GIS field-electrical fusion real-time condition awareness and early warning system and method

Abstract

A GIS field-electrical fusion real-time condition awareness and early warning system and method, comprising an embedded synchronized signal acquisition and measurement unit (100) and an industrial personal computer (200), wherein a simulator (201), a partial discharge source locator (202), and an early warning device (203) are deployed on the industrial personal computer (200); when the embedded unit (100) detects, according to a strongly correlated feature, that an electromagnetic field signal within a GIS comprises a partial discharge signal, the embedded unit (100) obtains a full time-domain waveform plot comprising a time-domain waveform plot and a frequency-domain waveform plot of the partial discharge signal; the simulator (201) performs coarse positioning on a partial discharge source on the basis of the full time-domain waveform plot of the GIS partial discharge, and performs simulation-enhanced computation by using a coarse positioning result as an initial injection point, so as to obtain simulation-enhanced data for the partial discharge source; the partial discharge source locator (202) performs iterative searching on a measured discharge intensity calculated according to the full time-domain waveform plot and a simulation-enhanced result, so as to obtain an actual discharge location of the partial discharge source; and the early warning device (203) performs early warning on the basis of a full waveform in a time domain and early warning information comprising the actual discharge location of the partial discharge source.

Description

GIS Field-Electricity Fusion Real-time Status Perception and Early Warning System and Method Technical Field

[0001] This invention relates to the field of power equipment technology, specifically to a GIS-based field-electricity fusion real-time status perception and early warning system and method. Background Technology

[0002] The high performance and reliability of high-voltage gas-insulated switchgear (GIS) are crucial for ensuring the stable operation of power systems, improving grid transmission efficiency, and ensuring the security of power supply.

[0003] With the widespread application of GIS equipment, frequent equipment failures caused by insulation aging and component loosening have seriously threatened the safe operation of substations at all levels. Partial discharge, as a key indicator of GIS equipment defects, requires research into its location and measurement technologies. This is crucial for effectively sensing the status of GIS equipment and providing early warning analysis. Generally, this involves collecting and monitoring high-frequency electromagnetic signals (UHF) within the GIS system. Once a partial discharge signal is detected, the source of the partial discharge is located, thereby achieving status awareness and long-term analysis and early warning, ensuring the safe and stable operation of GIS equipment.

[0004] When using UHF signals for discharge detection, the multiple reflections of UHF signals in the complex topology of GIS, with different propagation paths, can affect the accuracy of positioning. The multipath effect can cause different time delays in the arrival of the signal at the UHF sensor, thus introducing positioning errors and making it impossible for the positioning system to accurately identify the location of the discharge source. Furthermore, the high-precision positioning models used in related technologies (such as fuzzy neural networks and deep learning models) have complex structures, and the training and optimization process requires a lot of computing resources, making them difficult to apply in real time.

[0005] To this end, a method for rapidly iteratively searching for the optimal matching position by fusing UHF signal measurements with simulation models is proposed. However, the technical difficulties are as follows: (1) Most existing partial discharge signal acquisition devices use PRPD (Phase Resolved Partial Discharge) discharge spectra and other forms to transmit compressed partial discharge signals to the host computer. They cannot restore the high-frequency, complete full-time waveform partial discharge point time domain signal detected by the host computer, which cannot meet the need for accurate positioning of partial discharge signals inside GIS. In addition, existing UHF signal acquisition requires high sampling rate data acquisition equipment. Since the sampling rate is proportional to the positioning accuracy, improving the sampling accuracy will inevitably face problems such as insufficient chip computing resources. The high sampling rate brings a large amount of data, which increases the burden of processing and transmission. Transmission delay and data synchronization problems will affect real-time performance. Therefore, there is a lack of high-speed and high-performance real-time acquisition / matching / detection methods for partial discharge signals in related technologies.

[0006] (2) Traditional numerical methods face problems such as complex coupling between coarse and fine grids and difficulty in establishing boundary conditions when performing electromagnetic field non-uniform grid simulation. At the same time, non-uniform grid division simulation requires high computational resources. When the number of grid divisions and the density of grids are high, the calculation time is long and the simulation speed is slow. This will seriously affect the detection time when dealing with high-frequency GIS partial discharge detection problems. If coarse grid simulation is used, although the time cost can be reduced, the simulation accuracy is low, which will affect the quality of partial discharge detection. Therefore, there is a lack of fast and high-precision electromagnetic field distribution simulation technology applicable to multi-structure GIS equipment in related technologies. Electromagnetic field distribution simulation is difficult to provide a theoretical basis for high-performance real-time sensing of GIS.

[0007] (3) Since real-time measured UHF signals usually need to be preprocessed and cleaned to reduce the impact of noise and extract effective signals as much as possible, while the input data quality and clarity required for fast electromagnetic field simulation are usually higher than the actual measurement data, it is difficult to match the simulation results with the measured data, and fast electromagnetic field simulation cannot be organically combined with noisy measurement data. Summary of the Invention

[0008] The technical problem to be solved by this invention is how to achieve high-performance real-time status perception and early warning based on GIS field-electricity fusion.

[0009] This invention solves the above-mentioned technical problems through the following technical means: This invention proposes a GIS field-electricity fusion real-time status perception and early warning system. The system includes a signal synchronous acquisition and measurement embedded device and an industrial control computer. The industrial control computer deploys a simulator, a location searcher, and an early warning device. The signal synchronous acquisition and measurement embedded device is used to monitor the electromagnetic field signals inside the GIS for partial discharge signals based on pre-selected strongly correlated features, obtaining a full-time-domain waveform diagram containing both time-domain and frequency-domain waveforms of the partial discharge signals. The simulator is used to coarsely locate the partial discharge source based on the full-time-domain waveform diagram of the GIS partial discharge, and uses the coarse location result as the initial injection point for simulation enhancement calculation to obtain simulation enhancement data of the partial discharge source. The location searcher is used to iteratively search the measured discharge intensity calculated based on the full-time-domain waveform diagram and the simulation enhancement result to obtain the actual discharge location of the partial discharge source. The early warning device is used to issue an early warning based on the full-time-domain waveform and early warning information containing the actual discharge location of the partial discharge source.

