to adapt to the computational requirements of a thousand-node-level transmission network. The first k order singular values ​​are truncated in descending order of amplitude, and a truncated singular value diagonal matrix is ​​constructed from these singular values. Simultaneously extract the first k columns of the orthogonal singular vector matrix U to obtain the truncated orthogonal singular vector matrix. Singular values ​​whose order exceeds k and their corresponding orthogonal singular vectors are all discarded, thus completing matrix truncation and noise reduction. The approximate relationship of the truncated matrix satisfies: In the formula Store the first k sets of orthogonal singular vectors. Let be a diagonal matrix consisting of the first k non-zero singular values.

[0059] Truncation of orthogonal singular vector matrix Each row uniquely corresponds to a bus in the power grid. Row vectors matching a single bus are extracted row by row. Find the square root of each singular value in the truncated singular value diagonal matrix. The fusion is achieved by multiplying each orthogonal singular vector of order within the row vector with the square root of its corresponding singular value, resulting in a low-dimensional eigenvector of a single generatrix. The calculation expression is: In the formula, To truncate orthogonal singular vector matrices The i-th row, It is a diagonal matrix formed by taking the square root of each diagonal element of a singular value diagonal matrix.

[0060] The process iterates through all the matrix rows corresponding to all buses in the graph model, repeatedly performing vector extraction and vector-single-value square root fusion operations to generate k-dimensional low-dimensional feature vectors for all buses in the network. These low-dimensional feature vectors are then aggregated to form a low-dimensional feature vector matrix. The number of rows in the matrix is ​​consistent with the total number of nodes in the graph model, and the number of columns in the matrix is ​​equal to the preset order k.

[0061] Taking a 1354-node regional power transmission network as an example, the sensitivity matrix A has a dimension of 1354×1354. With a truncation order of k=64, after complete singular value decomposition and truncation, each bus can generate a 64-dimensional low-dimensional feature vector, ultimately yielding a feature vector of dimension [missing information]. low-dimensional eigenvector matrix .

[0062] Step 220: Calculate the correlation weight between any two buses based on the low-dimensional feature vectors of each bus.

[0063] The low-dimensional feature vectors of each bus obtained through step 210 have compressed the global electrical coupling law of the power grid into a unified dimension vector. However, the low-dimensional feature vector is only a carrier of the bus coupling characteristics and cannot directly reflect the contribution ratio of each associated bus when aggregating the features of neighboring nodes of a single bus. Furthermore, the original numerical values ​​of coupling obtained by vector operations for different bus pairs are not uniform and lack horizontal comparability, and cannot be directly used for weighted fusion calculation of node features. Therefore, it is necessary to further quantify and normalize the low-dimensional feature vector to generate standardized association weights, so as to accurately allocate the calculation weight of each associated bus in the feature aggregation process and conform to the actual electrical coupling strength relationship of the power grid.

[0064] In one implementation, the low-dimensional feature vectors corresponding to any two buses are selected, and the values ​​corresponding to the two sets of low-dimensional feature vectors are multiplied one by one and then all the values ​​are added together to obtain the original interaction values ​​between the two buses. The original interaction values ​​are subjected to exponential transformation to obtain a single set of transformed values. Taking one of the two buses as the reference bus, the transformed values ​​corresponding to all buses that are connected to the reference bus by physical electrical branches are accumulated to obtain a summation result. Based on the summation result, the single set of transformed values ​​is normalized to obtain the correlation weight between any two buses.

[0065] Specifically, the low-dimensional feature vector matrix K output in step 210 is retrieved, and the vector in the i-th row of the matrix is... Let be the low-dimensional feature vector of the i-th busbar, and be the vector of the j-th row. Let j be the low-dimensional eigenvector of the j-th busbar. , The dimensions are all equal to the truncation order k set in step 210. and After multiplying element by element and summing all elements, the vector inner product operation is completed, yielding the original interaction values ​​between bus i and bus j. The calculation formula is as follows: The original value can characterize the degree of coordination between the two bus dominant electrical modes, but the value has no fixed range and cannot be directly used as a weighting coefficient to participate in feature aggregation.

[0066] To distinguish the numerical differences between strong and weak couplings, an order scaling term is introduced to perform an exponential transformation on the original interaction values, resulting in transformed values. The operational expression is as follows: In the formula, k is the truncation order preset in step 210. This is used to balance the numerical offset caused by the vector dimension, and to prevent the numerical range of the exponential output from getting out of control.

[0067] Using bus i as the reference bus, traverse all bus m in the diagram model that are physically and electrically connected to bus i, and calculate the conversion values ​​between the reference bus i and bus m one by one. Summing up all transformed values ​​yields the summation term. ,in This represents the set of neighboring buses of the baseline bus i. The summation term is used to... Perform normalization to obtain standardized association weights. The calculation formula is as follows: After normalization, the values ​​of all associated weights corresponding to the same baseline bus are in the range of 0 to 1, and the sum of the weights of all neighboring buses under the same baseline bus is 1, which can be directly used for weighted aggregation of node features.

[0068] Since the dot product of vectors satisfies The final weights obtained naturally satisfy the following conditions. This aligns with the reciprocal electrical characteristics of undirected transmission lines in the power grid. By iteratively replacing all buses in the entire network as the reference bus, and repeatedly performing the entire set of operations—vector inner product, exponential transformation, and normalization—the pairwise correlation weights between all buses are obtained in batches.

[0069] Step 230: Based on the association weight, perform weighted fusion processing on the initial features of all buses one by one to obtain the node fusion features corresponding to all buses.

[0070] First, retrieve the initial feature vectors of all nodes output in step 130. Let the target bus be the i-th bus, and its initial feature vector is... .definition Given the set of all adjacent buses that are physically connected to the target bus i, for any adjacent bus... Retrieve the initial characteristics of bus j and the association weights calculated in step 220 .

[0071] The weighted feature components are obtained by multiplying the initial features of adjacent buses by their corresponding weights. traversal The node fusion features of the target bus i are obtained by summing the weighted feature components of all adjacent buses along each dimension. Calculation formula: , Maintaining the same dimension as the initial feature vector, adjacent buses with higher coupling strength have larger weights, and their measurement information accounts for a higher proportion in the fused features. This matches the reciprocity characteristics of bidirectional energy flow in undirected transmission lines.

[0072] Traverse all buses in the power grid, take each bus as the target bus in turn, repeat the neighbor feature retrieval, weighting, and aggregation operations, generate node fusion features corresponding to all buses in batches, and summarize them to form a fusion feature matrix. The number of rows in the matrix is ​​equal to the total number of nodes in the graph model, and the number of columns in the matrix is ​​consistent with the initial feature dimension.

[0073] Step 150: Combine the node connection relationships of the graph model to perform multi-layer feature transformation on the node fusion features of all buses to obtain the electrical state estimates of all buses.

