Online simulation method of LCC direct current transmission electromechanical electromagnetic hybrid system

Abstract

The invention relates to the technical field of direct-current power transmission system simulation, and discloses an online simulation method of an LCC direct-current power transmission electromechanical electromagnetic hybrid system. The method comprises the following steps: constructing a system diagram structure model, and respectively representing electrical components and electromagnetic coupling relations thereof by nodes and edges; mapping the real-time operation data into a space-time diagram data sequence; performing multi-resolution wavelet decomposition on the sequence, and extracting a low-frequency contour and a high-frequency detail electromagnetic characteristic mode; dynamically calculating a node weight according to a feature mode by using a graph attention network, and generating enhanced electromagnetic field distribution fusing local and global features; and finally, a simulation signal is output by solving an electromagnetic field equation. According to the method, through multi-scale feature separation and dynamic coupling relation modeling, finer and more adaptive simulation of the broadband electromagnetic transient process of the system is realized, and the simulation precision and the depicting capability of complex working conditions are improved.

Description

Technical Field

[0001] This invention relates to the field of DC transmission system simulation technology, specifically to an online simulation method for an LCC DC transmission electromechanical hybrid system. Background Technology

[0002] Currently, the mainstream methods for electromechanical-electromagnetic hybrid simulation of LCC DC transmission systems rely on models based on fixed network parameters or simplified equivalent circuits. When processing spatiotemporally correlated electromagnetic transient data, these methods typically analyze electrical quantities as independent time series, failing to fully reflect the complex electromagnetic interactions arising from the spatial connections of components within the system. Electromagnetic transient processes encompass physical characteristics at different time scales, but existing simulation techniques often process signals across the entire frequency band uniformly, lacking effective separation and targeted modeling of low-frequency components characterizing the dominant trends of the system and high-frequency components characterizing details such as switching and faults. This results in insufficient fidelity in simulating complex transient phenomena.

[0003] When calculating electromagnetic field distribution or the interactions between components, existing techniques typically determine weighting relationships based on physical topology or pre-calculated fixed admittance matrices. This static weighting method fails to reflect the dynamic characteristics of electromagnetic coupling strength as the system operates under abnormal conditions or at stable boundary conditions. When a local disturbance occurs in the system, some remote components may generate effects far exceeding those of the normal state through electromagnetic induction. Models based on fixed weights struggle to capture the deep correlations of this dynamic evolution, thus affecting the accurate assessment of the system's global behavior, particularly the fault evolution process. Summary of the Invention

[0004] The purpose of this invention is to provide an online simulation method for an LCC DC transmission electro-electromagnetic hybrid system to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, this invention provides an online simulation method for an LCC DC transmission electro-electromagnetic hybrid system, the method comprising: Construct a graph structure model of the LCC DC transmission electromechanical hybrid system, where nodes represent electrical components and edges represent electromagnetic coupling relationships, and define node attributes and edge weights; Real-time running data is collected and mapped to nodes and edges of a graph structure model to form a spatiotemporal graph data sequence, which includes timestamps and spatial coordinate information. Multi-resolution wavelet decomposition was performed on the spatiotemporal graph data sequence to extract electromagnetic feature modes at different scales, including low-frequency contour modes and high-frequency detail modes. Based on electromagnetic feature modes, a graph attention network is used to calculate the dynamic weights between nodes to generate an enhanced electromagnetic field distribution, which integrates local and global features. The enhanced electromagnetic field distribution is fused with the original spatiotemporal map data sequence, and a high-fidelity simulation signal is output by solving the electromagnetic field equations.

[0006] Preferably, the graphical structure model for constructing the LCC DC transmission electro-electromagnetic hybrid system specifically includes: Identify the key electrical components in the LCC DC transmission electro-electromagnetic hybrid system, including converter valves, transformers, and transmission lines, and abstract each component as a graph node; The connection edges between nodes are defined based on electromagnetic coupling relationships, and the connection edges include electrical connections and electromagnetic induction relationships. Assign an attribute vector to each node, the attribute vector including electrical parameters and physical parameters; A weight value is calculated for each edge, the weight value being determined based on the electromagnetic coupling strength and the distance attenuation factor; Combine nodes and edges into a graph structure model and store graph topology information.

[0007] Preferably, the step of collecting real-time operational data and forming a spatiotemporal graph data sequence includes: Real-time operational data is acquired from a sensor network, including voltage, current, and electromagnetic field strength. Sort real-time running data by timestamp to generate time series data; Map time series data to corresponding nodes in the graph structure model and update the node attribute vectors; Data from adjacent nodes is aggregated based on edge weights to form a data stream along the edge. By combining timestamps and spatial coordinates, a spatiotemporal graph data sequence is constructed, which is in the form of a multidimensional tensor.

[0008] Preferably, the step of performing multi-resolution wavelet decomposition on the spatiotemporal graph data sequence to extract electromagnetic feature modes includes: By selecting wavelet basis functions, wavelet transform is performed on the spatiotemporal graph data sequence to obtain coefficients at multiple resolution levels. Low-frequency subband coefficients are extracted from the coefficients to reconstruct the low-frequency profile modes, representing the macroscopic variation trend of the electromagnetic field; High-frequency subband coefficients are extracted from the coefficients to reconstruct high-frequency detail modes, representing the rapid fluctuation characteristics of the electromagnetic field; The low-frequency profile mode and the high-frequency detail mode are normalized to eliminate dimensional differences; The normalized modes are combined into a set of electromagnetic characteristic modes.

[0009] Preferably, the method of generating the enhanced electromagnetic field distribution using a graph attention network includes: The set of electromagnetic feature modes is input into a graph attention network, which contains a multi-layer attention mechanism. Calculate the attention weights of each node and its neighboring nodes, whereby the attention weights are based on node attributes and edge weights; The feature representation of each node is updated by weighted aggregation of the features of neighboring nodes; Repeated multi-layer attention calculations are performed to obtain enhanced node feature vectors. These enhanced node feature vectors are then mapped back to the physical space to generate an enhanced electromagnetic field distribution.

[0010] Preferably, the fusion of the enhanced electromagnetic field distribution with the original spatiotemporal map data sequence includes: The enhanced electromagnetic field distribution is interpolated to match the resolution of the original spatiotemporal map data sequence. The difference between the enhanced electromagnetic field distribution and the original data is calculated to generate a residual map. The residual map is then weighted and superimposed with the original spatiotemporal map data sequence to obtain fused data. The fused data is then smoothed and filtered to eliminate noise during the fusion process.