[0010] Furthermore, the embedded device for synchronous signal acquisition and measurement includes: a signal acquisition location determination module for determining the installation location of each sensor in the signal acquisition module; a signal acquisition module for acquiring electromagnetic field signals inside the GIS; a signal processing front end for conditioning the electromagnetic field signals transmitted by the signal acquisition module, and transmitting the conditioned signal to an analog-to-digital converter; an analog-to-digital converter for converting the conditioned signal from an analog signal to a digital signal and transmitting it to a signal measurement module; and a signal measurement module for monitoring the digital signal based on pre-selected strong correlation features, and when a partial discharge signal is detected, obtaining a full-time-domain waveform diagram including a time-domain waveform diagram and a frequency-domain waveform diagram of the partial discharge signal, wherein the strong correlation features include time-domain strong correlation features and frequency-domain strong correlation features.

[0011] Furthermore, the embedded device for synchronous signal acquisition and measurement also includes a feature filtering module, which is used to filter the time-frequency domain features used to detect partial discharge signals as nodes of a graph neural network to obtain the time-domain strong correlation features and the frequency-domain strong correlation features, wherein the lines between nodes represent the relationship between the features represented by the nodes.

[0012] Furthermore, the feature selection module includes: a feature map construction unit, used to construct a feature map from a complete feature map G = (V, E) of a graph neural network, where V is the set of nodes and E is the set of edges, and let x... v x co[v] h ne[v] x ne[v]Let h represent the feature vector of node v, the feature vector of the edge, the state vector of node v and its surrounding nodes, and the feature vector of the surrounding nodes of node v, respectively. The aggregation update unit is used to aggregate and update the information of the input node and edge according to the local transition function for updating the node state, and outputs the node label, expressed by the formula: h v =f(x) v ,x co[v] h ne[v] x ne[v] )

[0013] o v =g(h . v x v )

[0014] In the formula, f(·) is the local transition function for updating the node state, g(·) is the local output function, and h v o is the state vector learned through iterative learning by the graph neural network. v Node labels; Iteration units are used to calculate the association between a node and its neighboring nodes in the continuous iteration of the graph neural network, so as to obtain strong correlation features in the time domain and strong correlation features in the frequency domain.

[0015] Furthermore, the iterative unit is specifically used for: Let H, O, X and X N These are vectors constructed by superimposing the state vectors of all nodes, the output labels of all nodes, the feature vectors of all nodes and edges, and the feature vectors of all nodes, respectively. The formula can be written in a more compact form as: O = G(H, X) N )

[0016] H t+1 =F(H t (X)

[0017] Among them, H t H represents the t-th iteration of H. t+1 =F(H t X) represents the state vector of all nodes in the (t+1)th iteration obtained by passing the feature vector and the state vector of the tth iteration through the global transfer function. F(·) and G(·) are the global transfer function and the global output function, respectively, which are the stack of the local transfer function f(·) and the local output function g(·) of all nodes. During the iteration process, nodes with similar states, nodes with complementary states, and the influence of each node on the overall graph neural network are determined. One node with similar states is selected, nodes with complementary states are merged into one node, and the node with the greatest influence on the overall graph neural network is selected as the strongly correlated feature.

[0018] Further, the signal measurement module includes: a buffer unit for buffering the digital signal; a first monitoring unit for extracting the strong time-domain correlation features from the digital signal, determining the detected partial discharge signal based on the strong time-domain correlation features, and outputting a first time-domain waveform during discharge; a second monitoring unit for extracting the strong frequency-domain correlation features from the digital signal, determining the detected partial discharge signal based on the strong frequency-domain correlation features, and outputting a frequency-domain waveform during discharge; and a waveform synthesis unit for obtaining a full-time-domain waveform of the partial discharge signal based on the first time-domain waveform and the frequency-domain waveform.

[0019] Furthermore, the second monitoring unit includes: a Fourier transform subunit, used to perform at least two fast Fourier transforms on the electromagnetic field signal to obtain frequency domain data of at least two frequency bands; and a normalization subunit, used to extract the strong correlation features of the frequency domain data of each frequency band, and after analyzing and matching the strong correlation features, determine that when the frequency domain data of each frequency band exceeds the second discharge threshold, normalize the frequency domain waveforms corresponding to each frequency band to obtain the frequency domain waveform diagram during discharge.

[0020] Furthermore, the waveform synthesis unit includes: a waveform conversion subunit, used to convert the frequency domain waveform diagram into a second time domain waveform diagram during discharge; and a waveform synthesis subunit, used to synthesize the first time domain waveform diagram and the second time domain waveform diagram to obtain a full time domain waveform diagram of the partial discharge signal.

[0021] Furthermore, the simulator includes: a simulation module, used to use the coarse location result of the partial discharge power source obtained based on the full-time-domain waveform diagram as the initial injection point of the 3D simulation model of the GIS equipment to simulate the partial discharge phenomenon of the GIS equipment, and perform simulation calculations to obtain coarse mesh simulation results; and an enhancement module, used to enhance the coarse mesh simulation results using an enhancement model to obtain enhanced simulation data.

[0022] Furthermore, the simulation module includes: a spatial discretization unit, used to discretize the three-dimensional simulation model of the GIS equipment into a spatial grid, and determine the correspondence between the number of each grid node and the spatial coordinate data; and a differential simulation unit, used to inject the coarse positioning result as the initial injection point into the corresponding spatial grid to simulate the partial discharge phenomenon of the GIS equipment, convert the Maxwell equations used to describe the electromagnetic field of GIS into a fourth-order matrix, and use the four-step HIE-FDTD algorithm to solve the fourth-order matrix to calculate the field strength data of each grid node.

[0023] Furthermore, the differential simulation unit includes: a matrix reconstruction subunit, used to convert Maxwell's equations describing the electromagnetic field of GIS into a sixth-order matrix form as follows:

[0024] In the formula, Let [M] be a vector composed of the components of the electric and magnetic fields in a rectangular coordinate system, and let [M] be a sixth-order matrix. The matrix transformation subunit is used to convert the sixth-order matrix form into a fourth-order rectangular form:

[0025] In the formula, [A H ]for [B H ]for The sub-element is used to solve the fourth-order matrix form using the four-step HIE-FDTD algorithm, and to calculate the field strength data for each grid node.

[0026] Further, the solution sub-unit is used to perform the following steps: The four-step HIE-FDTD algorithm is decomposed in the time domain into four sub-steps: n→n+1 / 4, n+1 / 4→n+2 / 4, n+2 / 4→n+3 / 4, and n+3 / 4→n+1. For each sub-step, a semi-implicit difference scheme is used to calculate the tridiagonal implicit expression for the field strength data of the grid nodes; The chasing method is used to solve the tridiagonal implicit expression to obtain the field strength data of the grid nodes.