[0074] To fully explore the multi-level and cross-regional electrical coupling relationships in the power grid and improve the accuracy and robustness of electrical state recovery, this application, based on the node fusion features obtained in step 140, performs multi-level feature transformation processing in conjunction with the graph model topology. The final output of this multi-level feature transformation is the electrical state estimate, which is the estimated voltage amplitude and voltage phase angle of each bus in the power grid. This multi-level feature transformation processing consists of multiple structurally identical basic transformation units stacked in series. Each basic transformation unit follows the same operational logic and execution flow, achieving deep feature extraction and abstraction through layer-by-layer iteration. Figure 3 The diagram shows a flowchart of a single-layer feature transformation method provided in an embodiment of this application. The specific execution process of a single basic transformation unit (single layer) is described below.

[0075] Step 310: Use a preset weight matrix to perform a linear transformation on the node fusion features of each node to generate a linear transformation feature vector corresponding to each node.

[0076] This step, as the first computational stage of single-layer feature transformation, performs dimensionality mapping and feature recombination on the node fusion features through a learnable weight matrix. To avoid mutual interference between features with different physical attributes during transformation and to suppress the feature homogenization trend in deep networks, this step adopts a block-decoupled orthogonal composite linear transformation architecture. It explicitly splits the steady-state and perturbation components according to the physical semantics of the features, and performs independent projection, distance-preserving rotation, and then performs weighted fusion.

[0077] Specifically, this step performs linear mapping on the fused features of each node through a preset weight matrix, completing dimensional adaptation and feature reorganization. To reduce mutual interference caused by the mixed transformation of features with different physical attributes and improve the effectiveness of feature representation, the preset weight matrix in this application adopts a block-decoupled composite parameterized structure, and the overall operation is carried out in three stages: feature splitting, block projection, and cross-subspace fusion.

[0078] The node fusion features are output from step 230, maintaining the same dimensions as the initial features. Internally, they contain feature components derived from four physical quantities: active power injection, reactive power injection, voltage amplitude, and voltage phase angle. Active power injection and voltage amplitude primarily reflect the steady-state operating benchmark of the power grid, while reactive power injection and voltage phase angle reflect dynamic disturbance fluctuations. These two types of components differ in their physical properties and fluctuation scales. A uniform linear transformation could easily lead to feature information aliasing. Therefore, dimensional boundaries are fixed according to physical properties. The feature dimensions corresponding to active power injection and voltage amplitude are classified as steady-state fundamental components, while the feature dimensions corresponding to reactive power injection and voltage phase angle are classified as dynamic disturbance components. Taking the 4-dimensional fusion features as an example, the first and third dimensions are used to form the steady-state component sub-vector, and the second and fourth dimensions are used to form the disturbance component sub-vector. After splitting, the two sub-vectors do not overlap and are processed through independent projection channels.

[0079] Each of the two sets of sub-vectors is configured with an independent block projection sub-matrix. The parameters of the two sets of block projection sub-matrixes are independent of each other, and their gradients do not interfere with each other during training. Each completes the dimensionality upscaling mapping of its corresponding physical component. The shape of a single projection sub-matrix is ​​determined by the dimension of the input sub-vector and the preset hidden dimension. The number of rows in the matrix equals the dimension of the corresponding sub-vector, and the number of columns in the matrix equals the preset single-path subspace hidden dimension. For example, if the steady-state sub-vector dimension is 2 and the single-path subspace hidden dimension is set to 16, then the shape of the steady-state projection sub-matrix is ​​2 rows and 16 columns; if the perturbation sub-vector dimension is 2, the corresponding perturbation projection sub-matrix is ​​also 2 rows and 16 columns.

[0080] The initialization of the projection submatrix is ​​completed using a physics-guided prior approach. From the truncated singular vectors output in step 210, the top few singular vectors with the highest coupling degree to two types of steady-state quantities—active power injection and voltage amplitude—are selected and normalized to serve as the initial column vector basis for the steady-state projection submatrix. Similarly, the top few singular vectors with the highest coupling degree to two types of disturbance quantities—reactive power injection and voltage phase angle—are selected as the initial column vector basis for the disturbance projection submatrix. The remaining column vectors outside the basis are filled using a standard orthogonal initialization method to ensure that the projection directions are non-redundant in the initial stage. Through this initialization method, the initial principal direction of the projection submatrix aligns with the inherent mode direction of the electrical coupling of the power grid, avoiding training bias caused by random initialization and accelerating the model convergence speed.

[0081] After the two sets of subvectors are independently projected, they are linearly integrated across subspaces using a fusion weight matrix. The first step involves feature concatenation, where the two sets of projected features are concatenated end-to-end along their feature dimensions in a fixed order, forming a long vector containing both types of physical information: steady-state projected features are arranged in the first half, and perturbation projected features in the second half. The total dimension of the concatenated vector is the sum of the dimensions of the two subspaces. For example, if both sets of projected features are 16-dimensional, the concatenated vector will be 32-dimensional, with the first 16 dimensions corresponding to the steady-state component mapping and the last 16 dimensions corresponding to the perturbation component mapping. The second step involves linear weighted fusion, where the row and column dimensions of the fusion weight matrix match the dimensions of the concatenated long vector. For example, if the concatenated vector is 32-dimensional, the fusion weight matrix will be 32 rows and 32 columns. During computation, the concatenated feature vector is multiplied by the fusion weight matrix. Each dimension of the output feature is obtained by weighted summation of all dimensions in the concatenated vector according to their corresponding matrix weights. By utilizing learnable parameters of the weight matrix, the contribution ratios of steady-state and perturbation components in the final output features can be automatically adjusted, achieving an organic integration of the two types of physical information rather than a simple dimensional concatenation. After matrix multiplication, a global bias vector with the same output dimension is superimposed, and the features are numerically shifted globally to obtain a linearly transformed feature vector. The fusion weight matrix is ​​initialized using a Xavier uniform distribution, and the parameters are updated synchronously during training along with backpropagation of the loss function.

[0082] Based on the above structure, for any target bus node i within the graph model, its corresponding node fusion feature vector is first extracted. Steady-state component sub-vectors are obtained by splitting according to the aforementioned dimensional boundaries. With perturbation component subvector Then, the two sub-vectors are multiplied by their corresponding projection sub-matrices to achieve dimensionality increase, yielding the projected steady-state features. With disturbance characteristics Based on this, the two sets of projected features are concatenated end-to-end along the feature dimension to form a concatenated feature vector with a total dimension equal to the sum of the two. Then, the concatenated feature vector is multiplied by the fusion weight matrix, and the global bias vector is superimposed to obtain the linear transformation feature vector corresponding to that node. The overall operation expression is: In the formula, , It is a two-part block projection submatrix. This represents a vector concatenation operation along the feature dimension. To fuse the weight matrix, This is the global bias vector.

[0083] Traverse all bus nodes within the graph model, and sequentially perform dimensionality splitting, block projection, stitching and fusion, and bias overlay operations according to the same rules to generate linear transformation features for all nodes in batches. All linear transformation features are summarized to form a linear feature matrix, which has the same number of rows as the total number of nodes in the graph model and the same number of columns as the output feature dimension set for this layer.