[0011] Preferably, the step of solving the electromagnetic field equations includes: Establish the electromagnetic field equations, which include Maxwell's equations and boundary conditions; Using the fused data as initial conditions, we substitute it into the electromagnetic field equations and define physical constraints, including energy conservation and charge continuity. The electromagnetic field equations are solved iteratively using an optimization algorithm based on the gradient descent method. When the solution converges, the electromagnetic field distribution solution is output as the simulation result.

[0012] Preferably, after outputting the high-fidelity simulation signal, the method further includes: The simulation results are post-processed, including time-domain resampling and spatial-domain interpolation. The post-processed results are compared with the actual running data, and the error index is calculated. When the error index exceeds the threshold, the parameters or physical constraints of the graph attention network are adjusted, and the simulation process is re-executed until the error index meets the requirements.

[0013] Preferably, the method further includes an online learning mechanism: Collect historical data during the simulation process, including input data and output results, construct a training dataset, use it to update the model parameters of the graph attention network, retrain the graph attention network periodically to adapt to changes in the system's operating state, and apply the updated model to subsequent simulation tasks.

[0014] Compared with the prior art, the beneficial effects of the present invention are: Multi-resolution wavelet decomposition of spatiotemporal graph data sequences effectively separates different scale features in electromagnetic transient processes. This method clearly distinguishes between low-frequency profile modes characterizing the dominant dynamics of the system and high-frequency modes reflecting switching actions and fault details. This scale-level deconstruction makes the originally mixed electromagnetic signals more hierarchical, with low-frequency modes carrying the overall stability trend of the system and high-frequency modes accurately capturing locally rapidly changing transient information. Subsequent electromagnetic calculations can be processed separately according to the physical characteristics of different modes, avoiding the problem of smoothing or obscuring detailed features when calculating broadband signals uniformly. The simulation output can thus accurately reproduce both the macroscopic dynamic evolution and microscopic transient details of the system, improving the overall fidelity of the waveform in both the time and frequency domains.

[0015] By dynamically calculating node weights using a graph attention network, the electromagnetic coupling relationship is transformed from static to dynamic. The network generates attention scores between nodes based on real-time operating conditions, and the weight allocation adaptively adjusts according to system operating conditions. During normal and stable operation, the weight distribution reflects the basic physical connections; when disturbances occur or the system approaches a stability boundary, the network automatically identifies and amplifies the impact of critical propagation paths. This dynamic coupling mechanism accurately simulates the propagation of electromagnetic disturbances in real-world systems, especially under fault conditions, realistically depicting the spread path of fault currents and the electromagnetic induction intensity of remote equipment. The simulation model thus possesses the ability to capture complex state-dependent interactions, and its simulation of system transient behavior, particularly the fault evolution process, more closely resembles physical reality. Attached Figure Description

[0016] Figure 1 This is a schematic diagram illustrating the working principle of the online simulation method for the LCC DC transmission electro-electromagnetic hybrid system described in this invention. Figure 2 A flowchart for collecting real-time operational data and forming a spatiotemporal graph data sequence; Figure 3 A flowchart for extracting electromagnetic feature modes from multi-resolution wavelet decomposition; Figure 4 To compare the error rates at each stage of the simulation; Figure 5 This is a comparison chart showing the model accuracy versus relative processing time under different data augmentation methods. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Please see Figure 1 This invention provides an online simulation method for an LCC DC transmission electro-electromagnetic hybrid system. The method includes: constructing a graph structure model of the system, where nodes represent electrical components such as converter valves and transformers, and edges represent electromagnetic coupling relationships such as electrical connections and electromagnetic induction; node attributes include electrical parameters and physical parameters, and edge weights are determined based on electromagnetic coupling strength and distance attenuation factor. Real-time operating data, including voltage, current, and electromagnetic field strength, is collected from a sensor network. This data is sorted by timestamp and mapped to nodes and edges of the graph structure model, forming a spatiotemporal graph data sequence containing timestamp and spatial coordinate information, which is stored in the form of a multidimensional tensor. Multi-resolution wavelet decomposition is performed on the spatiotemporal graph data sequence. Appropriate wavelet basis functions are selected for wavelet transform to extract low-frequency contour modes and high-frequency detail modes, representing the macroscopic variation trend and rapid fluctuation characteristics of the electromagnetic field, respectively. The modes are normalized to eliminate dimensional differences. Based on the extracted electromagnetic feature modes, a graph attention network is used to calculate dynamic attention weights between nodes. Neighbor node features are weighted and aggregated through a multi-layer attention mechanism to update the node feature representation, generating an enhanced electromagnetic field distribution that integrates local and global features. The enhanced electromagnetic field distribution is fused with the original spatiotemporal map data sequence. Interpolation is used to ensure consistent resolution, and the residual map is calculated, then weighted and filtered to obtain the fused data. Using the fused data as initial conditions, Maxwell's equations and boundary conditions are established for the electromagnetic field equations. Physical constraints such as energy conservation and charge continuity are defined, and optimization algorithms such as gradient descent are used to iteratively solve the equations. A high-fidelity simulation signal is output when the solution converges. This method also includes an online learning mechanism to periodically update model parameters to adapt to system changes.

[0019] Example 1: See Figure 2In practical implementation, constructing a graph structure model of an LCC DC transmission electro-electromagnetic hybrid system requires systematically identifying and abstracting all key electrical components in the system. These key electrical components include, but are not limited to, converter valves, transformers, and transmission lines. Each independent electrical component is abstracted as an independent node in the graph structure model. The connections between nodes are strictly defined based on electromagnetic coupling relationships, which specifically include direct electrical connections and indirect electromagnetic induction relationships. Assigning attribute vectors to each node is a core step in model construction. Attribute vectors contain two types of parameters: electrical parameters, such as instantaneous voltage values, effective current values, and power factors at the node; and physical parameters, such as the node's three-dimensional spatial coordinates, geometric dimensions, and material properties. The weight of each edge needs to be calculated based on the electromagnetic coupling strength and the distance attenuation factor. The electromagnetic coupling strength can be quantified by measuring or simulating the mutual inductance coefficient or electric field strength between nodes, while the distance attenuation factor is typically inversely proportional to the physical distance between nodes. Finally, all the defined nodes and edges are combined according to the topological connections of the actual system to form a complete graph structure model, and the graph topology information, including the node list, edge list, attribute set and weight matrix, is persistently stored in the database.