[0027] Furthermore, the enhancement model employs a differential deep learning network model. The coarse-grid simulation results include coarse-grid structure data and coarse-grid field strength data. The differential deep learning network model comprises an enhancement network and a structural similarity network connected in sequence. The enhancement network includes a self-adjusting module and a differential convolution module. The self-adjusting module and the differential convolution module are used to calculate the coarse-grid structure data and the coarse-grid field strength data, respectively, to obtain the fine-grid enhanced grid structure features and the fine-grid enhanced field strength features. The structural similarity network is used to match the fine-grid enhanced grid structure features and the fine-grid enhanced field strength features to calculate similarity features. Based on the similarity features, the simulation enhancement data is calculated.

[0028] Furthermore, the self-adjusting module includes a first convolutional layer and a second convolutional layer connected in sequence. The output features of the first convolutional layer and the output features of the second convolutional layer are output to the activation function layer after a first addition operation. The coarse grid structure data is used as the input of the first convolutional layer, the grid size supplementary information of the coarse grid structure data is used as the input of the second convolutional layer, the deviation vector of the coarse grid structure data is used as the input of the first addition operation, and the coarse grid structure data and the grid size supplementary information of the coarse grid structure data are output to the first addition operation through residual connection.

[0029] Furthermore, the differential convolution module includes a differential convolutional layer, a size integration layer, a self-attention mechanism layer, and a third convolutional layer connected in sequence, with an activation function connected after the third convolutional layer; the coarse grid field strength data is used as the input of the differential convolutional layer, which is used to calculate field strength features of different sizes of the coarse grid field strength data using a differential algorithm; the size integration layer is used to sum the field strength features of different sizes calculated by the differential convolutional layer to obtain the renormalized field strength features.

[0030] Furthermore, the convolution kernel of the differential convolutional layer adopts any one of the following: five-point differential convolution kernel, weighted differential convolution kernel, multi-scale differential convolution kernel, directional differential convolution kernel, nine-point differential convolution kernel, and hybrid mode differential convolution.

[0031] Furthermore, the structural similarity network includes a first branch network, a second branch network, a second addition operation, and a first multilayer perceptron. The outputs of the first branch network and the second branch network are both connected to the second addition operation, and the output of the second addition operation is connected to the first multilayer perceptron. The multilayer perceptron is followed by an activation function.

[0032] Furthermore, the first branch network includes a convolutional neural network layer (CNN) and a batch regularization operation connected in sequence, wherein an activation function is followed by the regularization operation; the second branch network includes a second multilayer perceptron, wherein an activation function is followed by the second multilayer perceptron.

[0033] Further, the step of matching the fine-mesh enhanced mesh structural features and the fine-mesh enhanced field strength features using the structural similarity network to calculate similarity features includes: extracting structural feature information of the fine-mesh enhanced mesh structural features using the first branch network, expressed by the formula:

[0034] In the formula, The feature information for enhancing the mesh structure features of the fine mesh described in the k-th example. To enhance the mesh structure features of the fine mesh, CNN() is a stacked convolution operation, BatchNorm() is a batch regularization operation, and ReLU is the activation function; the second branch network is used to extract the field strength feature information of the fine mesh enhanced field strength features, expressed by the formula:

[0035] In the formula, This refers to the feature information of the fine-mesh enhanced field strength feature described in the k-th example. For the fine-mesh enhanced field strength feature, MLP() is the operation performed by the second multilayer perceptron; using the feature information of the fine-mesh enhanced mesh structure feature and the modulation weights corresponding to the feature information of the fine-mesh enhanced field strength feature, the field strength and structure combined feature is calculated, expressed by the formula:

[0036] In the formula, fk ，combined For the field strength and structure combined features, α and β represent the modulation weights corresponding to the feature information of the fine mesh enhanced mesh structure features and the feature information of the fine mesh enhanced field strength features, respectively; based on the field strength and structure combined features, the similarity features are calculated, expressed by the formula: f k，final =ReLU(MLP) final (fk, combined))

[0037] In the formula, f k，final For the aforementioned similarity features, MLP final () represents the operation performed by the first multilayer perceptron.

[0038] Further, the step of calculating the simulation enhancement data based on the similarity features includes: calculating, respectively, regularized coarse mesh structure data, regularized coarse mesh field strength data, regularized fine mesh enhanced mesh structure features, and regularized fine mesh enhanced field strength features based on the coarse mesh structure data, the coarse mesh field strength data, the fine mesh enhanced mesh structure features, and the fine mesh enhanced field strength features; calculating mesh structure consistency based on the regularized coarse mesh structure data and the regularized fine mesh enhanced mesh structure features; calculating mesh field strength consistency based on the regularized coarse mesh field strength data and the regularized fine mesh enhanced field strength features; calculating mesh field strength loss compensation based on the mesh structure consistency and the mesh field strength consistency; calculating fine mesh enhanced structure data based on the fine mesh enhanced mesh structure features, the mesh field strength loss compensation, and the similarity features; and calculating the simulation enhancement data based on the fine mesh enhanced field strength features, the mesh field strength loss compensation, and the similarity features.

[0039] Furthermore, the simulation enhancement data includes enhanced discharge location and enhanced discharge intensity. The location searcher includes: a state transition model construction module, which, based on the enhanced discharge location and enhanced discharge intensity at each time step, uses a Markov decision model to reconstruct the state transition model corresponding to the process of exploring the actual discharge location of the local discharge source using the coarse localization result of the local discharge source as the center of the ant colony algorithm; a heuristic spatial parameterization module, which uses a heuristic learner based on a graph neural network to generate a heuristic metric, converting the state transition model into a location exploration model that is affected by the heuristic metric and requires graph traversal; and an iterative search module, which uses a neural-guided perturbation interleaved local search algorithm to iteratively search the location exploration model to obtain the actual discharge location of the local discharge source.

[0040] Furthermore, the state transition model construction module is used to: construct the state space corresponding to the actual discharge location of the partial discharge source in the simulator, based on the simulated discharge location and simulated discharge intensity at each time step.

[0041]

[0042] In the formula: p represents the state corresponding to the local discharge power node i discovered at time t in the simulator. t E represents the simulated discharge position at time t. t This represents the simulated discharge intensity at time t. The initial state is defined as the initial discharge intensity E and initial discharge position P of the partial discharge source obtained from the full-time-domain waveform diagram of partial discharge based on GIS, and T represents the exploration period, t = 1, 2, ..., T; The action space corresponding to the exploration of the actual discharge position of the partial discharge source in the simulator is constructed.

