Secondary equipment electromagnetic risk assessment method and system based on T-GAT network

Abstract

The invention discloses a T-GAT network-based secondary equipment electromagnetic risk assessment method and system, and the method comprises the steps: taking a corresponding operation of primary equipment as an interference source, taking a transient overvoltage signal corresponding to the interference source as an excitation signal, carrying out the finite element simulation of a series compensation platform, and carrying out the calculation of an electromagnetic risk. The transient overvoltage and electromagnetic field intensity of the interference source, the secondary equipment and the related equipment are obtained; an interference source, secondary equipment and related equipment are used as nodes to construct a graph structure, node features and edge features are extracted, an edge weight is calculated according to the similarity of two nodes with edges connected, the edge weight of the edge connected with the interference source is corrected in real time, and the edge weight of the edge connected with the interference source is corrected in real time. All the corrected weights are recombined into a weight matrix, and symmetric normalization processing is carried out to obtain an adjacent matrix; and inputting the node features, the edge features and the adjacency matrix into a time sequence diagram attention network, calculating a risk value, and performing risk level assessment according to the risk value. According to the method, information fusion of two dimensions of space and time is realized, and the accuracy of risk level assessment is improved.

Description

Technical Field

[0001] This invention belongs to the field of electromagnetic interference protection technology for power systems, and more specifically, relates to a method and system for electromagnetic risk assessment of secondary equipment based on T-GAT networks. Background Technology

[0002] In ultra-high voltage (UHV) power transmission systems, series compensation devices (referred to as series compensation platforms) are typically configured in AC transmission lines to achieve efficient transmission of large-capacity power over long distances, thereby enhancing the transmission capacity and dynamic stability of the lines. During operation, the series compensation platform involves the switching of several key components, such as spark gaps, reactors, and parallel capacitors. These components generate high-intensity transient electromagnetic interference (EMI) signals during operation or breakdown. These EMI signals can be transmitted to various secondary devices within the platform via spatial radiation and coupling paths.

[0003] Against this backdrop, there is an urgent need to develop a risk assessment method for secondary equipment that is suitable for the electromagnetic environment characteristics of UHV series compensation platforms. However, existing technologies are insufficient to meet the requirements of accuracy, dynamism, and interpretability of risk assessment for secondary equipment in complex electromagnetic environments of UHV series compensation platforms. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a method and system for electromagnetic risk assessment of secondary equipment based on T-GAT networks.

[0005] The present invention adopts the following technical solution.

[0006] The first aspect of this invention proposes a method for assessing the electromagnetic risk of secondary equipment based on T-GAT networks, comprising: The transient overvoltage signal of the primary equipment when it performs the corresponding operation is obtained by modeling and simulating the series compensation platform. The corresponding operation of the primary equipment is used as the interference source, and the transient overvoltage signal corresponding to the interference source is used as the excitation signal. Finite element simulation is performed on the series compensation platform to obtain the transient overvoltage and electromagnetic field intensity at the locations of the interference source, the secondary equipment of the series compensation platform and related equipment. The interference source, the secondary equipment of the serial compensation platform and its related equipment are used as nodes to construct a graph structure. Node features and edge features are extracted, and the edge weights are calculated based on the similarity between the two nodes connected by the edge. The edge weights of the edges connected to the interference source are corrected in real time. All the corrected weights are combined into a weight matrix and then symmetrically normalized to obtain the adjacency matrix. The node features, edge features, and adjacency matrix are input into the pre-trained temporal graph attention network. The temporal graph attention network calculates the risk value through graph attention aggregation and temporal attention aggregation, and assesses the risk level based on the risk value.

[0007] Preferably, the step of constructing a graph structure using the interference source, the secondary equipment of the series compensation platform, and related equipment as nodes, and extracting node features and edge features, specifically involves: Secondary equipment includes bypass GAP control box and platform measurement equipment box. Interference sources include spark gap triggering and disconnect switch operation. The location of the interference source is the location of the corresponding primary equipment. Other equipment includes shielded cables, busbars and grounding nodes of CT and secondary equipment. The node features include the transient overvoltage measured at the corresponding node and the electromagnetic field strength at the corresponding node location obtained from finite element simulation; and each node feature has a fixed code representing time information. If two nodes are physically connected or connected on the electromagnetic interference propagation path, then an edge is established between the two corresponding nodes. The edge features include the electrical coupling strength, spatial distance, and historical operation and maintenance relationship between the two corresponding nodes. The historical operation counts of the two nodes are added together and then subtracted from the historical common operation counts of the two nodes to obtain the historical operation and maintenance relationship.