[0020] In practice, collecting real-time operational data and forming a spatiotemporal graph data sequence is a dynamic data processing flow. Real-time operational data is continuously acquired from a sensor network pre-deployed at key components of the LCC DC transmission electro-electromagnetic hybrid system. This data primarily includes voltage waveform data, current waveform data, and spatial electromagnetic field strength measurements. These raw data streams are first rigorously sorted according to their inherent high-precision timestamps to generate continuous time-series data. Subsequently, the sorted time-series data is mapped to corresponding nodes in the graph structure model using a mapping algorithm. This mapping process updates the node attribute vectors in real time; for example, the voltage measurement value at a certain moment is written into the voltage attribute field of the corresponding converter valve node. Based on the updated node data, the data of adjacent nodes is aggregated according to predefined edge weights in the graph structure. For example, a weighted average algorithm is used to calculate the equivalent electromagnetic field strength flowing along the edge, thus forming the edge data stream. Finally, the node attribute data with timestamps and the data flow on the edges are combined with the node spatial coordinate information extracted from the graph structure model to construct a unified spatiotemporal graph data sequence. This sequence is organized and stored in computer memory in the form of a multidimensional tensor, with dimensions typically including time dimension, node index dimension, and feature dimension.

[0021] In some embodiments, the construction process of the graph structure model can incorporate automated modeling tools. These tools can parse single-line diagrams or CAD design files of the power system, automatically identify component types and connection relationships, and generate initial graph nodes and edges. In specific implementations, the definition of node attribute vectors can be extended to include dynamic operating state parameters, such as equipment temperature and insulation aging levels, enabling the model to more comprehensively reflect the real-time status of components. Edge weights can be calculated using more accurate electromagnetic field numerical calculation methods, such as the finite element method, to solve for coupling strength under complex geometric layouts, thereby improving the accuracy of weight values. Graph topology information can be stored using a dedicated graph database system to achieve efficient graph traversal and query operations.

[0022] In some embodiments, real-time data acquisition can be achieved using a high-precision synchronous phasor measurement unit (HTP). The HTP provides high-sampling-rate data with time stamps, ensuring the synchronization and accuracy of the time-series data. In specific implementations, the data mapping process needs to handle the non-one-to-one correspondence between sensors and graph nodes. For example, when a sensor monitors multiple components or a component is equipped with multiple sensors, a data fusion algorithm needs to be designed to determine the final value of the node attribute to be written. The data stream aggregation algorithm on the edges can be differentiated according to the edge type. For edges representing strong electrical connections, direct weighted summation can be used; for edges representing weak electromagnetic induction, filtering can be introduced to highlight the characteristics of specific frequency bands. The multidimensional tensor format of the spatiotemporal graph data sequence can be a common data format such as HDF5 or TFRecord to facilitate direct reading by subsequent distributed processing and machine learning frameworks.

[0023] Optionally, when constructing the graph structure model, version identifiers can be added to nodes and edges. These identifiers allow the model to track the history of changes in component parameters or connections, thus supporting scenarios with dynamic changes in system topology. In practice, when system expansion or maintenance leads to the addition or removal of components or changes in connections, a new version of the graph structure model can be generated by updating the version identifiers, while retaining historical versions for backtracking analysis. This mechanism ensures that the graph structure model remains consistent with the physical system. During the formation of the spatiotemporal graph data sequence, a data quality verification step can be implemented. This step performs outlier detection, missing value imputation, and consistency checks on the original sensor data. In practice, outlier detection can employ statistical methods, such as Z-score detection, to identify and remove measurement points that significantly deviate from the normal range; missing value imputation can use time series prediction methods, such as linear interpolation or ARIMA models, to estimate the data at the missing time; and consistency checks verify whether the measurement results of different sensors for the same physical quantity are within the allowable error range. Through data quality verification, the reliability of input data can be significantly improved, providing a clean and consistent data foundation for subsequent simulation calculations.

[0024] It is understandable that the node attribute vector definition in a graph structure model needs to possess good scalability. Scalability allows for the easy introduction of new attribute types in subsequent development, such as vibration or acoustic signals used to assess the health status of equipment, without reconstructing the entire model. In practical implementation, attribute vectors can be stored using a flexible key-value pair data structure, where each attribute corresponds to a unique key name and a specific value. This design ensures that adding a new attribute only requires adding a new key-value pair, without affecting the storage and access of existing attributes. It is also understandable that the construction of spatiotemporal graph data sequences needs to consider the efficiency of large-scale data processing. Efficiency is reflected in the design of the data pipeline and the optimization of the storage format. In practical implementation, a stream processing framework can be used to receive and process sensor data streams in real time, achieving low-latency data sequence construction. Storage format optimization involves block compression of multidimensional tensors, retaining only active or recent time-slice data in memory, while archiving historical data to high-speed disks, thereby achieving a balance between computing resources and storage requirements.

[0025] Example 2: See Figure 3In practical implementation, multi-resolution wavelet decomposition is performed on spatiotemporal graph data sequences. The core of this process is selecting appropriate wavelet basis functions and transforming the input data. The choice of wavelet basis functions directly affects the feature extraction effect. Commonly used wavelet basis functions include the Daubechies wavelet family. The selection criteria are mainly based on the characteristics of the spatiotemporal graph data sequences themselves. For example, for electromagnetic field signals with smooth changes, the Daubechies wavelet is more suitable due to its tight support and regularity; while for signals containing abrupt changes, the simple structure of the Haar wavelet can better capture edge information. After selecting the wavelet basis functions, a discrete wavelet transform is performed on the spatiotemporal graph data sequences. The discrete wavelet transform convolves and downsamples the input signal through a series of low-pass and high-pass filters, decomposing the signal into different resolution levels. Each level generates a set of approximate coefficients (low frequency) and a set of detail coefficients (high frequency), thus obtaining wavelet coefficients at multiple resolution levels.