[0043] In the formula: This represents the action at time t, where the action is selected using an ε-greedy strategy. The state transition model corresponding to the process of the ant colony algorithm exploring the actual discharge location of the partial discharge source with the initial discharge location as the center is as follows:

[0044] In the formula: Indicates the action The probability of transitioning from discharge node i to discharge node j. τ represents the state corresponding to node j of the partial discharge source discovered at time t+1 in the simulator. ijη represents the pheromone concentration corresponding to the transition from node i to node j. i x represents the heuristic function corresponding to the transition from node i to node j, τ im This indicates a transition from node i to the node included in the allowed list. S The pheromone concentration corresponding to node m in the diagram, η im This indicates a transition from node i to the node included in the allowed list. S The heuristic function corresponding to node m in the diagram, where α and β represent the pheromone concentration and the weights of the heuristic function, respectively. S This represents the set of non-uniform mesh local discharge power nodes that can be selected in the next time step.

[0045] Furthermore, the reward function for the transition from node i to node j is: E′ t E represents the measured discharge intensity obtained by processing the full time-domain waveform acquired by the sensor at time t. t The simulated partial discharge intensity at time t is further represented by the heuristic spatial parameterization module, which is used to convert the edge features of the l-th layer of the graph neural network, which are connected by the i-th node and the j-th node. Mapping to the heuristic metric η θ The state transition model is then converted into a location exploration model that is influenced by heuristic metrics and requires T-step graph traversal:

[0046] In the formula: This represents a location exploration model. This indicates that the state corresponding to the local discharge power node i is discovered at time t in the simulator. This represents the state corresponding to the local discharge power node i at time t+1 in the simulator, where T represents the exploration period, and t = 1, 2, ..., T.

[0047] Furthermore, the location searcher also includes a training module for: training a graph neural network-based heuristic learner using a gradient strategy, wherein the objective function used during training is... for:

[0048] In the formula: This represents the objective function used to explore the actual discharge location of a partial discharge source using a neurally guided perturbation-interleaved local search algorithm, where W represents the equilibrium... and The parameters, In the heuristic metric η θ The expected value of the actual discharge location of the partial discharge source under the influence of [the following factors] Let f(·) represent the state space corresponding to the actual discharge location of the partial discharge source, and let f(·) represent the objective function; where the objective function is... The gradient is The formula is expressed as:

[0049] In the formula: This represents the average target value for directly exploring the actual discharge location of the partial discharge source. This represents the average target value used in a local search algorithm with neurally guided perturbation interleaving to explore the actual discharge location of the local discharge source. In the heuristic metric η θ The gradient of the actual discharge location of the partial discharge source is explored under the influence of the gradient; when the maximum number of iterations is reached... Or the maximum number of iterations has not been reached. But the objective function At that time, the training ended, among which, This represents the minimum threshold corresponding to the objective function.

[0050] Further, the iterative search module is configured as follows: a local search unit, used to iteratively search the location exploration model using a local search algorithm to obtain a local optimum; a perturbation unit, used to perform neural-guided perturbation on the local optimum obtained in the current iteration search, and obtain the optimal exploration scheme for the actual discharge location of the local discharge power source in the current iteration through interleaved local searches with neural-guided perturbation; a pheromone concentration update unit, used to update the pheromone concentration between each local discharge power source node after all local discharge power source nodes have been selected; and an exploration unit, used to determine the optimal exploration scheme for the actual discharge location of the local discharge power source when the iterative search reaches the iterative convergence condition, and to achieve matching of the actual discharge location of the local discharge power source based on the optimal exploration scheme.

[0051] Furthermore, the process by which the perturbation unit explores the optimal exploration scheme for the actual discharge position of the current iterative partial discharge power source is expressed by the following formula:

[0052] In the formula: This represents the optimal exploration scheme for the actual discharge position of the local discharge source in the current iteration, obtained by neural-guided perturbation of the local optimal solution. This is a locally optimal solution; T p η represents the number of moves in the disturbance. θ This represents a heuristic metric.

[0053] Furthermore, the early warning device is equipped with a fault classification module, which includes: a deep feature extraction network to obtain deep features corresponding to the UHF signal, wherein the deep feature extraction network includes a feature extraction network and an output network, the feature extraction network is formed by stacking several feature extraction layers, and each feature extraction layer includes a channel attention module and a spatial attention module connected in sequence; a reinforcement learning unit to construct a dataset using the deep features, and to train a GIS partial discharge feature matching Markov model using a deep reinforcement learning algorithm to obtain the optimal GIS partial discharge feature matching result; and an early warning unit to generate early warning information based on the optimal GIS partial discharge feature matching result and the actual discharge location of the partial discharge source, and to provide fault early warning.

[0054] Furthermore, this invention proposes a GIS field-electricity fusion real-time status perception and early warning method, used to realize GIS fault early warning using the GIS field-electricity fusion real-time status perception and early warning system as described above. The method includes: when monitoring the electromagnetic field signal inside the GIS containing partial discharge signals based on pre-screened strongly correlated features, obtaining a full-time domain waveform map containing time-domain waveform maps and frequency-domain waveform maps of the partial discharge signals; coarsely locating the partial discharge source based on the full-time domain waveform map of the GIS partial discharge, and using the coarse location result as the initial injection point for simulation enhancement calculation to obtain simulation enhancement data of the partial discharge source; iteratively searching the measured discharge intensity calculated based on the full-time domain waveform map and the simulation enhancement result to obtain the actual discharge location of the partial discharge source; and issuing an early warning based on the full-time domain waveform map and the early warning information containing the actual discharge location of the partial discharge source.

[0055] The advantages of this invention are: (1) This invention detects the digital signal converted from electromagnetic field signal based on the pre-screened strong correlation features as indicators, and when it is determined that a partial discharge signal is detected, it realizes the acquisition of a high-precision full-time-domain waveform of GIS partial discharge; then, based on the full-time-domain waveform of GIS partial discharge, it performs coarse positioning of the partial discharge source, injects the coarse positioning result into the simulator, and uses it to simulate the partial discharge phenomenon of GIS equipment and perform simulation enhancement calculation of the simulation enhancement data of the partial discharge source, so as to realize fast and high-precision GIS electromagnetic field simulation; finally, iterative search is performed based on the real-time acquired full-time-domain waveform and simulation enhancement data to obtain the actual discharge position of the partial discharge source; based on the actual discharge position and other information, early warning is realized. Therefore, this invention realizes high-performance real-time status perception and early warning based on GIS field-electric fusion.