[0008] Preferably, the step of calculating the edge weight based on the similarity between the two nodes connected by the edge specifically involves: Calculate the wavelet correlation coefficient between two nodes at each time point within the previously set short-time moving average window from the current time point. Calculate the sum of all wavelet correlation coefficients and divide it by the number of time points in the set short-time East China moving average window to obtain the wavelet coherence. The wavelet correlation coefficient between two nodes at each time step is obtained by: obtaining the wavelet transform coefficients of the overvoltage signals of the two nodes at the corresponding time step on a set scale, and calculating the similarity between the two wavelet transform coefficients as the wavelet correlation coefficient. The inherent electromagnetic coupling tendency between two nodes is defined. The inherent electromagnetic coupling tendency and the wavelet coherence of the two nodes are weighted and summed according to the set weights to obtain the basic coupling strength. The basic coupling strength is multiplied by the long-term degradation factor as the edge weight.

[0009] Preferably, the long-term degradation factor specifically includes: Calculate the difference between the current time and the system commissioning time, divide it by the set degradation rate, subtract the set curve midpoint parameter, and calculate the sigmoid function value of the subtraction result. Adding 1 to the set maximum degradation magnitude and multiplying it by the sigmoid function value yields the long-term degradation factor.

[0010] Preferably, for interference sources in the spark gap triggering operation, the edge weights of the edges connecting to the interference sources are corrected in real time, specifically as follows: For interference sources in spark gap triggering operation, at the moment of spark triggering, the corrected edge weight of the edge connecting the corresponding interference source is the edge weight calculated based on similarity plus a set peak value; then the weight within the spark triggering cycle is the edge weight calculated based on similarity plus the set peak value multiplied by the structure factor and then multiplied by the two-stage exponential decay coefficient. The two-stage exponential decay coefficient is the output calculated by inputting the current time into the two-stage exponential decay model. The structure factor is 1 plus the reciprocal of the degree of the corresponding interference source, where the degree of the corresponding interference source is the number of nodes connected to the corresponding interference source.

[0011] Preferably, for interference sources caused by the operation of the disconnecting switch, the edge weights of the edges connecting to the interference sources are corrected in real time, specifically as follows: For interference sources of disconnecting switch operation, during the duration of disconnecting switch operation, the corrected edge weight of the edge connecting the corresponding interference source is the edge weight calculated based on similarity plus the set operation intensity amplitude multiplied by the structure factor, then multiplied by the damping decay time term, and then multiplied by the interference pulse term; the set operation intensity amplitude is less than the set peak value. The damping decay time term is calculated by dividing the difference between the negative current time and the start time of the disconnecting switch operation by the set damping decay time constant as the exponent, and then calculating the natural base e to the power of the exponent. The interference pulse term is calculated by multiplying the arc reignition oscillation frequency during disconnection switch operation by 2π and multiplying the difference between the current time and the start time of disconnection switch operation as the phase, and then calculating the absolute value of the phase cos term.

[0012] Preferably, the temporal graph attention network is divided into an input layer, a first graph attention aggregation layer, a second graph attention aggregation layer, a temporal attention aggregation layer, and an output layer; The node features are input into the input layer; the first and second graph attention aggregation layers both map the output and edge features of the previous layer to obtain the attention score between each node and its neighboring nodes through the attention scoring function; the neighboring nodes are nodes that have edges with the corresponding node; The attention score is then normalized by combining it with the adjacency matrix.

[0013] Preferably, the result of uniformly mapping the attention scoring function is normalized by combining it with the adjacency matrix, specifically as follows: The attention score between node i and its neighbor node j is used as the exponent, and the natural base e is raised to the power of the exponent as the exponential attention score between node i and its neighbor node j. Divide the exponential attention score between node i and its neighbor node j by the sum of the exponential attention scores between node i and all its neighbors to obtain the exponential normalized attention score between node i and its neighbor node j. Multiply the exponentially normalized attention score between node i and its neighbor node j by the element in the i-th row and j-th column of the adjacency matrix to obtain the weighted attention score. Divide the weighted attention score between node i and its neighbor node j by the sum of the weighted attention scores between node i and all its neighbor nodes to obtain the final normalized attention between node i and its neighbor node j.

[0014] Preferably, the risk level assessment based on the risk value specifically includes: Set a first threshold and a second threshold. If the risk value is greater than or equal to 0 and less than the first threshold, it is a low-risk level; if the risk value is greater than or equal to the first threshold and less than the second threshold, it is a medium-risk level; if the risk value is greater than or equal to the second threshold, it is a high-risk level.

[0015] The second aspect of this invention proposes an electromagnetic risk assessment system for secondary equipment based on a T-GAT network, using the method described in the first aspect of this invention. The system includes a simulation module, an adjacency matrix calculation module, and a risk assessment module, specifically: Simulation module: Model and simulate the series compensation platform to obtain the transient overvoltage signal of the primary equipment when it performs the corresponding operation. Take the corresponding operation of the primary equipment as the interference source and the transient overvoltage signal corresponding to the interference source as the excitation signal. Perform finite element simulation on the series compensation platform to obtain the transient overvoltage and electromagnetic field intensity at the locations of the interference source, the secondary equipment of the series compensation platform and its related equipment. The adjacency matrix calculation module constructs a graph structure by using the interference source, the secondary device of the serial complement platform and its related devices as nodes, extracts node features and edge features, calculates the edge weights based on the similarity between the two nodes connected by the edge, and corrects the edge weights of the edges connected to the interference source in real time. All the corrected weights are combined into a weight matrix and then symmetrically normalized to obtain the adjacency matrix. Risk assessment module: Node features, edge features, and adjacency matrix are input into the pre-trained temporal graph attention network. The temporal graph attention network calculates the risk value through graph attention aggregation and temporal attention aggregation, and assesses the risk level based on the risk value.