[0026] In practical implementation, extracting electromagnetic feature modes from wavelet coefficients at multiple resolution levels is a crucial step. Low-frequency subband coefficients are extracted by retaining approximate coefficients at the coarsest resolution level; these low-frequency subband coefficients reflect the macroscopic profile and slow changing trends of the signal. Signal reconstruction using these low-frequency subband coefficients involves performing an inverse discrete wavelet transform to obtain low-frequency profile modes. These low-frequency profile modes characterize the overall evolution of the electromagnetic field in an LCC DC transmission system, such as the energy distribution near the power frequency. Extraction of high-frequency detail modes involves extracting detail coefficients from various finer resolution levels. These detail coefficients capture the local details and rapid fluctuations of the signal. Similarly, inverse transforms are used to reconstruct high-frequency detail modes at different scales. These high-frequency detail modes correspond to transient processes and harmonic components in the electromagnetic field, such as electromagnetic oscillations caused by commutation. In practical implementation, the reconstruction process requires precise timestamp alignment to ensure temporal consistency between the modes and the original sequence.

[0027] In practical implementation, normalization of the reconstructed low-frequency profile modes and high-frequency detail modes is essential. The purpose of normalization is to eliminate the influence of differences in dimensions and numerical ranges between different modes. Commonly used normalization methods include min-max normalization and Z-score normalization. Min-max normalization linearly scales the modal data to the [0,1] interval, as shown in the formula:

[0028] in: Represents the original modal data points. It is the minimum value of this modal data. It is the maximum value of this modal data. This is the result after normalization. Normalization ensures that the subsequent graph attention network can treat features of different scales equally, preventing certain modalities from dominating the learning process due to excessively large values.

[0029] In some embodiments, the number of wavelet decomposition layers (i.e., the number of resolution levels) is a parameter that needs to be determined, as it directly affects the granularity of feature extraction. In specific implementations, the number of wavelet decomposition layers can be determined based on the sampling frequency of the spatiotemporal graph data sequence and the highest frequency component of interest. A rule of thumb is that the number of decomposition layers... Should meet ,in: It is the sampling frequency. This refers to the highest frequency that needs to be analyzed. For example, when focusing on electromagnetic interference with a maximum frequency of 1 kHz, if the sampling frequency is 10 kHz, the maximum number of decomposition layers is approximately 3. By adjusting the number of layers, the feature scale can be controlled from coarse to fine, achieving layered analysis of the macroscopic trends and microscopic fluctuations of the electromagnetic field.

[0030] In some embodiments, after extracting low-frequency and high-frequency subband coefficients from the coefficients, a coefficient thresholding step can be introduced before reconstruction. Coefficient thresholding aims to suppress noise and enhance effective features. In specific implementations, a threshold function, such as a hard thresholding or soft thresholding function, is applied to the obtained wavelet coefficients. A hard thresholding function accepts values ​​whose absolute values ​​are less than a certain threshold. The coefficients of the hard threshold are set to zero, while the coefficients greater than the threshold remain unchanged; the soft threshold function, based on the hard threshold, further shrinks the retained coefficients towards zero. Size. Threshold The choice can be based on noise estimation, for example, using a universal threshold. ,in: It is an estimate of the noise standard deviation. This refers to the signal length. After thresholding, the coefficients are reconstructed to obtain the denoised low-frequency contour modes and high-frequency detail modes, thus improving feature quality.

[0031] In some embodiments, when combining normalized modes into an electromagnetic feature mode set, a tensor concatenation method can be used, which can preserve the scale information of each mode. In a specific implementation, low-frequency contour modes and high-frequency detail modes at different scales are treated as multiple feature maps and concatenated according to the scale dimension to form a three-dimensional tensor as the electromagnetic feature mode set. The three dimensions of this three-dimensional tensor represent time points, spatial nodes, and feature modes, respectively. This organization ensures that each spatial node corresponds to a feature vector at each time step, which integrates multi-scale electromagnetic information from macroscopic to microscopic levels, providing rich input features for the graph attention network.

[0032] Optionally, for online simulation applications with extremely high real-time requirements, the Mallat algorithm, a fast algorithm based on discrete wavelet transform, can be used. The Mallat algorithm achieves multi-resolution analysis through iterative use of filter banks, resulting in high computational efficiency. In its implementation, the Mallat algorithm uses low-pass and high-pass filters corresponding to the wavelet basis functions to convolve and downsample the input signal, decomposing it step by step. The reconstruction process is the inverse of the decomposition, achieved through upsampling and filtering. Using the Mallat algorithm effectively reduces computational complexity and meets the stringent latency requirements of online processing. It is understandable that the choice of wavelet basis functions is not static; it can be optimized as a hyperparameter. In practice, a representative validation dataset can be prepared to evaluate the impact of different wavelet basis functions on the final simulation accuracy, thereby selecting the most suitable wavelet basis function for the specific LCC DC transmission system simulation task. This data-driven selection method can improve the targeting of feature extraction.

[0033] It is understandable that the dimensionality of the electromagnetic feature mode set may be high, potentially leading to a computational burden. In practical implementation, feature selection or dimensionality reduction steps can be added after mode combination and before the input graph attention network. For example, principal component analysis (PCA) or an autoencoder can be used to compress the feature dimensionality, remove redundant information, and improve the processing efficiency of the subsequent graph attention network while maintaining the feature expressiveness. This approach helps to achieve a balance between computational resources and simulation accuracy.

[0034] Example 3: In specific implementation, the process of generating an enhanced electromagnetic field distribution based on electromagnetic feature modes using a graph attention network begins by inputting a set of electromagnetic feature modes into the graph attention network architecture. The set of electromagnetic feature modes is a three-dimensional tensor containing multi-scale low-frequency contour modes and high-frequency detail modes. The graph attention network is designed as a computational graph containing multiple layers of attention mechanisms, each layer consisting of multiple parallel attention heads to capture different feature interaction patterns. Calculating the attention weights of each node and its neighboring nodes is the core operation. The calculation of attention weights depends on the node attribute vector and edge weight information. The node attribute vector comes from the feature slices of the corresponding nodes in the electromagnetic feature modes, while the edge weights originate from the electromagnetic coupling strength defined in the pre-constructed graph structure model. In specific calculations, for each node in the graph, its own feature vector is concatenated or dot-productd with the feature vectors of each neighboring node, and the edge weights are incorporated as a bias term. An unnormalized attention score is generated through a learnable weight matrix and activation function. Subsequently, a softmax function is used to normalize all these scores, so that the sum of the attention weights of each node to all its neighbors is 1, thus obtaining a dynamic attention weight distribution. This dynamic weight reflects the importance of the interactions between nodes at a specific electromagnetic characteristic scale.