[0056] (2) This invention substitutes various time-frequency domain features of electromagnetic field signals into a graph neural network, uses each time-frequency domain feature as a node of the graph neural network, and the connection between nodes represents the correlation between the features represented by the nodes. The correlation strength is filtered through the iteration of the graph neural network to obtain the feature with the highest correlation, i.e. the strongly correlated feature. Then, the partial discharge signal is detected based on the strongly correlated feature obtained by the filtering. Although the amount of computation is reduced, the accuracy of partial discharge signal detection is not affected because the correlation between the indicators is considered. Therefore, this invention improves the accuracy of partial discharge signal detection while greatly reducing the amount of computation.

[0057] (3) Since GIS electromagnetic field simulation focuses more on the accurate simulation of the detailed interaction between equipment and electromagnetic field, this invention designs and constructs a differential deep learning network model. This differential deep learning network model uses convolution combined with differential algorithm to process coarse grid field strength data, which can accurately simulate the electromagnetic field distribution of GIS, especially when dealing with complex geometric structures and boundary conditions. Furthermore, the deep learning network is used to optimize the simulation process, improve the simulation accuracy, and accelerate the simulation process, effectively shortening the design and evaluation cycle, thus taking into account both the accuracy and efficiency of GIS electromagnetic simulation.

[0058] (4) This invention takes the preliminarily calculated local discharge source as the center and proposes a heuristic fast iterative search algorithm for the optimal matching of the measurement value in its neighborhood. The fast iterative search problem of the optimal matching position of the discharge source is input into a graph neural network to obtain its heuristic metric, which is used as a substitute for the expert design heuristic. Based on this, the ant colony algorithm is iterated. Under the framework of the ant colony algorithm, the local search of the constructed solution will obtain a better solution. The deep ant colony algorithm uses a graph neural network to generate heuristic metrics, reducing the need for expert knowledge. Combined with probabilistic local search, the local search method with neural-guided perturbation will achieve better performance to ensure efficient solution and realize the fast iterative search of the optimal matching position of the discharge source.

[0059] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0060] Figure 1 is a structural schematic diagram of a GIS field-electricity fusion real-time state perception and early warning system proposed in Embodiment 1 of the present invention; Figure 2 is a principle block diagram of the embedded device for signal synchronous acquisition and measurement in Embodiment 1 of the present invention; Figure 3 is a schematic diagram of the calculation process of the full-domain waveform diagram of the partial discharge signal in Embodiment 1 of the present invention; Figure 4 is a complete feature map of the graph neural network in Embodiment 1 of the present invention; Figure 5 is a schematic diagram of the iterative process of the graph neural network in Embodiment 1 of the present invention; Figure 6 is a principle block diagram of the GIS electromagnetic field distribution simulation enhancement implemented by the finite-difference time-domain method in Embodiment 2 of the present invention; Figure 7 is a principle block diagram of the GIS electromagnetic field distribution simulation enhancement implemented by the finite-element time-domain method in Embodiment 3 of the present invention; Figure 8 is a principle block diagram of the GIS electromagnetic field coarse grid data enhancement based on the differential network in Embodiment 4 of the present invention; Figure 9 is a structural schematic diagram of the enhancement network in the differential deep learning network model in Embodiment 4 of the present invention; Figure 10 is a schematic diagram of the design principle of the differential convolution module in Embodiment 4 of the present invention; Figure 11 is a structural schematic diagram of the structural similarity network in the differential deep learning network model in Embodiment 4 of the present invention; Figure 12 is a principle block diagram of the GIS electromagnetic field simulation data enhancement based on PU-GAN in Embodiment 5 of the present invention; Figure 13 is a schematic diagram of the generator in Embodiment 5 of the present invention; Figure 14 is a schematic diagram of the feature extraction module in Embodiment 5 of the present invention; Figure 15 is a schematic diagram of the structural similarity module in Embodiment 5 of the present invention; Figure 16 is a schematic diagram of the field strong consistency module in Embodiment 5 of the present invention; Figure 17 is a schematic diagram of the self-attention unit in Embodiment 5 of the present invention; Figure 18 is a block diagram of the principle of GIS partial discharge optimal location matching based on deep ant colony algorithm in Embodiment 6 of the present invention; Figure 19 is a schematic diagram of the GIS partial discharge optimal location search process based on deep ant colony algorithm in Embodiment 6 of the present invention; Figure 20 is a schematic diagram of the GIS partial discharge feature extraction process based on deep reinforcement learning in Embodiment 7 of the present invention; Figure 21 is a flowchart of a GIS field-electricity fusion real-time state perception and early warning method proposed in Embodiment 8 of the present invention. Detailed Implementation

[0061] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, 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.

[0062] As shown in Figure 1, Embodiment 1 of this invention proposes a GIS field-electric fusion real-time status perception and early warning system. The system includes a signal synchronous acquisition and measurement embedded device 100 and an industrial control computer 200. The industrial control computer 200 deploys a simulator 201, a location searcher 202, and an early warning device 203. The output of the signal synchronous acquisition and measurement embedded device 100 is connected to the simulator 201 and the location searcher 202, respectively. The output of the simulator 201 is connected to the location searcher 202, and the output of the location searcher 202 is connected to the early warning device 203. The signal synchronous acquisition and measurement embedded device 100 is used to detect partial discharge signals within the electromagnetic field signals of the GIS based on pre-selected strongly correlated features, obtaining a full-time-domain waveform diagram containing both time-domain and frequency-domain waveform diagrams of the partial discharge signals. The simulator 201 is used to coarsely locate the partial discharge source based on the full-time-domain waveform diagram of the partial discharge in the GIS, and uses the coarse location result as the initial injection point for simulation enhancement calculations to obtain simulation enhancement data of the partial discharge source. The location searcher 202 is used to iteratively search the measured discharge intensity calculated based on the full-time waveform diagram and the simulation enhancement results to obtain the actual discharge location of the partial discharge source; the early warning device 203 is used to issue an early warning based on the full-time waveform diagram and early warning information containing the actual discharge location of the partial discharge source.

[0063] This embodiment detects the digital signal converted from electromagnetic field signal based on pre-selected strong correlation features. When a partial discharge signal is detected, it acquires a high-precision full-time-domain waveform of the GIS partial discharge. Then, based on the full-time-domain waveform of the GIS partial discharge, it performs coarse localization of the partial discharge source. The coarse localization result is injected into the simulator to simulate the partial discharge phenomenon of the GIS equipment and perform simulation enhancement calculations to obtain simulation enhancement data of the partial discharge source, achieving fast and high-precision GIS electromagnetic field simulation. Finally, iterative search is performed based on the real-time acquired full-time-domain waveform and simulation enhancement data to obtain the actual discharge location of the partial discharge source. Early warning is achieved based on information such as the actual discharge location. Therefore, this invention achieves high-performance real-time status perception and early warning based on GIS field-electricity fusion.