[0084] Step 320: Perform nonlinear mapping processing on each element in the linear transformation feature vector to obtain the output features of the single-layer feature transformation.

[0085] After the linear transformation, the uniform nonlinear activation method cannot adapt to the differences in numerical distribution and physical properties of steady-state derived and perturbation derived features, which can easily lead to impaired smoothness of steady-state operating features or decreased recognizability of perturbation fluctuation features. To address this, this step adopts a dual-modal adaptive piecewise nonlinear mapping mechanism, performing differentiated element-by-element nonlinear transformations grouped according to the feature source dimension, while embedding soft constraints of physical boundaries. This introduces nonlinear expressive power while maintaining the physical rationality of the features. The specific implementation method is as follows.

[0086] The linear transformation feature vector output in step 310 can be divided into a steady-state derived feature group and a perturbation derived feature group according to their source. The two groups of dimensions are arranged continuously along the vector order, and the boundary position corresponds to the proportion of the subspace dimensions before fusion. Taking a 32-dimensional linear transformation feature as an example, the first 16 dimensions are the steady-state feature group derived by projection fusion of the steady-state fundamental component, and the last 16 dimensions are the perturbation feature group derived by projection fusion of the dynamic perturbation component. The two groups are configured with independent learnable nonlinear parameter sets, and the parameters are updated synchronously with the model training process without interfering with each other.

[0087] For the steady-state derived feature set, a smooth saturated nonlinear mapping with learnable operating points is applied element-wise. Each dimension is independently configured with a set of learnable parameters, including a gain coefficient, a threshold offset, and a saturation coefficient. During mapping, the linear transformation value of that dimension is first multiplied by the gain coefficient, then the threshold offset is subtracted to center the value, and then the smooth saturated activation function is input for nonlinear transformation. Finally, the nonlinear output of that dimension is obtained by scaling with the saturation coefficient. The mapping function expression is: In the formula, Let be the input value of the j-th dimension of the steady-state group in the linear transformation feature of node i. This is the gain coefficient for this dimension. This is the threshold offset. The saturation coefficient is... This represents the nonlinear output value corresponding to this dimension. The function possesses global differentiability and one-sided smooth saturation characteristics, which can conform to the physical laws of the finite and gradual changes in the steady-state quantity of the power grid, avoiding meaningless and drastic fluctuations in the steady-state characteristics. During parameter initialization, the gain coefficient is initialized to 1, the threshold offset is initialized to the standardized nominal value of the corresponding steady-state physical quantity, and the saturation coefficient is initialized to the normal fluctuation amplitude of the corresponding physical quantity, making the initial nonlinear curve conform to the typical operating range of the power grid.

[0088] For the perturbation-derived feature set, an adaptive piecewise dead-zone nonlinear mapping is applied element-wise. Each dimension is independently configured with three sets of learnable parameters: dead-zone threshold, linear gain, and saturation upper limit. The mapping rule is divided into three segments: when the absolute value of the input value is less than the dead-zone threshold, the output is proportionally attenuated to near zero to suppress interference from small-amplitude measurement noise; when the absolute value of the input value is between the dead-zone threshold and the saturation upper limit, it is proportionally amplified by the linear gain to retain effective perturbation features; when the absolute value of the input value exceeds the saturation upper limit, it enters the smooth saturation region to limit the output amplitude of extreme outliers. To ensure that the function is differentiable everywhere, a smooth transition function is used at the segmentation points. The overall mapping expression is: In the formula, Let be the input value of the j-th dimension of the perturbation group in the linear transformation feature of node i. This is the dead zone threshold for this dimension. For linear gain, The upper limit of saturation, To prevent division by zero of extremely small constants, This represents the nonlinear output value corresponding to this dimension. This mapping can automatically filter out small fluctuations with amplitudes below the noise level, amplify effective disturbance signals, and limit the range of extreme values, adapting to the distribution characteristics of power grid disturbance signals. During parameter initialization, the dead zone threshold is initialized to the typical noise amplitude of the corresponding measurement type, the linear gain is initialized to 1, and the saturation upper limit is initialized to the fault-level fluctuation amplitude of the corresponding physical quantity.

[0089] After completing the element-wise nonlinear mapping of the two sets of features, the steady-state output group and the disturbance output group are concatenated in their original dimensional order to obtain the preliminary nonlinear feature vector. Based on this, element-wise physical soft truncation is performed. According to the reasonable operating range of the power grid electrical quantities, corresponding physical upper and lower limit reference values ​​are set for each feature dimension. When the feature value is within the upper and lower limit range, the output remains unchanged; when the value exceeds the range, the hyperbolic tangent function is used to smoothly compress the value to near the range boundary, thus limiting unreasonable extreme values ​​while ensuring the function is differentiable throughout, without affecting backpropagation training. The soft truncation expression is: In the formula, The initial nonlinear characteristic is represented by the j-th dimension value. , These are the physical upper and lower limit reference values ​​for this dimension, respectively. This is the final output value after soft truncation. Traverse all bus nodes within the graph model, and sequentially perform dimensional grouping, bimodal element-wise nonlinear mapping, concatenation, and physical soft truncation operations on the linear transformation feature vector of each node according to the same rules, generating single-layer output features for all nodes in batches. All output features are summarized to form a single-layer output feature matrix, with the number of rows matching the total number of nodes in the graph model and the number of columns matching the linear transformation feature dimensions. After generation, this matrix is ​​passed down to subsequent residual stacking or the next layer of transformation.

[0090] Step 330: Repeat the single-layer feature transformation and concatenate them layer by layer. Each layer uses the output features of the single-layer feature transformation of the previous layer as the input features of the current layer, thereby realizing multi-layer feature transformation.

[0091] After the above single-layer linear transformation and nonlinear mapping, the node features only complete the information recombination and nonlinear expression at a single level, which is insufficient to fully characterize the complex electrical coupling relationships of the power grid across multiple levels and regions. Therefore, the features output from the single-layer feature transformation are used as the input for the next layer of feature transformation. The complete operation process of block decoupling linear transformation and dual-modal nonlinear mapping is repeatedly executed to progressively extract deeper levels of electrical coupling features.

[0092] The total number of network layers is pre-set based on the power grid scale and task complexity. For a thousand-node-level transmission grid, it can be set to 8 layers. Each layer is independently configured with a corresponding block projection sub-matrix, fusion weight matrix, and nonlinear parameter set. The parameters of each layer are updated independently during the training process. The single-layer output feature matrix of the previous layer is directly used as the input feature matrix of the linear transformation of the next layer. The feature dimension remains consistent during the multi-layer transmission process. Shallow networks focus on extracting electrical coupling details in the local neighborhood, while deep networks focus on integrating global operational correlation information across regions.