[0016] A third aspect of the present invention provides an apparatus comprising a processor and a storage medium, characterized in that: the storage medium is used to store instructions; and the processor is used to operate according to the instructions to perform the steps of the method described in the first aspect of the invention.

[0017] A fourth aspect of the invention provides a computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the steps of the method described in the first aspect of the invention.

[0018] The beneficial effects of this invention are as follows: Compared with the prior art, this invention abstracts the series compensation platform into a dynamically evolving graph structure. The edge weights are dynamically calculated by fusing wavelet coherence within a short time window and degradation factors reflecting the long-term aging effects of the equipment. This accurately captures the time-frequency correlation characteristics of transient processes while considering long-term degradation factors. Furthermore, this invention designs a real-time weight adjustment formula for two types of interference sources: spark gap triggering and disconnector operation. This formula, based on nonlinear modeling driven by their physical mechanisms, greatly improves the accuracy of describing complex transient processes. This invention constructs an adjacency matrix through edge weights. When using a time-series graph attention network for risk assessment, it not only fuses the node features of nodes and their neighbors but also introduces edge features as physical priors. Additionally, it uses the adjacency matrix for normalization, ensuring the physical rationality and interpretability of the risk assessment. Ultimately, the time-series graph attention network achieves information fusion in both spatial and temporal dimensions, improving the accuracy of risk level assessment. Attached Figure Description

[0019] Figure 1 This is a simulation diagram of a primary system overvoltage signal; Figure 2 Finite element model of the UHV series compensation platform; Figure 3 This is a schematic diagram of the structure of an ultra-high voltage series compensation platform equipment. Figure 4 Risk assessment framework for secondary equipment of the T-GAT network UHV series compensation platform. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of this invention. The embodiments described in this application are merely some embodiments of this invention, and not all embodiments. Based on the spirit of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of this invention.

[0021] Embodiment 1 of the present invention proposes a method for electromagnetic risk assessment of secondary equipment based on T-GAT networks, including: The transient overvoltage signal of the primary equipment when it performs the corresponding operation is obtained by modeling and simulating the series compensation platform. The corresponding operation of the primary equipment is used as the interference source, and the transient overvoltage signal corresponding to the interference source is used as the excitation signal. Finite element simulation is performed on the series compensation platform to obtain the transient overvoltage and electromagnetic field intensity at the locations of the interference source, the secondary equipment of the series compensation platform and related equipment. The interference source, the secondary equipment of the serial compensation platform and its related equipment are used as nodes to construct a graph structure. Node features and edge features are extracted, and the edge weights are calculated based on the similarity between the two nodes connected by the edge. The edge weights of the edges connected to the interference source are corrected in real time. All the corrected weights are combined into a weight matrix and then symmetrically normalized to obtain the adjacency matrix. The node features, edge features, and adjacency matrix are input into the pre-trained temporal graph attention network. The temporal graph attention network calculates the risk value through graph attention aggregation and temporal attention aggregation, and assesses the risk level based on the risk value.

[0022] It should be noted that the electromagnetic environment in the UHV series compensation platform is extremely complex. Transient overvoltages generated by primary equipment operations such as spark gap breakdown and disconnector operation can reach amplitudes several times the rated voltage, with rise times less than 1 μs, pulse widths of only a few hundred nanoseconds, and a spectrum covering tens of kHz to tens of MHz. These overvoltages exhibit multi-pulse characteristics and complex spatial distribution, generating strong electromagnetic interference in the secondary equipment area. Due to these complex characteristics, the electromagnetic signals of the primary and secondary equipment on the series compensation high-voltage platform cannot be directly measured. Therefore, it is necessary to construct a simulation model and combine it with on-site transient overvoltage waveform data to analyze the characteristic distribution on the platform and the risk of interference to the secondary equipment.

[0023] Primary equipment includes an MOV (Metal Oxide Valves), spark gap, disconnector, and series capacitor bank, etc., and adopts a wideband equivalent circuit model (the MOV can be equivalently represented by a voltage-controlled current source; the spark gap can be equivalently represented by a capacitor and a time-varying resistor in parallel), and the series disconnector can be equivalently represented by a time-varying resistor, a lossless transmission line, and a capacitance to ground in parallel. The series capacitor bank and damping circuit are represented by lumped elements in ATP-EMTP.