[0035] In practice, the feature representation of each node is updated by weighted aggregation of the features of its neighbors. The weighted aggregation operation uses calculated attention weights as coefficients to linearly combine the feature vectors of each neighbor node. For the target node, its new feature representation is a weighted sum of the feature vectors of all its neighbors, plus a residual connection of its own features, to preserve the original information. The aggregation function can be expressed as:

[0036] in: It is the updated feature vector of node u. It is a non-linear activation function such as ELU. Let u represent the set of its neighboring nodes. It is the attention weight of node u to its neighbor v. It is a trainable weight matrix. It is the original feature vector of the neighbor node v. It is a trainable bias vector. Repeated multi-layer attention computation means stacking multiple such graph attention layers, typically 2 to 4 layers, with the output of each layer serving as the input to the next layer. Through this deep structure, node features can fuse information from multi-hop neighbors, thereby capturing a broader graph structure context and ultimately obtaining enhanced node feature vectors that encode local and global electromagnetic field properties.

[0037] In practice, the enhanced node feature vectors are mapped back to physical space to generate an enhanced electromagnetic field distribution. This mapping process is typically achieved through a fully connected neural network or a linear transformation layer, which projects the high-dimensional node feature vectors to a dimension that matches the actual physical field measurements, such as three-dimensional electromagnetic field strength components. The parameters of the mapping layer are learned during training to ensure that the output is aligned with physical reality. The generated enhanced electromagnetic field distribution is spatially discrete, with each graph node location corresponding to a field strength value; therefore, it needs to be converted into a continuous field distribution for subsequent use. In some embodiments, the graph attention network can employ a multi-head attention mechanism, which allows the model to focus on different feature subspaces simultaneously. In practice, each attention head independently computes a set of attention weights and generates a set of node feature updates. Then, the output feature vectors of all heads are concatenated or averaged, and then fused through a linear transformation layer. This design improves the model's expressiveness and robustness, enabling it to capture more complex dependencies.

[0038] In some embodiments, the calculation of attention weights can incorporate explicit edge feature encoding, which uses edge attributes as additional input. In specific implementations, in addition to node features and edge weights, edge feature vectors can also be included in the attention score calculation. For example, edge features can be concatenated with node features and input together into the attention calculation function. This allows the attention weights to be based not only on the feature similarity between nodes but also on the specific physical characteristics of the connecting edges, thus more accurately modeling electromagnetic coupling effects. Graph normalization layers or residual connections can be added to multi-layer attention calculations. These layers stabilize the training process and prevent gradient vanishing. In specific implementations, after each layer of graph attention operations, graph normalization can be applied to the updated node features. Graph normalization is similar to batch normalization but designed for graph-structured data; it standardizes features based on node degree or neighbor information. Simultaneously, residual connections directly add features from the previous layer to the current layer output, ensuring that the network does not lose fundamental information as it deepens. These techniques facilitate the training of deeper graph attention networks.

[0039] In practice, fusing the enhanced electromagnetic field distribution with the original spatiotemporal graph data sequence is a multi-step data processing flow. First, the enhanced electromagnetic field distribution needs interpolation. Since the enhanced electromagnetic field distribution is generated based on the discrete locations of graph nodes, while the original spatiotemporal graph data sequence may have higher spatial resolution or different grid points, the interpolation process uses spatial interpolation algorithms such as bilinear interpolation or inverse distance weighted interpolation to estimate the discrete node field strength values ​​onto each spatial coordinate point of the original data sequence, ensuring consistency in spatial resolution between the two. Next, the differences between the enhanced electromagnetic field distribution and the original spatiotemporal graph data sequence at corresponding points are calculated. These differences are typically residuals obtained by subtracting points one by one. All these residuals form a residual map, which highlights inconsistencies between the original data and the network enhancement results, potentially corresponding to noise, anomalies, or physical phenomena that the model has not fully learned. Then, the residual map and the original spatiotemporal graph data sequence are weighted and superimposed. The weighted superposition uses an adjustable weight parameter λ (0 ≤ λ ≤ 1) to control the contribution of the residual map. The fused data can be expressed as: Fused Data = Original Data + λ × Residual Map. The weight λ can be dynamically adjusted according to the confidence level of the residual map or the application scenario. For example, when the residuals are small and uniform, λ can be set small to reduce introduced disturbances. The fused data is then smoothed using a smoothing filter to eliminate high-frequency noise that may be introduced by the interpolation or weighting process. Linear filters such as Gaussian filters or nonlinear filters such as median filters are used for smoothing. The size of the filter kernel is selected based on the noise characteristics and the required smoothness, ensuring that the final fused data retains the authenticity of the original data while incorporating the enhanced information brought by the graph attention network.

[0040] Optionally, an adversarial training strategy can be introduced when generating the enhanced electromagnetic field distribution. This strategy aims to improve the physical plausibility of the generated distribution. In practice, a discriminator network is introduced while training the graph attention network. The discriminator network attempts to distinguish the generated enhanced electromagnetic field distribution from the actual electromagnetic field measurement distribution, while the graph attention network (as the generator) strives to generate a distribution that can deceive the discriminator. This adversarial process drives the generator to produce a field distribution that is more in line with physical laws, thereby indirectly improving the quality of the fused data.

[0041] Example 4: In specific implementation, the process of outputting a high-fidelity simulation signal by solving the electromagnetic field equations begins with establishing a complete electromagnetic field mathematical model. The core of the electromagnetic field mathematical model is Maxwell's equations, which contain four partial differential equations describing the fundamental laws of electric and magnetic fields: Ampere's circuital law describing the generation of magnetism from electromagnetism, Faraday's law of electromagnetic induction describing the generation of electromagnetism from magnetism, Gauss's law describing electric flux, and Gauss's law of magnetism describing the absence of magnetic fields. Besides Maxwell's equations themselves, the boundary conditions of the solution domain must be clearly defined. Boundary conditions describe the behavior of field quantities at the boundaries of the computational domain. For example, the boundary of an ideal conductor requires the tangential component of the electric field to be zero, while the radiation boundary simulates the propagation of energy to infinity. Using the fused data obtained in the previous step as the initial conditions for solving the electromagnetic field equations means that at the initial moment, the spatial distribution of field quantities such as electric field strength and magnetic induction intensity within the entire computational domain is directly assigned by the fused data, providing a starting point for transient solutions. At the same time, physical constraints need to be defined to limit the solution space and make it conform to the basic physical laws. The physical constraints mainly include the law of conservation of energy and the charge continuity equation. The law of conservation of energy requires that the rate of change of the total energy of the system equals the input power minus the power loss. The charge continuity equation describes the relationship between the change of charge density with time and the divergence of current density, ensuring that charge will not be created or disappear out of thin air.