[0093] After all preset feature transformation layers are completed, the final deep feature vectors corresponding to each node in the entire network are obtained. These features reside in a high-dimensional hidden space and cannot directly correspond to the physical operating quantities of the power grid. Therefore, an end-point output projection layer is configured to complete the mapping from the feature space to the physical space. The output projection layer adopts a dual-channel decoupled mapping structure, corresponding to two types of state quantities to be recovered: voltage amplitude and voltage phase angle. The two sets of channel parameters are trained independently, and the mapping logic is independent of each other, avoiding the accuracy loss caused by sharing mapping parameters for physical quantities with different dimensions and fluctuation ranges. The output projection layer is configured with two independent output weight matrices and corresponding bias vectors. The number of rows in a single output weight matrix is ​​consistent with the dimension of the final deep feature, and the number of columns is 1. For example, if the dimension of the deep feature is 32, then the dimension of both output weight matrices is 32×1, and the dimension of the corresponding bias vector is 1.

[0094] For any bus node i within the graphical model, its corresponding final deep feature vector is extracted. This vector is then input into two channels, multiplied by the voltage amplitude output weight matrix, and superimposed with the amplitude channel bias to obtain the original estimated value of the bus's voltage amplitude. Similarly, multiplying this vector by the voltage phase angle output weight matrix and superimposing with the phase angle channel bias yields the original estimated value of the bus's voltage phase angle.

[0095] Physical boundary soft constraint processing is configured at the ends of the two output channels. The voltage amplitude channel sets upper and lower limits based on the per-unit value range of normal grid operation, and uses a smooth saturation function to constrain the amplitude estimate to a reasonable fluctuation range to avoid abnormal values ​​that deviate from the actual operating range; the voltage phase angle channel performs periodic mapping processing based on the phase angle periodicity to normalize the phase angle estimate to the standard value range and eliminate numerical ambiguity caused by phase periodicity.

[0096] Traverse all bus nodes in the graph model, and sequentially perform dual-channel projection mapping and physical boundary constraints according to the bus number order. Arrange the voltage amplitude estimate and voltage phase angle estimate corresponding to each bus one by one, and finally form an electrical state estimate set that corresponds one-to-one with the entire network bus, that is, the electrical state estimate value corresponding to all buses.

[0097] While the stacked structure of multi-layer feature transformations can progressively extract deeper levels of electrical coupling representation, the original measurement information carried by the initial input gradually decays with each layer as the number of layers increases. Furthermore, gradient vanishing during backpropagation training can lead to increased training difficulty for deep networks and hinder the stable transfer of effective information from shallow layers to deeper layers. To alleviate these problems, this application simultaneously establishes a two-level identity transfer mechanism: intra-layer and block-level. This mechanism combines short and long paths to ensure the effective flow of original information within deep networks. Specifically, within a single-layer feature transformation, the input features are directly transferred to the output of the single-layer feature transformation, where they are superimposed to form a local identity transfer path. Furthermore, feature transformations of a predetermined number of consecutive layers are divided into a set of feature transformations. The input features of the first layer within this set are directly transferred to the output of the last layer within the set, where they are superimposed to form a block-level identity anchor.

[0098] Specifically, for the local identity transfer path, the input features of a single-layer feature transformation have the same dimensions as the output features obtained after linear transformation and nonlinear mapping, thus possessing the operational conditions for element-wise superposition. For each layer of feature transformation, an independent identity branch is set within the layer. This branch does not perform any transformation operations on the input features, but directly passes the complete input feature vector to the output of the current layer. After the current layer completes the linear transformation and nonlinear mapping operations and obtains the transformed output features, the original input feature vector and the transformed output feature vector are added element-wise according to their corresponding dimensions. The sum is used as the final output feature of the single-layer feature transformation, thus forming the local identity transfer path within the layer. Taking the l-th layer feature transformation as an example, its input features are denoted as... The transformation result obtained by the linear and nonlinear transformations of this layer is denoted as Then the final output feature of this layer satisfy: This path allows gradients to be directly backpropagated to preceding network layers via the identity branch, alleviating the gradient vanishing problem in deep network training. At the same time, the measurement features carried by the original input can be continuously retained during layer-by-layer propagation, avoiding complete dilution by multiple transformations.

[0099] For block-level identity anchors, a pre-defined number of layers, *s*, is set for each feature transformation set. All feature transformation layers are divided into sets of *s* layers each, representing a residual block. These residual blocks are then concatenated, with the total number of blocks determined by the ratio of the total number of feature transformation layers to the number of layers in a single block. For each residual block, the input features of the first feature transformation layer within that block are extracted as the block-level baseline input. This baseline input bypasses any transformation operation within the block and is directly passed along an independent long path to the output of the last feature transformation layer. After the last layer of the block completes local identity superposition and obtains its single-layer output, the block-level baseline input features and the output features of that layer are added element-wise along their corresponding dimensions. The sum is used as the final output feature of the entire residual block, thus forming the block-level identity anchor. Throughout the multi-layer feature transformation process, the dimensions of the input and output features of each layer remain consistent. The dimensions of the block-level baseline input and the output of the last layer of the block are perfectly matched, eliminating the need for additional dimension mapping operations; element-wise superposition is performed directly.

[0100] Taking a 1354-node regional power transmission network as an example, the total number of feature transformation layers is set to 8, with a preset single-block layer number s=4, thus dividing the network into two sequentially connected residual blocks. Each layer has 32-dimensional input and output features. The local identity path within a layer directly adds the 32-dimensional input features to the transformed 32-dimensional output features element-wise, resulting in a 32-dimensional output. Within each residual block, the 32-dimensional input features of layer 1 are directly passed to the output of layer 4, where they are superimposed element-wise with the 32-dimensional output features of layer 4, becoming the output feature of that residual block, which is then passed to the next residual block for further computation. This combination of two-level identity propagation paths ensures information flow and gradient transfer between adjacent layers while also preserving original information across multiple layers, enabling stable training of deep networks and balancing the representation of shallow local electrical features with deep global coupling features.

[0101] While the two-level identity propagation path can effectively ensure the cross-layer flow of original information and alleviate gradient vanishing, the learnable weight matrix of the linear transformation may still gradually cause distortion of the feature space structure during iterative training if there is a lack of geometric constraints. This leads to a continuous narrowing of feature differences between different nodes as the number of layers increases, weakening the discriminative ability of deep features. To further maintain the stability of the intrinsic geometric structure of the feature space and ensure that the relative distance relationship between node features before and after the linear transformation is not distorted, this application introduces orthogonalization constraint rules to normalize and control the weight matrix of each layer.

[0102] In one implementation, the preset weight matrices corresponding to each layer of linear transformation are orthogonally constrained based on preset constraint rules to generate corresponding correction offsets. The preset constraint rule is that the difference between any two sets of node features before and after the linear transformation calculation of the same weight matrix remains equidistant. Based on the correction offsets and the deviation between the node features output by the multi-layer feature transformation and the actual electrical measurement data, the preset weight matrices corresponding to each layer of linear transformation are adjusted so that the linear transformation calculation of each layer meets the orthogonal equidistant transformation requirements.