[0024] Based on the above design, a simulation model of the UHV series compensation platform is established, such as... Figure 1 The setup uses spark gap triggering or disconnector switch operation as excitation sources. Multiple measurement points (such as bus end, platform access point, and secondary equipment interface) are set up between the low-voltage bus and the high-potential platform to acquire fast transient overvoltages at different locations. U distributed.

[0025] Transient overvoltage signal UAs the excitation signal input, it is used for finite element simulation of the series-compensated high-voltage platform to analyze the electric field distribution and intensity in the secondary equipment area and on the PCB board surface. The platform structure was simplified as necessary in the simulation, retaining the electrical paths between the busbar, grounding, and equipment to ensure that transient overvoltage excitation can be transmitted to the secondary equipment through a reasonable path. The secondary equipment is abstracted into an equivalent electrical model, simplifying internal circuit details and ensuring the electrical response characteristics of the ports. Geometrically, the actual relative layout is maintained to ensure that electromagnetic interference is reflected in the simulation.

[0026] like Figure 2 As shown, the series compensation platform is modeled and simulated using the aforementioned modeling method. In the model, the transient overvoltage signal generated by the primary system is equivalent to an electromagnetic excitation source, which propagates to the secondary equipment location through the platform's main conductors and spatial coupling paths. By extracting the port induced voltage and local electric field intensity, the interference effects experienced by the secondary equipment can be effectively reproduced, achieving equivalent simulation and analysis under simplified conditions. However, due to the large amount of data and high feature dimensionality generated by the simulation, it is difficult to draw effective conclusions by direct processing. It is necessary to introduce graph structure models and T-GAT network methods to realize rapid risk assessment of secondary equipment of UHV series compensation platform.

[0027] In this preferred embodiment, the step of constructing a graph structure using the interference source, the secondary device of the serial compensation platform, and related devices as nodes, and extracting node features and edge features, specifically involves: Secondary equipment includes bypass GAP control box and platform measurement equipment box. Interference sources include spark gap triggering and disconnect switch operation. The location of the interference source is the location of the corresponding primary equipment. Other equipment includes shielded cables, busbars and grounding nodes of CT and secondary equipment. The node features include the transient overvoltage measured at the corresponding node and the electromagnetic field strength at the corresponding node location obtained from finite element simulation; and each node feature has a fixed code representing time information. If two nodes are physically connected or connected on the electromagnetic interference propagation path, then an edge is established between the two corresponding nodes. The edge features include the electrical coupling strength, spatial distance, and historical operation and maintenance relationship between the two corresponding nodes. The historical operation counts of the two nodes are added together and then subtracted from the historical common operation counts of the two nodes to obtain the historical operation and maintenance relationship.

[0028] It should be noted that the electrical coupling strength is the ratio of the transient overvoltage value of one node to the transient overvoltage value of another node when a transient overvoltage signal exists in the simulation. The historical operation and maintenance relationship formula is:

[0029] in, The historical operation and maintenance relationship between node i and its neighbor node j, wherein the neighbor node is a node that has an edge with the corresponding node; The number of times node i and its neighbor node j have run together in history; , The nodes i and their neighbor j represent the historical number of runs.

[0030] Specifically, the graph structure constructed in this embodiment is as follows: Figure 3 As shown.

[0031] It should be noted that the fixed encoding is:

[0032] in, This represents the time offset corresponding to the current sampling time of the node; for The corresponding node currently uses the time. The encoding; , This represents the current sampling time of the node. The set time window length; For the k-th angular frequency of the encoding, , This represents the total number of sampling points set within the time window.

[0033] The time window length and sampling rate are set according to the overvoltage signal spectrum width on the platform. In this embodiment, the time window length is set as follows: The window type uses Hanning, ensuring that it can capture single transient pulses without sacrificing too much frequency resolution, while effectively reducing spectral leakage, and the sampling rate... To ensure coverage of interference frequencies, therefore the number of sampling points... .

[0034] In this preferred embodiment, the step of calculating the edge weight based on the similarity between the two nodes connected by the edge specifically involves: Calculate the wavelet correlation coefficient between two nodes at each time point within the previously set short-time moving average window from the current time point. Calculate the sum of all wavelet correlation coefficients and divide it by the number of time points in the set short-time East China moving average window to obtain the wavelet coherence. The wavelet correlation coefficient between two nodes at each time step is obtained by: obtaining the wavelet transform coefficients of the overvoltage signals of the two nodes at the corresponding time step on a set scale, and calculating the similarity between the two wavelet transform coefficients as the wavelet correlation coefficient. The inherent electromagnetic coupling tendency between two nodes is defined. The inherent electromagnetic coupling tendency and the wavelet coherence of the two nodes are weighted and summed according to the set weights to obtain the basic coupling strength. The basic coupling strength is multiplied by the long-term degradation factor as the edge weight.