[0042] In practical implementation, iteratively solving the electromagnetic field equations using optimization algorithms is key to achieving numerical solutions. These algorithms are based on the gradient descent method. They construct a loss function for the field quantities, which measures the degree to which the current solution satisfies Maxwell's equations and boundary conditions. The gradient of the loss function with respect to the field quantities is then calculated, and the field quantity values ​​are updated along the gradient descent direction to gradually reduce the residuals. Commonly used algorithms include steepest descent, conjugate gradient method, or more advanced adaptive moment estimators (AMEs). The iterative process begins with an initial guess. In each iteration, the residuals under the current field quantity distribution are first calculated. Then, the field quantity values ​​across the entire computational grid are updated according to the step size and direction determined by the optimization algorithm. The update process needs to consider physical constraints, such as projecting the solution after each update to ensure it satisfies the charge continuity equation. The iteration continues until the preset convergence condition is met. The convergence condition is usually set to the L2 norm of the residual being less than a very small positive threshold or the change in the solution between two consecutive iterations being sufficiently small. When the solution converges, the final electromagnetic field distribution solution is output as a high-fidelity simulation signal. This signal contains detailed electromagnetic information about the evolution of the system under given initial and boundary conditions.

[0043] In practical implementation, post-processing of simulation results is a crucial step in improving their usability. Post-processing operations mainly include time-domain resampling and spatial-domain interpolation. Time-domain resampling is performed on the output time-series simulation signal. Because the time step of numerical solutions can be very small to meet stability requirements, the data volume is enormous. Time-domain resampling resamples the simulation signal to a lower frequency that matches the actual observation or analysis needs, such as resampling from microsecond-level sampling to millisecond-level sampling. Resampling algorithms typically employ anti-aliasing filtering before downsampling to prevent frequency aliasing distortion. Spatial-domain interpolation transforms the simulation results at discrete grid points to any spatial location of interest to the user. For example, when obtaining the field distribution along a specific path, spatial-domain interpolation uses methods such as bilinear interpolation and cubic spline interpolation to estimate the field strength value at the target point based on the values ​​of surrounding grid points, filling in the gaps between grid points.

[0044] In practical implementation, referring to Table 1, comparing the post-processing results with the actual operating data and calculating the error index is the core step in verifying the simulation accuracy. The actual operating data comes from the voltage, current, or electromagnetic field strength data measured by the sensors deployed in the system during actual operation. The comparison process needs to ensure that the simulated signal and the real signal are strictly aligned in time and space. The error index is used to quantify the difference between the two. Commonly used error indices include root mean square error (RMSE), which measures the overall magnitude of the deviation; mean absolute error (MAE), which is less sensitive to outliers; and the coefficient of determination (COP). The error index reflects the degree to which the simulation results interpret changes in real data. The calculated error index is compared to a pre-set threshold, which is determined based on the accuracy requirements of the simulation application. For example, for protection system design, the threshold may be set very strictly. When the calculated error index exceeds the threshold, it indicates that the current simulation results fail to meet the accuracy requirements, and certain parameters or conditions in the simulation process need to be adjusted. Adjustments may include hyperparameters of the graph attention network or weights of physical constraints. After adjustment, the subsequent simulation process, starting from data fusion, needs to be re-executed to generate new simulation results and perform error evaluation again. This iterative optimization process continues until the error index falls below the threshold, meeting the preset accuracy requirements, thereby ensuring the reliability of the simulation output.

[0045] Table 1: Commonly Used Optimization Algorithms and Their Characteristics in Solving Electromagnetic Field Equations

[0046] Optionally, when defining physical constraints, the Lagrange multiplier method can be introduced for stricter constraints. The Lagrange multiplier method introduces the constraints as an additional term into the original loss function, forming an augmented Lagrange function. In practice, for equality constraints such as the charge continuity equation, they can be multiplied by a Lagrange multiplier and added to the loss function formed by the residuals of Maxwell's equations. Then, optimization is performed simultaneously on the field quantities and the Lagrange multiplier. This method can more strictly guarantee that the solution satisfies the physical constraints, and is especially suitable for scenarios with complex or mandatory constraints.

[0047] Optionally, in the spatial domain interpolation stage, a physically based neural network can be used as the interpolator. This network takes spatial coordinates as input and electromagnetic field quantities as output, incorporating the residuals of the physical equations as a penalty term in the training loss. In practice, a fully connected neural network can be trained to learn the mapping relationship from spatial coordinates to electromagnetic field quantities. The training data comes from simulation results on a discrete grid. Furthermore, the loss function requires the network's output to approximately satisfy Maxwell's equations. This results in interpolation results that are not only smooth and continuous but also possess greater physical consistency, outperforming traditional purely mathematical interpolation methods.

[0048] See Figure 4 This study demonstrates the changes in root mean square error (RMSE) and mean absolute error (MAE) at different stages during the online simulation of an LCC DC transmission electromechanical hybrid system. From the raw data to the final simulation result, the error rate shows a significant decreasing trend with the sequential implementation of wavelet decomposition, GAT enhancement, and data fusion. In the raw data stage, the RMSE reaches 8.5% and the MAE reaches 6.8%; after wavelet decomposition, the RMSE decreases to 5.2% and the MAE to 4.1%; after GAT enhancement, the RMSE further decreases to 3.1% and the MAE to 2.5%; after data fusion, the RMSE is 1.8% and the MAE is 1.5%; and in the final simulation result, the RMSE is only 0.9% and the MAE is only 0.8%. This change clearly demonstrates the effectiveness of the multi-stage processing flow in improving simulation accuracy. Each optimization measure significantly reduces the error, ultimately achieving high-fidelity simulation output and providing reliable support for accurate system modeling and analysis.