[0103] Specifically, the preset constraint rules use row orthogonality as the basis for quantification. For any set of linear transformation weight matrices, the product of itself and its transpose is constrained to approximate an identity matrix of the same order. This ensures that the difference in the magnitude of any two sets of node features remains equidistant before and after the linear transformation calculation of the same weight matrix. Since node features participate in the calculation in the form of row vectors, for any two sets of node feature row vectors... , After right multiplication by the weight matrix W, it always satisfies The relative geometric structure of the feature space does not undergo distortion.

[0104] The preset weight matrix corresponding to each layer of linear transformation includes three sets: a steady-state block projection sub-matrix, a perturbation block projection sub-matrix, and a fused weight matrix. These three sets of matrices undergo orthogonal constraint processing independently, each generating its corresponding correction offset. For a single set of weight matrices... First, calculate its orthogonal deviation matrix: multiply the weight matrix by its transpose, and subtract the identity matrix of the corresponding order to obtain the deviation matrix. ,in Let be the identity matrix of the same order as the number of rows in the weight matrix. The squared Frobenius norm of the bias matrix is ​​used as the orthogonal constraint loss term, expressed as: The value of this loss term directly reflects the degree to which the current weight matrix deviates from the isometric transformation condition. The orthogonal constraint loss term with respect to the weight matrix... The gradient vector obtained by taking the derivative is the correction offset corresponding to the weight matrix. This offset represents the direction and magnitude of the correction of the weight matrix to the orthogonal equidistant state.

[0105] Weight adjustment is performed synchronously during the iterative model training process. In the forward propagation phase of each training round, the input node features undergo multi-layer feature transformation to output electrical state estimates. The deviation between the estimated values ​​and the actual electrical measurement data is calculated, forming the task loss term. In the backpropagation phase, the task gradient is first obtained by differentiating the weight matrices of each layer based on the task loss term. Then, the correction offset corresponding to each set of weight matrices is calculated. The task gradient and the correction offset are weighted and superimposed according to a preset regularization coefficient to obtain the final weight update gradient. Based on the update gradient, parameter iteration is performed on each layer's weight matrix, repeatedly executing the forward calculation and backpropagation update process to gradually reduce the orthogonal deviation of each set of weight matrices, ultimately satisfying the orthogonal isometric transformation requirements.

[0106] Taking the feature transformation layer of a 1354-node regional power grid as an example, the steady-state projection sub-matrix has a dimension of 2×16, the disturbance projection sub-matrix has a dimension of 2×16, corresponding to a 2nd-order identity matrix; the fusion weight matrix has a dimension of 32×32, corresponding to a 32nd-order identity matrix. The orthogonality deviation and correction offset are calculated for the three sets of matrices respectively. During training, the regularization coefficients are taken as fixed small values ​​to ensure that the accuracy of the state recovery task is not disturbed while correcting the orthogonality of the weight matrices. After multiple iterations, the row vectors of each set of weight matrices are pairwise orthogonal and their magnitudes approach 1. The difference in the node features before and after the linear transformation is controlled within a very small range, and the geometric structure of the feature space remains stable.

[0107] In summary, this application transforms the low-rank characteristics and symmetric physical laws of the power grid sensitivity matrix into associated weights for feature aggregation. This reduces storage and computational complexity from quadratic to linear, significantly lowering the memory overhead for large-scale power grid modeling, while ensuring the feature fusion process closely aligns with real electrical coupling mechanisms. Combined with a block-decoupled linear transformation architecture and a dual-modal differentiated nonlinear mapping mechanism, it effectively reduces aliasing losses between steady-state operation and dynamic disturbance physical features, improving the precision and physical rationality of feature representation. Two-level identity propagation paths—one within the superposition layer and one at the block level—ensure stable cross-layer flow of original measurement information, mitigating gradient vanishing during deep network training. Furthermore, orthogonalization and equidistant constraints on the weight matrix maintain the intrinsic geometric structure of the feature space, suppressing feature homogenization during deep propagation. The entire scheme works synergistically to achieve high-precision and robust recovery of the electrical state of large-scale power grids with thousands of nodes under complex conditions such as missing measurements and noisy interference, while simultaneously considering model training efficiency and engineering deployment feasibility.

[0108] When a power grid fault occurs, on-site measurement and communication interruptions, sensor damage, and incomplete or distorted measurements are common. Relying directly on raw measurements for fault diagnosis can easily lead to misjudgments of operating conditions, and the handling procedures may introduce secondary risks. The electrical state estimates of all busbars obtained through graph modeling, weighted fusion of node features, and multi-layer feature transformation processing described in this application can accurately reconstruct the true electrical operating level of the power grid even under conditions of missing or noisy measurements. Therefore, this application uses the aforementioned calculated electrical state estimates to perform fault diagnosis and simulate candidate handling schemes, achieving reliable fault handling. Figure 4 This paper illustrates a flowchart of a fault handling method provided in an embodiment of this application. The specific handling steps are as follows: Figure 4 As shown.

[0109] Step 410: Obtain the electrical state estimate of all buses in the target power grid.

[0110] The electrical state estimates of all buses in the target power grid obtained in this step are the set of voltage amplitude estimates and voltage phase angle estimates output after processing in steps 110-150 above. The specific calculation process for each step is described in the corresponding embodiments above and will not be repeated here.

[0111] Step 420: Compare the estimated values ​​of each electrical state with the preset safety thresholds corresponding to each electrical state to obtain the results of the power grid parameters exceeding the limit.

[0112] For each busbar electrical state estimate obtained in step 410, the estimated voltage amplitude and voltage phase angle of that busbar are extracted and compared with preset safety thresholds one by one. The voltage amplitude corresponds to preset upper and lower voltage limits. When the estimated voltage amplitude of a busbar is higher than the upper voltage limit or lower than the lower voltage limit, it is determined that the busbar has exceeded the voltage limit. The voltage phase angle corresponds to a preset limit for the phase angle difference between adjacent buses, and the phase angle difference is calculated from the estimated voltage phase angles of two connected buses. When the phase angle difference exceeds the preset limit, it is determined that the connected branch has exceeded the phase angle limit.

[0113] Besides directly comparing voltage amplitude and phase angle, the estimated voltage amplitude and phase angle of all buses can be used to calculate the power flow estimate of each branch in conjunction with the grid branch parameters. Specifically, for any AC transmission branch connecting bus i and bus j, the resistance of that branch can be obtained. Reactance The branch admittance is calculated based on the resistance and reactance. Let the estimated amplitude of the voltage at bus i be... The voltage phase angle estimate is The estimated amplitude of the bus voltage j is The voltage phase angle estimate is The formulas for calculating the active power flow estimate and reactive power flow estimate of branch i to j are as follows: In the formula To take the real part of a complex number, For complex numbers, take the imaginary part. Traverse all electrical branches of the power grid, substituting the corresponding bus voltage magnitudes, phase angle estimates, and branch impedance parameters, and solve for the active power flow estimate of each branch. Reactive power flow estimation value The absolute value of active power flow and reactive power flow of each branch are compared with the pre-stored thermal stability active power flow limit and reactive power flow limit of that branch. If the estimated value of the branch power flow exceeds the corresponding thermal stability limit, the branch is determined to be overloaded.