[0035] Specifically, edge weight The formula is:

[0036] in, The weights are set; The inherent electromagnetic coupling tendency between node i and its adjacent node j is defined. It represents wavelet coherence and reflects the instantaneous correlation of electromagnetic signals between nodes in real time; The long-term degradation factor is used; the wavelet coherence calculation formula is:

[0037] in, The set sliding interval; It is a short-time moving average window used to smooth transient coherence; Indicates time; As a scale; For at any time and scale The wavelet correlation coefficient is given by the following formula:

[0038] in, Indicates at time and scale Above, the wavelet transform coefficients of the transient overvoltage signal at node i; Indicates at time and scale Above, the wavelet transform coefficients of the transient overvoltage signal at node i; It is a norm; express and .

[0039] In this preferred embodiment, the long-term degradation factor is specifically: Calculate the difference between the current time and the system commissioning time, divide it by the set degradation rate, subtract the set curve midpoint parameter, and calculate the sigmoid function value of the subtraction result. Adding 1 to the set maximum degradation magnitude and multiplying it by the sigmoid function value yields the long-term degradation factor.

[0040] Specifically, long-term degradation factors The formula is:

[0041] in, This represents the maximum degree of degradation; This refers to the system's commissioning time; The rate of degradation; The parameters for the midpoint of the curve are set.

[0042] Preferably, for interference sources in the spark gap triggering operation, the edge weights of the edges connecting to the interference sources are corrected in real time, specifically as follows: For interference sources in spark gap triggering operation, at the moment of spark triggering, the corrected edge weight of the edge connecting the corresponding interference source is the edge weight calculated based on similarity plus a set peak value; then the weight within the spark triggering cycle is the edge weight calculated based on similarity plus the set peak value multiplied by the structure factor and then multiplied by the two-stage exponential decay coefficient. The two-stage exponential decay coefficient is the output calculated by inputting the current time into the two-stage exponential decay model. The structure factor is 1 plus the reciprocal of the degree of the corresponding interference source, where the degree of the corresponding interference source is the number of nodes connected to the corresponding interference source.

[0043] For interference sources in spark gap triggering operations, the corrected edge weights for:

[0044] in, For the set peak value, Indicates the degree of interference source; Indicates the corresponding interference source; This is the two-stage exponential decay coefficient; Specifically, the two-stage exponential decay model is as follows:

[0045] in, It is the fast decay time constant, which serves as the rapid decay time of the interference pulse; The slow decay time constant is set to 1ms. A weighting factor of 0.7 indicates that the fast decay component dominates in the early stages of decay. The moment when the spark is triggered.

[0046] It should be noted that after the spark is triggered, the intensity of electromagnetic interference will decay over time and space. Therefore, this embodiment uses a two-stage exponential decay model to accurately simulate this physical process, which more accurately reflects the oscillation decay waveform of actual electromagnetic interference than simple linear or single exponential decay.

[0047] In this preferred embodiment, for interference sources caused by the operation of the disconnecting switch, the edge weights of the edges connecting to the interference sources are corrected in real time, specifically as follows: For interference sources of disconnecting switch operation, during the duration of disconnecting switch operation, the corrected edge weight of the edge connecting the corresponding interference source is the edge weight calculated based on similarity plus the set operation intensity amplitude multiplied by the structure factor, then multiplied by the damping decay time term, and then multiplied by the interference pulse term; the set operation intensity amplitude is less than the set peak value. The damping decay time term is calculated by dividing the difference between the negative current time and the start time of the disconnecting switch operation by the set damping decay time constant as the exponent, and then calculating the natural base e to the power of the exponent. The interference pulse term is calculated by multiplying the arc reignition oscillation frequency during disconnection switch operation by 2π and multiplying the difference between the current time and the start time of disconnection switch operation as the phase, and then calculating the absolute value of the phase cos term.

[0048] For interference sources in disconnector switch operation, the corrected edge weights The formula is:

[0049] in, The operating intensity amplitude is set to 1.8, which is lower than the peak value set for spark gap triggering. This is the start time of the disconnector switch operation; The damping decay time constant has a value of 50ms; The arc reignition oscillation frequency during disconnection switch operation; This represents the interference pulse for each reignition.

[0050] It should be noted that, for the interference source of the disconnecting switch operation, the disconnecting switch operation is a process accompanied by arc reignition and lasts for tens to hundreds of milliseconds. Therefore, this embodiment regards it as a damped oscillation enhancement model to simulate its multiple arc reignition characteristics.

[0051] In this preferred embodiment, the time-series graph attention network is divided into an input layer ( ), First image attention aggregation layer ( The second image shows the attention aggregation layer. ), temporal attention aggregation layer ( ) and output layer ( );in Number of floors; The node features are input into the input layer, as shown in the formula:

[0052] in, The input layer outputs at time t for node i. This is the mapping of node features after passing through the input layer; Both the first and second graph attention aggregation layers map the output and edge features of the previous layer to obtain the attention score between each node and its neighboring nodes through an attention scoring function; the neighboring nodes are nodes that have edges with the corresponding node; and the attention scores are normalized by combining them with the adjacency matrix.