[0064] As a further preferred technical solution, the embedded device for synchronous signal acquisition and measurement includes a signal acquisition location determination module, a signal acquisition module, a signal processing front-end, an analog-to-digital converter, and a signal measurement module, wherein: the signal acquisition location determination module is used to determine the installation location of each sensor in the signal acquisition module; the signal acquisition module is used to acquire electromagnetic field signals inside the GIS; the signal processing front-end is used to perform signal conditioning on the electromagnetic field signals transmitted by the signal acquisition module, and the conditioned signal is transmitted to the analog-to-digital converter; the analog-to-digital converter is used to convert the conditioned signal from an analog signal to a digital signal and then transmit it to the signal measurement module; the signal measurement module is used to monitor the digital signal based on pre-selected strong correlation features, and when a partial discharge signal is detected, obtain a full-time-domain waveform diagram including a time-domain waveform diagram and a frequency-domain waveform diagram of the partial discharge signal, wherein the strong correlation features include time-domain strong correlation features and frequency-domain strong correlation features.

[0065] This embodiment achieves multi-channel acquisition of electromagnetic field signals inside the GIS by setting up a multi-channel acquisition module, and conditions the acquired electromagnetic field signals to obtain high-performance electromagnetic field signals. Then, based on the pre-selected strong correlation features as indicators, the digital signals converted from the electromagnetic field signals are detected. When a partial discharge signal is detected, time-domain waveforms and frequency-domain waveforms are obtained according to the strong correlation features in the time domain and frequency domain, respectively. The time-domain waveforms and frequency-domain waveforms are combined to obtain a full-time-domain waveform. Therefore, this invention can achieve high-precision acquisition of full-time-domain waveforms of partial discharge in GIS.

[0066] As a further preferred technical solution, as shown in Figure 2, the signal processing front end includes an amplification and filtering circuit with the same number of sensors as the signal acquisition module, and the amplification and filtering circuit is connected to each sensor in a one-to-one correspondence; the amplification and filtering circuit includes a radio frequency gain circuit, a bandpass filter and a bandstop filter connected in sequence, wherein the frequency of the bandpass filter is 300MHz to 1500MHz and the frequency of the bandstop filter is 800MHz to 1000MHz.

[0067] It should be noted that this embodiment has implemented bandpass and bandstop filtering according to the effective frequency band and common interference frequency band of UHF signals in GIS, which is more suitable for UHF signal acquisition in GIS. The bandpass filter is used to extract useful signals in specific frequency bands and filter out irrelevant interference signals and noise. A low-noise, wide-bandwidth, cascadeable and high-linearity signal amplifier can be connected after the bandstop filter.

[0068] As a preferred technical solution, the analog-to-digital converter uses an ADC chip with a data conversion rate of 4Gsps for 4 channels or 5Gsps for 1 channel.

[0069] It should be noted that the synchronous acquisition and measurement embedded device set in this embodiment can realize high-speed data acquisition and real-time processing of 4 channels at 4GSPS, and supports multi-channel sampling. It can realize simultaneous acquisition of 4 channels of data, with a signal input range of 1.4V and 12-bit data at a 4GSPS sampling rate. At the same time, the signal processing logic with fixed function is placed in the hardware FPGA to improve the parallel speed of the system and realize multi-channel, high-precision acquisition of electromagnetic field signals inside GIS.

[0070] As a further preferred technical solution, the signal measurement module includes a buffer unit, a first monitoring unit, a second monitoring unit, and a waveform synthesis unit, wherein: the buffer unit is used to buffer the digital signal; the first monitoring unit is used to extract the strong time-domain correlation features from the digital signal, and determine that a partial discharge signal is detected based on the strong time-domain correlation features, and output a first time-domain waveform diagram during discharge; the second monitoring unit is used to extract the strong frequency-domain correlation features from the digital signal, and determine that a partial discharge signal is detected based on the strong frequency-domain correlation features, and output a frequency-domain waveform diagram during discharge; the waveform synthesis unit is used to obtain a full-time-domain waveform diagram of the partial discharge signal based on the first time-domain waveform diagram and the frequency-domain waveform diagram.

[0071] As a further preferred technical solution, the first monitoring unit is used to analyze the strong correlation features in the time domain and determine whether the first discharge threshold is exceeded. If so, it is determined that a partial discharge signal has been detected; otherwise, it is determined that no partial discharge signal has been detected. The second monitoring unit is used to analyze the strong correlation features in the frequency domain and determine whether the second discharge threshold is exceeded. If so, it is determined that a partial discharge signal has been detected; otherwise, it is determined that no partial discharge signal has been detected.

[0072] As a further preferred technical solution, the digital signal stored in the cache unit is deleted when no partial discharge signal is detected.

[0073] It should be noted that the collected time-domain data is completely sent to the high-performance storage buffer. If it is determined that no partial discharge UHF signal is detected, the detected time-domain data is deleted from the buffer to free up storage space for the next time-domain data to be collected.

[0074] As a further preferred technical solution, the second monitoring unit includes: a Fourier transform subunit, used to perform at least two fast Fourier transforms on the electromagnetic field signal to obtain frequency domain data of at least two frequency bands; and a normalization subunit, used to extract the strong correlation features of the frequency domain data of each frequency band, and after analyzing and matching the strong correlation features, determine that when the frequency domain data of each frequency band exceeds the second discharge threshold, normalize the frequency domain waveforms corresponding to each frequency band to obtain the frequency domain waveform diagram during discharge.

[0075] When the amount of acquired signal data is large, direct FFT calculation would be slow due to the large data volume, which is not conducive to real-time signal detection and processing. Therefore, this application adopts array processing, as shown in Figure 3, dividing the FFT results into four frequency bands for parallel processing (frequency bands are 200M-525M, 525M-850M, 850M-1175M, and 1175M-1500M respectively). The correlation of the four frequency bands is calculated separately, which can increase the number of correlation judgment criteria, thereby greatly increasing the amount of information reflecting the relative deformation of the waveform. At the same time, it can also fully reflect the changes of the frequency response curve in different frequency ranges due to different degrees of discharge. Finally, the waveforms of multiple frequency bands are fused.

[0076] Therefore, the array processing mechanism of this application is a frequency domain segmentation and combination mechanism. Each part of the array only processes a part of the frequency domain, and then they are merged together to restore the complete frequency domain waveform. This mechanism disperses the noise signal superimposed on the frequency domain data to the segmented part, reducing the stacking of noise signals, which will help to achieve high-sensitivity detection of partial discharge UHF signals.