[0114] After the above comparison is completed, the bus numbers, parameter types, and magnitudes of all instances of exceeding limits are summarized to form the power grid parameter limit exceedance results. These results indicate which locations and types of electrical parameters in the power grid deviated from the safe operating range after a fault occurred; the larger the magnitude of the exceedance, the more severely the location was affected by the fault.

[0115] Step 430: Generate multiple candidate fault handling schemes based on the results of grid parameter exceeding limits.

[0116] This step determines the fault handling operation that can eliminate the corresponding over-limit based on the bus number, over-limit parameter type and over-limit magnitude of all over-limits recorded in the power grid parameter over-limit results, and combines them to form multiple candidate fault handling schemes.

[0117] Specifically, the results of grid parameter overruns are first categorized according to the type of overrun parameter, resulting in voltage overrun sets, branch overload sets, and phase angle overrun sets. For each overrun bus in the voltage overrun set, optional operations to restore the bus voltage to a safe range are listed, including connecting capacitor banks, disconnecting reactors, and increasing generator excitation current, forming the voltage overrun optional operation set for that bus. For each overloaded branch in the branch overload set, optional operations to reduce the load rate of that branch are listed, including reducing the active power output of the sending-end units, increasing the active power output of the receiving-end units, switching part of the load to an adjacent branch, disconnecting the overloaded branch and transferring the load to downstream loads through tie switches, forming the overload optional operation set for that branch. For each pair of overrun bus pairs in the phase angle overrun set, optional operations to reduce the phase angle difference are listed, including adjusting the power injected at both ends and transferring load, forming the phase angle overrun optional operation set for that bus pair.

[0118] Subsequently, one operation is selected from each set of optional operations, and the selected operations are combined. Each combination constitutes a candidate fault handling scheme. Since each set of optional operations contains multiple optional operations, different selection methods correspond to different combinations, thus obtaining multiple candidate fault handling schemes. The multiple candidate fault handling schemes are initially sorted according to the total number of operations contained in each scheme, and schemes with a total number of operations exceeding a preset upper limit are eliminated, retaining candidate fault handling schemes with a reasonable number of operations.

[0119] Step 440: Using the electrical state estimate as the simulation benchmark, simulate each candidate fault handling scheme, determine the target fault handling scheme based on the simulation results, and execute the target fault handling scheme.

[0120] For any candidate fault handling plan, the handling operations are analyzed one by one according to the order of operations recorded in the plan. If it is a unit output adjustment operation, the active and reactive power output settings of the corresponding unit in the power grid simulation model are modified. If it is a capacitor or reactor switching operation, the reactive power compensation connection capacity of the corresponding bus is modified. If it is a load transfer operation, the power supply topology of the load node is changed. If it is a switching operation, the on / off state of the corresponding switch is modified, and the power grid topology connection relationship in the power grid simulation model is updated. After each operation parameter or topology modification is completed, the other parameters of the power grid simulation model remain unchanged. Using the existing bus voltage and phase angle as the calculation starting point, the voltage amplitude and voltage phase angle values ​​of each bus are repeatedly adjusted. After each round of adjustment, the deviation between the power input value and the power output value of each node in the power grid is determined. When the power input and output deviations of all nodes are less than the preset convergence threshold, the numerical adjustment is stopped, and the voltage amplitude and voltage phase angle of all buses after the operation is completed are obtained. Next, the active and reactive power flows of each branch are calculated. The specific calculation formulas for active and reactive power flows are the same as in step 420, and will not be repeated here. According to the preset execution order of the candidate fault handling scheme, the above-mentioned process of modifying the parameters or topology of the power grid simulation model, numerical adjustment calculation, and obtaining electrical results is repeated until all handling operations within the candidate fault handling scheme are completed. Finally, the simulated electrical state of the target power grid's bus voltage and branch power flows is obtained after all operations of the candidate fault handling scheme have been executed.

[0121] Next, referring to the comparison logic in step 420, the simulated bus voltage and branch power flow are compared with the corresponding preset safety thresholds. It is then determined whether the original over-limits are eliminated after the scheme is implemented, and whether any new bus voltage over-limits, branch overloads, or phase angle over-limits are generated in the simulation results. All candidate fault handling schemes are iterated through, and the scheme that can completely eliminate all original parameter over-limits identified in step 420 and will not generate new safety over-limits is selected as the qualified handling scheme.

[0122] When multiple qualified fault handling plans exist, the total number of equipment operations for each plan is counted, and the operational complexity is assessed based on the operation type. The plan with the fewest total equipment operations and the lowest operational complexity is selected as the target fault handling plan. After determining the target fault handling plan, all operation objects, operation types, and operation execution sequences included in the plan are analyzed. Corresponding control commands are issued to the dispatch control system, and the operations in the target fault handling plan are executed sequentially on-site to complete the power grid fault handling.

[0123] In summary, this application addresses the realities of missing or noisy measurements under fault conditions by directly using the aforementioned calculated electrical state estimate to handle fault-related processes, eliminating reliance on incomplete on-site measurements to obtain the true electrical state of the power grid. The electrical state estimate is compared with safety thresholds to obtain power grid parameter exceedance results. Based on these exceedance results, multiple candidate fault handling schemes are generated. The electrical state estimate is then used as a benchmark for offline simulation, from which the target fault handling scheme is selected. This avoids misjudgments of operating conditions caused by original measurement failures, pre-verifies whether a handling scheme will trigger new safety exceedances, selects the less costly handling scheme, reduces the risk of secondary faults induced by on-site equipment actions, shortens power grid fault handling and recovery time, and improves the safety and reliability of fault handling.

[0124] It is understood that, in order to achieve the functions in the above embodiments, the computer device includes hardware structures and / or software modules corresponding to the execution of each function. Those skilled in the art should readily recognize that, based on the units and method steps described in conjunction with the embodiments disclosed in this application, this application can be implemented in hardware or a combination of hardware and computer software. Whether a function is executed by hardware or by computer software driving hardware depends on the specific application scenario and design constraints of the technical solution.

[0125] Furthermore, as a response to the above Figure 1 The implementation of the method embodiment shown in this application provides a device for determining the electrical state of a power grid. The embodiment of this device corresponds to the foregoing method embodiments. For ease of reading, this embodiment will not repeat the details of the foregoing method embodiments one by one, but it should be clear that the device in this embodiment can correspondingly implement all the contents of the foregoing method embodiments. Specifically, as shown... Figure 5 As shown, the power grid electrical condition determination device 500 includes: The acquisition module 510 is used to acquire the topology of the target power grid, the measurement data of each bus in the target power grid, and the sensitivity matrix corresponding to the power flow equation of the target power grid. Module 520 is used to construct a graph model of the target power grid based on the topology. The graph model uses buses as nodes and electrical branches as connecting edges. The initial feature module 530 is used to generate corresponding initial features for each node in the graph model based on the measurement data of each bus. The feature fusion module 540 is used to determine the electrical coupling strength between different buses based on the sensitivity matrix, and to perform weighted fusion processing on the initial features of all buses according to the electrical coupling strength to obtain the node fusion features corresponding to all buses. The determination module 550 is used to perform multi-layer feature transformation processing on the node fusion features of all buses in combination with the node connection relationship of the graph model to obtain the electrical state estimate value corresponding to all buses.