[0049] Example 5: In specific implementation, the core of the online learning mechanism lies in constructing a closed-loop model adaptive optimization system. This system dynamically updates the parameters of the graph attention network by continuously collecting historical data generated during the simulation process. Historical data includes input data such as spatiotemporal graph data sequences and output results such as high-fidelity simulation signals. A rolling window strategy is used to collect historical data. This strategy defines a fixed-length time window, for example, retaining the simulation data from the most recent 30 days. When new data is generated, the oldest data is removed, ensuring that the training dataset always reflects the latest operating state of the system. When constructing the training dataset, it is necessary to cover various operating scenarios to improve the model's generalization ability. Operating scenarios include normal operating conditions, fault operating conditions, and transition processes, such as the system switching from light load to heavy load. The construction process of the training dataset also includes data augmentation techniques. Data augmentation techniques increase data diversity and quantity by performing supervised transformations on the original historical data. Common techniques include adding Gaussian noise to simulate measurement errors and slightly scaling or shifting the spatiotemporal graph data sequences along the time axis to simulate dynamic response changes. Parameter updates utilize the backpropagation algorithm, which is based on gradient descent. It calculates the gradient of the loss function with respect to the graph attention network parameters and adjusts the parameter values ​​in the reverse direction of the gradient to minimize the loss. The loss function is typically chosen as mean squared error or cross-entropy loss, used to measure the difference between the model output and the true signal. The frequency of periodic retraining of the graph attention network needs to be set according to the system's rate of change. The system rate of change refers to how quickly the system topology, equipment parameters, or operating environment undergo significant changes. For example, for systems with frequent load fluctuations, the retraining frequency might be set to once a day; for stable systems, the frequency might be reduced to once a week. Retraining trigger conditions can be based on performance monitoring results. Performance monitoring calculates an error index by comparing the simulation output with real-time sensor data. When the error index exceeds a preset threshold, the retraining process is automatically triggered. The updated graph attention network model is applied to subsequent simulation tasks to achieve smooth switching of model parameters, ensuring that the simulation system can adapt to changes in system operating conditions, such as parameter drift caused by equipment aging or seasonal load fluctuations.

[0050] In practical implementation, the historical data collection process needs to be integrated into the simulation pipeline. Each time the simulation pipeline completes a full simulation calculation, it packages and stores the input and output data of that simulation task, along with timestamps and runtime scenario labels, into a distributed database or file system. The implementation of the rolling window strategy relies on an efficient data management module. This module periodically scans the storage system, automatically cleaning up old data exceeding the window period based on the data's timestamp attributes, and maintains a data index to accelerate data retrieval during training. When constructing the training dataset, data batches from different scenarios are randomly sampled from the data retained in the rolling window. Data augmentation operations are performed in real-time after the data batches are loaded into memory. For example, for each spatiotemporal graph data sequence sample, Gaussian white noise with a signal-to-noise ratio of 30dB is added with a certain probability, or a random translation with a delay of no more than 10 milliseconds is performed along the time dimension. The parameter update process is executed on a dedicated model training server. The training server loads the current version of the graph attention network model and the latest training dataset, using the backpropagation algorithm for multiple rounds of iterative training until the loss function on the validation set converges, finally generating the updated model parameter file. The retraining trigger condition is continuously evaluated by an independent monitoring service. The monitoring service subscribes to the real-time running data stream and simulation output stream, calculates the average error within the sliding window at regular intervals, and immediately sends a retraining request to the training server when the error exceeds the threshold for multiple consecutive time intervals.

[0051] In some embodiments, the time window length of a rolling window strategy can be dynamically adjusted based on data value. This dynamic adjustment mechanism is based on the data's information entropy or freshness weight. In specific implementations, each data point can be assigned an importance score, calculated based on the rarity of the corresponding operational scenario or the magnitude of the error. For example, fault data is generally more valuable than normal data, and the window length can be appropriately extended to retain more fault samples. Simultaneously, changes in data distribution are monitored; if the recent data distribution differs significantly from historical data distributions, the window length is automatically reduced to adapt to the new distribution more quickly. Data augmentation techniques can introduce transformations based on physical laws, ensuring that the augmented data still conforms to the fundamental equations of electromagnetic fields. In specific implementations, besides adding noise and time shifting, affine transformations in the spatial domain can be performed on the spatiotemporal map data sequence, such as simulating small sensor position shifts. However, the transformation parameters must be constrained within a physically reasonable range to avoid generating unrealistic data. Another augmentation method is to use generative adversarial networks to generate synthetic data, but a discriminator is needed to ensure the physical plausibility of the synthetic data.

[0052] Optionally, the parameter update process can employ incremental learning, which avoids training from scratch, thus saving computational resources. In practice, when retraining is triggered, the model is not retrained using the entire training dataset, but rather fine-tuned using only the most recently collected batch of data. A small learning rate is used during fine-tuning to prevent the model from forgetting previously learned knowledge. Simultaneously, elastic weight consolidation methods can be combined to apply regularization penalties to important parameters, protecting existing knowledge. Retraining trigger conditions can incorporate multi-source indicators, including model uncertainty estimates or external system events. In practice, in addition to threshold judgment based on output error, the rate of change of attention weights within the graph attention network can be monitored. A drastic change in the attention pattern may indicate that the model is no longer adapted to the current system state. Furthermore, event notifications from the power grid dispatching system, such as planned equipment maintenance or load switching, can be integrated as auxiliary signals to trigger retraining.

[0053] Optionally, the model application process can include a shadow mode phase. This shadow mode phase is used to verify the performance of the new model without immediately affecting the online simulation. In practice, after the updated graph attention network model is trained, it does not directly replace the old model in the production environment. Instead, it is deployed in a parallel environment, processing the same input data simultaneously. However, the output is only used for comparison and evaluation with the old model's output and real data. Only when the new model runs in shadow mode for a period of time and its performance is consistently superior to the old model is the formal model switch performed. It is understood that the overall efficiency of the online learning mechanism depends on the rational planning of computing and storage resources. Resource planning needs to balance data retention, training frequency, and real-time requirements. In practice, a cloud-native architecture can be used to elastically allocate resources, such as automatically expanding storage capacity during peak data collection periods and dynamically allocating GPU computing instances when training tasks are triggered. Simultaneously, historical data can be stored in layers, storing frequently accessed hot data on high-speed storage media and archiving cold data in low-cost storage to optimize overall cost-effectiveness.

[0054] See Figure 5This chart illustrates the changes in model accuracy (%) and relative processing time under five data augmentation methods: original data, Gaussian noise augmentation, temporal translation augmentation, spatial affine transformation, and GAN-synthesized data. The blue bars represent model accuracy (%), and the red line graph represents relative processing time. In terms of model accuracy, GAN-synthesized data achieved the highest accuracy, exceeding 94%; temporal translation augmentation was second, at approximately 93%; spatial affine transformation accuracy was above 92%; original data accuracy was approximately 92.3%; and Gaussian noise augmentation accuracy was slightly lower than original data, at approximately 91.8%. In terms of relative processing time, the relative processing time for original data was 1.0 (which can be used as a baseline); Gaussian noise augmentation slightly increased the processing time to approximately 1.2; temporal translation augmentation was slightly lower than Gaussian noise augmentation, at approximately 1.05; spatial affine transformation processing time increased to 1.6; and GAN-synthesized data had the highest relative processing time, exceeding 2.2. Overall, different data augmentation methods improved model accuracy while having varying impacts on processing time. For example, while GAN-synthesized data can significantly improve model accuracy, it also incurs higher time costs; whereas time-shift augmentation, while maintaining a certain level of accuracy, offers relatively controllable processing time. This difference provides a reference for balancing model performance and efficiency when selecting data augmentation methods.