[0053] Specifically, the attention scoring function is as follows:

[0054] in, This indicates that node i and its neighbor node j are... The result of a unified mapping between the layer's output and edge features through an attention scoring function; This indicates that for node i at time t, the first... Layer output; This indicates that for neighbor node j at time t, the first... Layer output; Edge features; , They represent respectively to , Perform mapping, which is learnable and obtained through training; The scoring matrix is ​​obtained through training; The linear rectifier function with leakage is given by the following formula:

[0055] in, Let be the independent variable of the linear rectifier function with leakage. The leakage coefficient is a trainable parameter that is learned and optimized during model training using backpropagation and gradient descent algorithms. In this embodiment, it is given an initial value. And train the best in the overall evaluation. value; In this preferred embodiment, the normalization of the result after unified mapping of the attention scoring function combined with the adjacency matrix is ​​specifically as follows: The attention score between node i and its neighbor node j is used as the exponent, and the natural base e is raised to the power of the exponent as the exponential attention score between node i and its neighbor node j. The exponentially normalized attention score between node i and its neighbor node j is calculated by dividing the exponentially normalized attention score between node i and all its neighbors by the sum of their exponentially normalized attention scores. The formula is:

[0056] in, Let i be the set of all neighboring nodes of node i. To represent the relationship between node i and its neighbor node z, The result of a unified mapping between the layer's output and edge features through an attention scoring function; The weighted attention score is calculated by multiplying the exponentially normalized attention score between node i and its neighbor node j by the element in the i-th row and j-th column of the adjacency matrix. The weighted attention score between node i and its neighbor node j is then divided by the sum of the attention scores between node i and all its neighbors. The sum of weighted attention scores is used as the final normalized attention between node i and its neighbor node j; the final normalized attention... The formula is:

[0057] in, The element in the i-th row and j-th column of the adjacency matrix Let be the element in the i-th row and z-th column of the adjacency matrix; The exponentially normalized attention score between node i and its neighbor node j. Let z be the exponentially normalized attention score between node i and its neighbor node z. The output of the corresponding layer is:

[0058] in, For the first Layer output; This is a nonlinear risk response function, designed to ensure that the model can learn and simulate complex, nonlinear risk accumulation and response mechanisms. It is a trainable linear transformation matrix used to map the feature vectors of neighboring device j; These are trainable parameters that represent the basic risk level of the node itself.

[0059] Specifically, the temporal attention aggregation layer uses temporal attention to achieve spatial aggregation across the temporal dimension, enabling risk accumulation modeling in the temporal dimension. The temporal attention aggregation layer first sets the current time... and time The output of the previous layer is uniformly mapped to obtain the attention scores for historical moments:

[0060] in, For the mapping result, As a learnable vector, it represents historical time. The degree of contribution to the current risks of secondary equipment; ; , , The current time step of the attention aggregation layer in the second graph is shown below. and time The output; , It is a trainable mapping that maps the representations of current time and historical time to the same space.

[0061] Attention scores at historical moments are normalized using the following formula:

[0062] in, Score the attention given to historical moments at time d. , This represents the time interval between two sampling points.

[0063] The output of the temporal attention aggregation layer is:

[0064] in, , This is the output of the attention aggregation layer at time d in the second graph; This is the output of the temporal attention aggregation layer; It should be noted that time aggregation can reflect the impact of transient events such as spark gaps and disconnect switches, as well as capture the cumulative increase in equipment risk level under continuous interference or multiple triggers.

[0065] The risk value output by the output layer is:

[0066] in, , The risk value output by the output layer. It is an activation function. and These are the learnable mapping and bias, respectively.

[0067] In this preferred embodiment, the risk level assessment based on the risk value specifically includes: Set a first threshold and a second threshold. If the risk value is greater than or equal to 0 and less than the first threshold, it is a low-risk level; if the risk value is greater than or equal to the first threshold and less than the second threshold, it is a medium-risk level; if the risk value is greater than or equal to the second threshold, it is a high-risk level.

[0068] Embodiment 3 of this invention proposes a secondary equipment electromagnetic risk assessment system based on a T-GAT network, using the method described in Embodiment 1 of this invention. The system includes a simulation module, an adjacency matrix calculation module, and a risk assessment module, specifically: Simulation module: Model and simulate the series compensation platform to obtain the transient overvoltage signal of the primary equipment when it performs the corresponding operation. Take the corresponding operation of the primary equipment as the interference source and the transient overvoltage signal corresponding to the interference source as the excitation signal. Perform finite element simulation on the series compensation platform to obtain the transient overvoltage and electromagnetic field intensity at the locations of the interference source, the secondary equipment of the series compensation platform and its related equipment. The adjacency matrix calculation module constructs a graph structure by using the interference source, the secondary device of the serial complement platform and its related devices as nodes, extracts node features and edge features, calculates the edge weights based on the similarity between the two nodes connected by the edge, and corrects the edge weights of the edges connected to the interference source in real time. All the corrected weights are combined into a weight matrix and then symmetrically normalized to obtain the adjacency matrix. Risk assessment module: Node features, edge features, and adjacency matrix are input into the pre-trained temporal graph attention network. The temporal graph attention network calculates the risk value through graph attention aggregation and temporal attention aggregation, and assesses the risk level based on the risk value.