[0126] Furthermore, such as Figure 5 As shown, the fusion feature module 540 is specifically used to perform truncated singular value decomposition on the sensitivity matrix, construct low-dimensional feature vectors corresponding to each bus based on the decomposition results, calculate the correlation weight between any two buses based on the low-dimensional feature vectors of each bus, and the correlation weight is used to characterize the electrical coupling strength between the buses; and perform weighted fusion processing on the initial features of all buses one by one based on the correlation weight to obtain the node fusion features corresponding to all buses.

[0127] Furthermore, such as Figure 5 As shown, the feature fusion module 540 is specifically used to perform complete singular value decomposition on the sensitivity matrix to obtain an orthogonal singular vector matrix and a set of singular values; select singular values ​​of the first preset order and orthogonal singular vectors corresponding to the preset order from the set of singular values, and remove the singular values ​​and orthogonal singular vectors corresponding to the remaining orders; for the row of the orthogonal singular vector matrix corresponding to a single bus line, extract the orthogonal singular vectors of the first preset order corresponding to the row of the matrix, and multiply each order of orthogonal singular vector by the square root of the corresponding singular value element by element to obtain the low-dimensional feature vector corresponding to the single bus line.

[0128] Furthermore, such as Figure 5 As shown, the feature fusion module 540 is specifically used to select the low-dimensional feature vectors corresponding to any two buses, multiply the corresponding values ​​of the two sets of low-dimensional feature vectors one by one, and then add all the values ​​to obtain the original interaction values ​​between the two buses; perform exponential transformation on the original interaction values ​​to obtain a single set of transformed values; take one of the two buses as the reference bus, accumulate the transformed values ​​corresponding to all buses that are connected to the reference bus by physical electrical branches, and obtain the summation result; perform normalization processing on the single set of transformed values ​​based on the summation result to obtain the correlation weight between any two buses.

[0129] Furthermore, such as Figure 5 As shown, the determination module 550 is specifically used to perform linear transformation on the node fusion features of each node using a preset weight matrix to generate linear transformation feature vectors corresponding to each node; to perform nonlinear mapping processing on each element in the linear transformation feature vector to obtain the output features of the single-layer feature transformation; to repeatedly execute the single-layer feature transformation and connect them layer by layer, with each layer taking the output features of the single-layer feature transformation of the previous layer as the input features of the current layer, thereby realizing multi-layer feature transformation.

[0130] Furthermore, such as Figure 5As shown, the determining module 550 is also used to, when performing single-layer feature transformation, directly transmit the input features of the single-layer feature transformation to the output end of the single-layer feature transformation and superimpose them with the features output by the single-layer feature transformation; divide the feature transformations of a continuous preset number of layers into a set of feature transformations, and directly transmit the input features of the first layer feature transformation in the feature transformation set to the output end of the last layer feature transformation in the feature transformation set and superimpose them with the features output by the last layer feature transformation.

[0131] Furthermore, such as Figure 5 As shown, the determining module 550 is also used to perform orthogonal constraint processing on the preset weight matrix corresponding to each layer of linear transformation based on preset constraint rules, and generate the corresponding correction offset. The preset constraint rule is that before and after the linear transformation calculation of the same weight matrix, the difference amplitude of any two sets of node features maintains an equidistant correspondence. Based on the correction offset and the deviation between the node features output by the multi-layer feature transformation and the actual electrical measurement data, the preset weight matrix corresponding to each layer of linear transformation is adjusted so that the linear transformation calculation of each layer meets the orthogonal equidistant transformation requirements.

[0132] Optionally, the power grid electrical state determination device may be an electronic device with data processing capabilities, or a functional module within the electronic device, without limitation.

[0133] For example, the electronic device can be a server, which can be a single server or a server cluster consisting of multiple servers. As another example, the electronic device can be a mobile phone, tablet computer, desktop computer, laptop computer, handheld computer, notebook computer, ultra-mobile personal computer (UMPC), netbook, as well as cellular phones, personal digital assistants (PDAs), augmented reality (AR) devices, virtual reality (VR) devices, and other terminal devices. As yet another example, the electronic device can also be a recording device, video surveillance equipment, etc. This application does not impose any special limitations on the specific form of the electronic device.

[0134] The following example uses an electronic device to determine the electrical state of the power grid. Figure 6 As shown, Figure 6 The hardware structure of an electronic device 600 provided in this application.

[0135] like Figure 6 As shown, the electronic device 600 includes a processor 610, a communication line 620, and a communication interface 630.

[0136] Optionally, the electronic device 600 may also include a memory 640. The processor 610, memory 640, and communication interface 630 can be connected via a communication line 620.

[0137] The processor 610 can be a central processing unit (CPU), a general-purpose processor, a network processor (NP), a digital signal processor (DSP), a microprocessor, a microcontroller, a programmable logic device (PLD), or any combination thereof. The processor 610 can also be any other device with processing capabilities, such as a circuit, device, or software module, without limitation.

[0138] In one example, processor 610 may include one or more CPUs, for example Figure 6 CPU0 and CPU1 in the CPU.

[0139] As an optional implementation, the electronic device 600 may include multiple processors, for example, in addition to processor 610, it may also include processor 670. A communication line 620 is used to transmit information between the components included in the electronic device 600.

[0140] Communication interface 630 is used for communicating with other devices or other communication networks. This other communication network can be Ethernet, Radio Access Network (RAN), Wireless Local Area Network (WLAN), etc. Communication interface 630 can be a module, circuit, transceiver, or any device capable of enabling communication.

[0141] The memory 640 is used to store instructions. These instructions can be computer programs.

[0142] The memory 640 can be a read-only memory (ROM) or other type of static storage device capable of storing static information and / or instructions; it can also be a random access memory (RAM) or other type of dynamic storage device capable of storing information and / or instructions; it can also be an electrically erasable programmable read-only memory (EEPROM), a compact disc read-only memory (CD-ROM) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), disk storage media, or other magnetic storage devices, etc., without limitation.

[0143] It should be noted that the memory 640 can exist independently of the processor 610, or it can be integrated with the processor 610. The memory 640 can be used to store instructions, program code, or some data, etc. The memory 640 can be located inside or outside the electronic device 600, without restriction.

[0144] The processor 610 is configured to execute instructions stored in the memory 640 to implement the communication method provided in the following embodiments of this application. For example, when the electronic device 600 is a terminal or a chip in a terminal, the processor 610 can execute instructions stored in the memory 640 to implement the steps performed by the transmitting end in the following embodiments of this application.