[0055] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0056] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An online simulation method for an LCC DC transmission electro-electromagnetic hybrid system, characterized in that, The method is implemented through the following processing flow: Construct a graph structure model of the LCC DC transmission electromechanical hybrid system, where nodes represent electrical components and edges represent electromagnetic coupling relationships, and define node attributes and edge weights; Real-time running data is collected and mapped to nodes and edges of a graph structure model to form a spatiotemporal graph data sequence, which includes timestamps and spatial coordinate information. Multi-resolution wavelet decomposition was performed on the spatiotemporal graph data sequence to extract electromagnetic feature modes at different scales, including low-frequency contour modes and high-frequency detail modes. Based on electromagnetic feature modes, a graph attention network is used to calculate the dynamic weights between nodes to generate an enhanced electromagnetic field distribution, which integrates local and global features. The enhanced electromagnetic field distribution is fused with the original spatiotemporal map data sequence, and a high-fidelity simulation signal is output by solving the electromagnetic field equations.

2. The online simulation method for the LCC DC transmission electro-electromagnetic hybrid system according to claim 1, characterized in that, The graphical structure model for constructing the LCC DC transmission electro-electromagnetic hybrid system specifically includes: Identify the key electrical components in the LCC DC transmission electro-electromagnetic hybrid system, including converter valves, transformers, and transmission lines, and abstract each component as a graph node; The connection edges between nodes are defined based on electromagnetic coupling relationships, and the connection edges include electrical connections and electromagnetic induction relationships. Assign an attribute vector to each node, the attribute vector including electrical parameters and physical parameters; A weight value is calculated for each edge, the weight value being determined based on the electromagnetic coupling strength and the distance attenuation factor; Combine nodes and edges into a graph structure model and store graph topology information.

3. The online simulation method for the LCC DC transmission electromechanical hybrid system according to claim 2, characterized in that, The process of collecting real-time operational data and forming a spatiotemporal graph data sequence includes: Real-time operational data is acquired from a sensor network, including voltage, current, and electromagnetic field strength. Sort real-time running data by timestamp to generate time series data; Map time series data to corresponding nodes in the graph structure model and update the node attribute vectors; Data from adjacent nodes is aggregated based on edge weights to form a data stream along the edge. By combining timestamps and spatial coordinates, a spatiotemporal graph data sequence is constructed, which is in the form of a multidimensional tensor.

4. The online simulation method for the LCC DC transmission electromechanical hybrid system according to claim 1, characterized in that, The step of performing multi-resolution wavelet decomposition on the spatiotemporal graph data sequence to extract electromagnetic feature modes includes: By selecting wavelet basis functions, wavelet transform is performed on the spatiotemporal graph data sequence to obtain coefficients at multiple resolution levels. Low-frequency subband coefficients are extracted from the coefficients to reconstruct the low-frequency profile modes, representing the macroscopic variation trend of the electromagnetic field; High-frequency subband coefficients are extracted from the coefficients to reconstruct high-frequency detail modes, representing the rapid fluctuation characteristics of the electromagnetic field; The low-frequency profile mode and the high-frequency detail mode are normalized to eliminate dimensional differences; The normalized modes are combined into a set of electromagnetic characteristic modes.

5. The online simulation method for the LCC DC transmission electro-electromagnetic hybrid system according to claim 4, characterized in that, The method of generating enhanced electromagnetic field distribution using graph attention networks includes: The set of electromagnetic feature modes is input into a graph attention network, which contains a multi-layer attention mechanism. Calculate the attention weights of each node and its neighboring nodes, whereby the attention weights are based on node attributes and edge weights; The feature representation of each node is updated by weighted aggregation of the features of neighboring nodes; Repeated multi-layer attention calculations are performed to obtain enhanced node feature vectors. These enhanced node feature vectors are then mapped back to the physical space to generate an enhanced electromagnetic field distribution.

6. The online simulation method for the LCC DC transmission electro-electromagnetic hybrid system according to claim 5, characterized in that, The process of fusing the enhanced electromagnetic field distribution with the original spatiotemporal map data sequence includes: The enhanced electromagnetic field distribution is interpolated to match the resolution of the original spatiotemporal map data sequence. The difference between the enhanced electromagnetic field distribution and the original data is calculated to generate a residual map. The residual map is then weighted and superimposed with the original spatiotemporal map data sequence to obtain fused data. The fused data is then smoothed and filtered to eliminate noise during the fusion process.

7. The online simulation method for the LCC DC transmission electro-electromagnetic hybrid system according to claim 1, characterized in that, The process of solving the electromagnetic field equations includes: Establish the electromagnetic field equations, which include Maxwell's equations and boundary conditions; Using the fused data as initial conditions, we substitute it into the electromagnetic field equations and define physical constraints, including energy conservation and charge continuity. The electromagnetic field equations are solved iteratively using an optimization algorithm based on the gradient descent method. When the solution converges, the electromagnetic field distribution solution is output as the simulation result.

8. The online simulation method for the LCC DC transmission electro-electromagnetic hybrid system according to claim 7, characterized in that, After outputting the high-fidelity simulation signal, the following is also included: The simulation results are post-processed, including time-domain resampling and spatial-domain interpolation. The post-processed results are compared with the actual running data, and the error index is calculated. When the error index exceeds the threshold, the parameters or physical constraints of the graph attention network are adjusted, and the simulation process is re-executed until the error index meets the requirements.

9. The online simulation method for the LCC DC transmission electro-electromagnetic hybrid system according to claim 1, characterized in that, The method also includes an online learning mechanism: Collect historical data during the simulation process, including input data and output results, construct a training dataset, use it to update the model parameters of the graph attention network, retrain the graph attention network periodically to adapt to changes in the system's operating state, and apply the updated model to subsequent simulation tasks.

Images