[0069] A third aspect of the present invention provides an apparatus comprising a processor and a storage medium, characterized in that: the storage medium is used to store instructions; and the processor is used to operate according to the instructions to perform the steps of the method described in the first aspect of the invention.

[0070] Embodiment 4 of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described in Embodiment 1 of the present invention.

[0071] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the claims of the present invention.

Claims

1. A method for electromagnetic risk assessment of secondary equipment based on T-GAT networks, characterized in that, include: The transient overvoltage signal of the primary equipment when it performs the corresponding operation is obtained by modeling and simulating the series compensation platform. The corresponding operation of the primary equipment is used as the interference source, and the transient overvoltage signal corresponding to the interference source is used as the excitation signal. Finite element simulation is performed on the series compensation platform to obtain the transient overvoltage and electromagnetic field intensity at the locations of the interference source, the secondary equipment of the series compensation platform and related equipment. The interference source, the secondary equipment of the serial compensation platform and its related equipment are used as nodes to construct a graph structure. Node features and edge features are extracted, and the edge weights are calculated based on the similarity between the two nodes connected by the edge. The edge weights of the edges connected to the interference source are corrected in real time. All the corrected weights are combined into a weight matrix and then symmetrically normalized to obtain the adjacency matrix. The node features, edge features, and adjacency matrix are input into the pre-trained temporal graph attention network. The temporal graph attention network calculates the risk value through graph attention aggregation and temporal attention aggregation, and assesses the risk level based on the risk value.

2. The electromagnetic risk assessment method for secondary equipment based on T-GAT networks according to claim 1, characterized in that: The process of constructing a graph structure using interference sources, secondary devices of the serial compensation platform, and related devices as nodes, and extracting node and edge features, specifically involves: Secondary equipment includes bypass GAP control box and platform measurement equipment box. Interference sources include spark gap triggering and disconnect switch operation. The location of the interference source is the location of the corresponding primary equipment. Other equipment includes shielded cables, busbars and grounding nodes of CT and secondary equipment. The node features include the transient overvoltage measured at the corresponding node and the electromagnetic field strength at the corresponding node location obtained from finite element simulation; and each node feature has a fixed code representing time information. If two nodes are physically connected or connected on the electromagnetic interference propagation path, then an edge is established between the two corresponding nodes. The edge features include the electrical coupling strength, spatial distance, and historical operation and maintenance relationship between the two corresponding nodes. The historical operation counts of the two nodes are added together and then subtracted from the historical common operation counts of the two nodes to obtain the historical operation and maintenance relationship.

3. The electromagnetic risk assessment method for secondary equipment based on T-GAT networks according to claim 2, characterized in that: The calculation of edge weights based on the similarity between the two nodes connected by the edge is specifically as follows: Calculate the wavelet correlation coefficient between two nodes at each time point within the previously set short-time moving average window from the current time point. Calculate the sum of all wavelet correlation coefficients and divide it by the number of time points in the set short-time East China moving average window to obtain the wavelet coherence. The wavelet correlation coefficient between two nodes at each time step is obtained by: obtaining the wavelet transform coefficients of the overvoltage signals of the two nodes at the corresponding time step on a set scale, and calculating the similarity between the two wavelet transform coefficients as the wavelet correlation coefficient. The inherent electromagnetic coupling tendency between two nodes is defined. The inherent electromagnetic coupling tendency and the wavelet coherence of the two nodes are weighted and summed according to the set weights to obtain the basic coupling strength. The basic coupling strength is multiplied by the long-term degradation factor as the edge weight.

4. The electromagnetic risk assessment method for secondary equipment based on T-GAT networks according to claim 3, characterized in that: The long-term degradation factor is specifically: Calculate the difference between the current time and the system commissioning time, divide it by the set degradation rate, subtract the set curve midpoint parameter, and calculate the sigmoid function value of the subtraction result. Adding 1 to the set maximum degradation magnitude and multiplying it by the sigmoid function value yields the long-term degradation factor.