[0077] As a further preferred technical solution, the waveform synthesis unit includes: a waveform conversion subunit, used to convert the frequency domain waveform diagram into a second time domain waveform diagram during discharge; and a waveform synthesis subunit, used to synthesize the first time domain waveform diagram and the second time domain waveform diagram to obtain a full time domain waveform diagram of the partial discharge signal.

[0078] Furthermore, this embodiment selects the Xilinx UltraScale+MPSoC FPGA chip to process digital signals. It should be noted that this embodiment receives sampled data from the analog-to-digital converter (ADC) based on the buffer unit, calculates feature indicators for the transmitted time-domain waveform data, matches the data, performs FFT to convert the time-domain waveform into a frequency-domain waveform, arrays the frequency-domain waveform, calculates feature indicators, matches the data, normalizes the obtained frequency-domain waveform, performs IFFT after normalization, obtains the time-domain waveform, and combines it with the waveform directly obtained in the time domain to output a full-time-domain waveform.

[0079] This embodiment proposes a frequency domain arraying method and a real-time digital matching method, and develops an embedded device for synchronous acquisition and measurement of GIS partial discharge UHF signals based on an ultra-high-speed ADC and FPGA architecture, to obtain high-precision full-time domain waveform diagrams of partial discharge UHF signals.

[0080] It should be noted that since there are various time-frequency domain indicators for detecting partial discharge signals, if all time-frequency domain indicators are calculated to detect partial discharge signals, the detection results will theoretically be very accurate. However, if detection and judgment are based on all time-frequency domain indicators, it will consume a large amount of computing resources, which current hardware computing resources cannot perform. On the other hand, randomly selecting certain features for calculation cannot meet the accuracy requirements.

[0081] This embodiment uses a feature selection module to select strongly correlated features through a neural network for subsequent partial discharge signal detection, which improves the accuracy of partial discharge signal detection while greatly reducing the amount of computation.

[0082] Specifically, the time-frequency domain features include margin, skewness, root mean square (RMS), mean, variance, peak-to-peak value, shape factor, kurtosis, impulse factor, and C-index; the frequency domain features include mean frequency, frequency concentration, frequency center, RMS frequency, and variation in the position of the main frequency band.

[0083] Specifically, the time-frequency domain characteristics of the detected partial discharge signal are shown in Table 1: Table 1: Selected Time-Frequency Domain Characteristics

[0084] In the formula: L is the margin index, x p For the maximum absolute value (peak value), x r s is the square root amplitude. k σ is the skewness, μ3 is the third center moment of the time-domain waveform data x, σ is the standard deviation of the time-domain waveform data x, and x rms x is the root mean square. n Let be the nth time-domain data point, N be the total number of time-domain data points, SF be the shape factor, and x be the value of x. i For deviation: kurtosis is the kurtosis, m is the impulse factor, x is the time-domain waveform data, p1 is the mean frequency, y is the amplitude of each frequency in the spectrum, k is a fixed coefficient used to adjust the ratio in the formula, p2 is the frequency concentration index 1, p3 is the frequency concentration index 2, p4 is the frequency concentration index 3, p5 is the frequency center index, f is the frequency of each point in the spectrum, p6 is the frequency concentration index 4, p7 is the root mean square frequency index, p8 is the main frequency band position change index 1, p9 is the main frequency band position change index 2, p 10 For the frequency concentration index 5, p 11 For the frequency concentration index 6, p12 For the frequency concentration index 7, p 14 8 represents the frequency concentration index, and * represents the multiplication sign.

[0085] It should be noted that the indicators listed in Table 1 are common time-frequency domain indicators, some of which contain certain characteristics of UHF and are common UHF signal indicators. These indicators contain some information that can be used for UHF signal detection, but this information may be repeated or complementary in multiple features. Therefore, this embodiment uses a graph neural network to filter out indicators with strong correlations among these indicators.

[0086] As a further preferred technical solution, the feature selection module includes: a feature map construction unit, used to construct a feature map from a complete feature map G = (V, E) of a graph neural network, where V is the set of nodes and E is the set of edges, and let x... v x co[v] h ne[v] x ne[v] Let h represent the feature vector of node v, the feature vector of the edge, the state vector of node v and its surrounding nodes, and the feature vector of the surrounding nodes of node v, respectively. The aggregation update unit is used to aggregate and update the information of the input node and edge according to the local transition function for updating the node state, and outputs the node label, expressed by the formula: h v =f(x) v ,x co[v] h ne[v] ,x ne[v] )

[0087] o v =g(h v x v )

[0088] In the formula, f(·) is the local transition function for updating the node state, g(·) is the local output function, and h v o is the state vector learned through iterative learning by the graph neural network. v Node labels; Iteration units are used to calculate the association between a node and its neighboring nodes in the continuous iteration of the graph neural network, so as to obtain strong correlation features in the time domain and strong correlation features in the frequency domain.

[0089] As a further preferred technical solution, the iterative unit is specifically used for: Let H, O, X and X N These are vectors constructed by superimposing the state vectors of all nodes, the output labels of all nodes, the feature vectors of all nodes and edges, and the feature vectors of all nodes, respectively. The formula can be written in a more compact form as: O = G(H, X) N )

[0090] H t+1=F(H t (X)

[0091] Among them, H t H represents the t-th iteration of H. t+1 =F(H t Let X represent the state vector of all nodes in the (t+1)th iteration obtained by passing the feature vector and the state vector of the tth iteration through the global transfer function. F(·) and G(·) are the global transfer function and the global output function, respectively, which are the stack of the local transfer functions f(·) and the local output functions g(·) of all nodes. During the iteration process, nodes with similar states, nodes with complementary states, and the influence of each node on the overall graph neural network are determined. One node with similar states is selected, nodes with complementary states are merged into one node, and the node with the greatest influence on the overall graph neural network is selected as the strongly correlated feature.

[0092] It should be noted that the feature vectors of nodes represent certain inherent states of the nodes, which do not change during the iteration process of the graph neural network. Specifically, in this embodiment, they represent certain parameters in the time-frequency domain feature calculation formula of the electromagnetic field signal, such as standard deviation and peak value. The feature vectors of edges refer to the relationships between two nodes. Specifically, in this embodiment, they refer to nodes sharing the same parameters or similar processing procedures in the calculation formula. Connecting these nodes with similar features helps the model learn the relationships and mutual influences between these features. The state vector refers to the node vector that is continuously updated based on the calculations during the iteration process; the node label is the target value used for supervised learning tasks. In this embodiment, each node represents a signal feature, with corresponding labels to indicate the category and state of these signal features.