[0145] As an optional implementation, the electronic device 600 also includes an output device 650 and an input device 660. The output device 650 can be a display screen, speaker, or other device capable of outputting data from the electronic device 600 to the user. The input device 660 can be a keyboard, mouse, microphone, joystick, or other device capable of inputting data into the electronic device 600.

[0146] It should be pointed out that, Figure 6 The structure shown does not constitute a limitation on the electronic device, except... Figure 6 In addition to the components shown, the electronic device may include more or fewer components than illustrated, or combine certain components, or have different component arrangements.

[0147] The power grid electrical state determination device and application scenarios described in the embodiments of this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided in the embodiments of this application. As those skilled in the art will know, with the evolution of power grid electrical state determination devices and the emergence of new business scenarios, the technical solutions provided in the embodiments of this application are also applicable to similar technical problems.

[0148] This application provides a storage medium storing a program that, when executed by a processor, implements the method for determining the electrical state of the power grid.

[0149] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0150] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0151] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A method for determining the electrical state of a power grid, characterized in that, The method includes: The topology of the target power grid, the measurement data of each bus in the target power grid, and the sensitivity matrix corresponding to the power flow equation of the target power grid are obtained. A graph model of the target power grid is constructed based on the aforementioned topology, wherein the graph model uses buses as nodes and electrical branches as connecting edges. Based on the measurement data of each busbar, generate corresponding initial features for each node in the graph model; Based on the sensitivity matrix, the electrical coupling strength between different buses is determined. The initial features of all buses are then weighted and fused one by one according to the electrical coupling strength to obtain the node fusion features corresponding to all buses. By combining the node connection relationships of the graph model, multi-level feature transformation processing is performed on the node fusion features of all buses to obtain the electrical state estimates corresponding to all buses.

2. The method according to claim 1, characterized in that, Based on the sensitivity matrix, the electrical coupling strength between different buses is determined. Then, the initial features of all buses are weighted and fused according to the electrical coupling strength to obtain the node fusion features corresponding to all buses, including: The sensitivity matrix is ​​subjected to truncated singular value decomposition, and low-dimensional feature vectors corresponding to each bus are constructed based on the decomposition results. The association weight between any two buses is calculated based on the low-dimensional feature vectors of each bus, and the association weight is used to characterize the electrical coupling strength between the buses. Based on the aforementioned association weights, the initial features of all buses are weighted and fused one by one to obtain the node fusion features corresponding to all buses.

3. The method according to claim 2, characterized in that, The sensitivity matrix is ​​subjected to truncated singular value decomposition, and low-dimensional feature vectors corresponding to each bus are constructed based on the decomposition results, including: Perform a complete singular value decomposition on the sensitivity matrix to obtain an orthogonal singular vector matrix and a set of singular values; Select singular values ​​of the first preset order and orthogonal singular vectors corresponding to the preset order from the set of singular values, and remove singular values ​​and orthogonal singular vectors corresponding to the remaining orders. For each row of the orthogonal singular vector matrix corresponding to a single busbar, extract the orthogonal singular vectors of the first preset order corresponding to the row of the matrix, and multiply each orthogonal singular vector of the first order by the square root of the corresponding singular value element by element to obtain the low-dimensional feature vector corresponding to the single busbar.

4. The method according to claim 2, characterized in that, Calculate the association weight between any two buses based on the low-dimensional feature vectors of each bus, including: Select any two low-dimensional feature vectors corresponding to each of the two sets of low-dimensional feature vectors, multiply the corresponding values ​​of the two sets of low-dimensional feature vectors one by one, and then add all the values ​​to obtain the original interaction values ​​between the two gene lines. The original interaction values ​​are subjected to exponential transformation to obtain a single set of transformed values; Using one of the two busbars as the reference busbar, the conversion values ​​corresponding to all busbars that are connected to the reference busbar by physical electrical branches are summed to obtain a summation result. Based on the summation result, the single set of conversion values ​​is normalized to obtain the correlation weight between the two busbars.

5. The method according to any one of claims 1-4, characterized in that, Based on the node connection relationships of the graph model, multi-layer feature transformation processing is performed on the node fusion features of all buses, including: The node fusion features of each node are linearly transformed using a preset weight matrix to generate linearly transformed feature vectors for each node. By performing nonlinear mapping on each element of the linear transformation feature vector, the output features of the single-layer feature transformation are obtained. The single-layer feature transformation is repeatedly executed and chained layer by layer. Each layer uses the output features of the single-layer feature transformation of the previous layer as the input features of the current layer, thereby realizing multi-layer feature transformation.

6. The method according to claim 5, characterized in that, The method further includes: When performing single-layer feature transformation, the input features of the single-layer feature transformation are directly passed to the output of the single-layer feature transformation and superimposed with the features output by the single-layer feature transformation. The feature transformations of a continuous preset number of layers are divided into a set of feature transformations. The input features of the feature transformation of the starting layer in the feature transformation set are directly passed to the output of the feature transformation of the last layer in the feature transformation set, and then superimposed with the features output by the feature transformation of the last layer.

7. The method according to claim 5, characterized in that, The method further includes: Based on preset constraint rules, the preset weight matrices corresponding to the linear transformation of each layer are orthogonally constrained to generate corresponding correction offsets. The preset constraint rule is that before and after the linear transformation calculation of the same weight matrix, the difference magnitude of any two sets of node features maintains an equidistant correspondence. Based on the correction offset and the deviation between the node features output by the multi-layer feature transformation and the actual electrical measurement data, the preset weight matrix corresponding to each layer of linear transformation is adjusted so that the calculation of each layer of linear transformation meets the requirements of orthogonal equidistant transformation.

8. A device for determining the electrical state of a power grid, characterized in that, The device includes: The acquisition module is used to acquire the topology of the target power grid, the measurement data of each bus in the target power grid, and the sensitivity matrix corresponding to the power flow equation of the target power grid. A construction module is used to construct a graph model of the target power grid based on the topology, wherein the graph model uses buses as nodes and electrical branches as connecting edges; The initial feature module is used to generate corresponding initial features for each node in the graph model based on the measurement data of each bus. The feature fusion module is used to determine the electrical coupling strength between different buses based on the sensitivity matrix, and to perform weighted fusion processing on the initial features of all buses one by one according to the electrical coupling strength to obtain the node fusion features corresponding to all buses. The determination module is used to perform multi-layer feature transformation processing on the node fusion features of all buses in combination with the node connection relationship of the graph model to obtain the electrical state estimate value corresponding to all buses.

9. A storage medium, characterized in that, The storage medium includes a stored program, wherein, when the program is executed, it controls the device containing the storage medium to perform the power grid electrical state determination method as described in any one of claims 1-7.

10. An electronic device, characterized in that, The device includes at least one processor, at least one memory connected to the processor, and a bus; wherein the processor and the memory communicate with each other through the bus; the processor is used to call program instructions in the memory to execute the power grid electrical state determination method as described in any one of claims 1-7.

Images