5. The electromagnetic risk assessment method for secondary equipment based on T-GAT networks according to claim 4, characterized in that: For interference sources in spark gap triggering operations, the edge weights of the edges connecting to the interference sources are corrected in real time, specifically as follows: For interference sources in spark gap triggering operation, at the moment of spark triggering, the corrected edge weight of the edge connecting the corresponding interference source is the edge weight calculated based on similarity plus a set peak value; then the weight within the spark triggering cycle is the edge weight calculated based on similarity plus the set peak value multiplied by the structure factor and then multiplied by the two-stage exponential decay coefficient. The two-stage exponential decay coefficient is the output calculated by inputting the current time into the two-stage exponential decay model. The structure factor is 1 plus the reciprocal of the degree of the corresponding interference source, where the degree of the corresponding interference source is the number of nodes connected to the corresponding interference source.

6. The electromagnetic risk assessment method for secondary equipment based on T-GAT networks according to claim 4, characterized in that: For interference sources caused by disconnecting switch operation, the edge weights of the edges connected to the interference sources are corrected in real time, specifically as follows: For the interference source of the disconnecting switch operation, during the duration of the disconnecting switch operation, the corrected edge weight of the edge connecting the corresponding interference source is the edge weight calculated based on similarity plus the set operation intensity amplitude multiplied by the structure factor, then multiplied by the damping decay time term, and then multiplied by the interference pulse term. The set operation intensity amplitude is less than the set peak value; The damping decay time term is calculated by dividing the difference between the negative current time and the start time of the disconnecting switch operation by the set damping decay time constant as the exponent, and then raising the natural base e to the power of the exponent. The interference pulse term is calculated by multiplying the arc reignition oscillation frequency during disconnection switch operation by 2π and multiplying the difference between the current time and the start time of disconnection switch operation as the phase, and then calculating the absolute value of the phase cos term.

7. The electromagnetic risk assessment method for secondary equipment based on T-GAT networks according to claim 1, characterized in that: The temporal graph attention network is divided into an input layer, a first graph attention aggregation layer, a second graph attention aggregation layer, a temporal attention aggregation layer, and an output layer. Input node features into the input layer; Both the first and second attention aggregation layers map the output and edge features of the previous layer to a unified attention score between each node and its neighboring nodes using an attention scoring function. The neighbor node is a node that has an edge with the corresponding node; The attention score is then normalized by combining it with the adjacency matrix.

8. The electromagnetic risk assessment method for secondary equipment based on T-GAT networks according to claim 7, characterized in that: The process of normalizing the result after uniformly mapping the attention scoring function with the adjacency matrix is ​​as follows: The attention score between node i and its neighbor node j is used as the exponent, and the natural base e is raised to the power of the exponent as the exponential attention score between node i and its neighbor node j. Divide the exponential attention score between node i and its neighbor node j by the sum of the exponential attention scores between node i and all its neighbors to obtain the exponential normalized attention score between node i and its neighbor node j. Multiply the exponentially normalized attention score between node i and its neighbor node j by the element in the i-th row and j-th column of the adjacency matrix to obtain the weighted attention score. Divide the weighted attention score between node i and its neighbor node j by the sum of the weighted attention scores between node i and all its neighbor nodes to obtain the final normalized attention between node i and its neighbor node j.

9. The electromagnetic risk assessment method for secondary equipment based on T-GAT networks according to claim 1, characterized in that: The risk level assessment based on the risk value is specifically as follows: Set a first threshold and a second threshold. If the first threshold is less than the second threshold, and the risk value is greater than or equal to 0 and less than the first threshold, it is considered a low-risk level. If the risk value is greater than or equal to the first threshold and less than the second threshold, it is classified as a medium risk level. If the risk value is greater than or equal to the second threshold, it is classified as a high-risk level.

10. A secondary equipment electromagnetic risk assessment system based on a T-GAT network according to the method of any one of claims 1-9, comprising a simulation module, an adjacency matrix calculation module, and a risk assessment module, characterized in that: Simulation module: Model and simulate the series compensation platform to obtain the transient overvoltage signal of the primary equipment when it performs the corresponding operation. Take the corresponding operation of the primary equipment as the interference source and the transient overvoltage signal corresponding to the interference source as the excitation signal. Perform finite element simulation on the series compensation platform to obtain the transient overvoltage and electromagnetic field intensity at the locations of the interference source, the secondary equipment of the series compensation platform and its related equipment. The adjacency matrix calculation module constructs a graph structure by using the interference source, the secondary device of the serial complement platform and its related devices as nodes, extracts node features and edge features, calculates the edge weights based on the similarity between the two nodes connected by the edge, and corrects the edge weights of the edges connected to the interference source in real time. All the corrected weights are combined into a weight matrix and then symmetrically normalized to obtain the adjacency matrix. Risk assessment module: Node features, edge features, and adjacency matrix are input into the pre-trained temporal graph attention network. The temporal graph attention network calculates the risk value through graph attention aggregation and temporal attention aggregation, and assesses the risk level based on the risk value.

11. A device comprising a processor and a storage medium, characterized in that: The storage medium is used to store instructions; the processor is used to operate according to the instructions to perform the steps of the method according to any one of claims 1-9.

12. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the steps of the method according to any one of claims 1-9.

Images