[0093] This embodiment utilizes a graph neural network (GNN) to calculate the relationship between a node and its neighboring nodes. Through multiple iterations, it can also find the relationship between the neighboring nodes of neighboring nodes and the current node, making it very suitable for filtering strongly correlated features. In this embodiment, all features of the GIS partial discharge time-frequency domain signal (as shown in Table 1 above) are imported into the graph neural network as nodes in the network. The complete feature map constructed is shown in Figure 4. In Figure 4, each node represents a single feature, and the lines between nodes represent the relationship between the features represented by the nodes. Some features have a strong correlation, such as p2 and p3 both containing p1, while the relationship between some features is relatively weak, so no connection is made. Blank nodes represent features that are not closely related to other features.

[0094] As shown in Figure 5, during the iteration process of a graph neural network, the feature vector inherent to each node does not change with the network iteration; however, the state vector between nodes changes with the iteration of the GNN network. Therefore, a graph neural network can be used to calculate the connection between a node and its neighboring nodes. Through multiple iterations, the connection between the neighboring nodes of a neighboring node and the current node can also be found. Through the continuous iteration of the graph neural network, it can be found that the states of some nodes in the graph neural network become more and more similar, the states of some nodes are complementary, and the influence of some nodes on the overall graph neural network decreases. By selecting one of the similar nodes, merging the complementary nodes, and then selecting the node with the greatest impact on the overall system, the strong correlation characteristics of the system can be obtained.

[0095] In this embodiment, the graph neural network is pre-constrained by a loss function, and then the gradient descent algorithm is used to minimize the loss, thereby obtaining the optimal parameters of the network model and a trained graph neural network. This embodiment substitutes the features from Table 1 above into the graph neural network and performs correlation strength filtering to obtain the features with the highest correlation, i.e., the strongly correlated features. The results are the root mean square (RMS), impulse factor, and margin index in the time domain, and the RMS frequency index, main frequency band position change index 1, and main frequency band position change index 2 in the frequency domain. Then, partial discharge signals are detected based on these filtered strongly correlated features. Although the computational load is reduced, the accuracy of partial discharge signal detection is not affected because the correlation between the indices is considered. Therefore, this invention improves the accuracy of partial discharge signal detection while significantly reducing the computational load.

[0096] Example 2, based on the content disclosed in Example 1, provides a detailed description of the process by which the simulator in Example 1 enhances the simulation of GIS electromagnetic field distribution through finite-difference time-domain calculus. The simulator includes: a simulation module, used to use the coarse location result of the partial discharge source obtained based on the full-time-domain waveform as the initial injection point of the 3D simulation model of the GIS equipment to simulate the partial discharge phenomenon of the GIS equipment and perform simulation calculations to obtain coarse-grid simulation results; and an enhancement module, used to enhance the coarse-grid simulation results using an enhancement model to obtain enhanced simulation data.

[0097] As a further preferred technical solution, as shown in Figure 6, the simulation module includes: a spatial discretization unit, used to discretize the three-dimensional simulation model of the GIS equipment into a spatial grid, and determine the correspondence between the number of each grid node and the spatial coordinate data; and a differential simulation unit, used to inject the coarse positioning result as the initial injection point into the corresponding spatial grid to simulate the partial discharge phenomenon of the GIS equipment, convert the Maxwell equations used to describe the electromagnetic field of GIS into a fourth-order matrix, and use the four-step HIE-FDTD algorithm to solve the fourth-order matrix to calculate the field strength data of each grid node.

[0098] As a further preferred technical solution, the differential simulation unit includes: a matrix reconstruction subunit, used to convert Maxwell's equations used to describe the electromagnetic field of GIS into a sixth-order matrix form:

[0099] In the formula, Let be a vector composed of the components of the electric and magnetic fields in a rectangular coordinate system. in:

[0100]

[0101]

[0102]

[0103]

[0104]

[0105] Where [M] is a sixth-order matrix, represented as:

[0106] In the formula, ε is the dielectric constant and μ is the permeability.

[0107] It should be noted that this embodiment represents Maxwell's equations in matrix form, so that the electromagnetic field components are expressed in vector form, and the differential operations in the equations are described by matrix operations. This form provides a more unified and compact description method, making the equations easier to understand and derive. Furthermore, in the numerical solution process, numerical methods usually involve solving discretized equations, and the matrix form makes these solution methods more direct and effective.

[0108] The matrix transformation subunit is used to convert a sixth-order matrix into a fourth-order rectangular form as follows:

[0109] In the formula, [A H ] and [B H The representations are as follows:

[0110]

[0111] The sub-element is used to solve the fourth-order matrix form using the four-step HIE-FDTD algorithm, and to calculate the field strength data for each grid node.

[0112] This embodiment significantly improves the accuracy and stability of numerical solutions by introducing higher-order time and spatial integration schemes. Traditional FDTD methods may be affected by numerical dissipation and numerical dispersion, while this invention reduces these problems through more accurate integration methods, especially effective in handling long-term simulations or high-frequency electromagnetic wave propagation. For electromagnetic field simulations of complex structures, such as those involving media, non-uniform media, or non-uniform structures, this embodiment employs a four-step hybrid implicit finite-difference time-domain (HIE-FDTD) method to better handle these complexities. Its accurate integration and numerical methods can effectively simulate electromagnetic field propagation and interaction between different media.

[0113] As a further preferred technical solution, the solution sub-unit is used to perform the following steps: The four-step HIE-FDTD algorithm is decomposed in the time domain into four sub-steps: n→n+1 / 4, n+1 / 4→n+2 / 4, n+2 / 4→n+3 / 4, and n+3 / 4→n+1. For each sub-step, a semi-implicit difference scheme is used to calculate the tridiagonal implicit expression for the field strength data of the grid nodes; The chasing method is used to solve the tridiagonal implicit expression to obtain the field strength data of the grid nodes.

[0114] Specifically, the four-step HIE-FDTD algorithm is decomposed in the time domain into four sub-steps: n→n+1 / 4, n+1 / 4→n+2 / 4, n+2 / 4→n+3 / 4, and n+3 / 4→n+1, where: (1) n→n+1 / 4 sub-step:

[0115] (2) Step-by-step process from n+1 / 4 to n+2 / 4:

[0116] (3) Step-by-step process from n+2 / 4 to n+3 / 4:

[0117] (4) n+3 / 4 → n+1 in steps:

[0118] For the step n→n+1 / 4, an auxiliary vector is introduced using a semi-implicit difference scheme. To keep the stability condition of the algorithm unchanged, we can obtain:

[0119] in and The equations are mutually coupled, and the substitution method is used to eliminate them. The tridiagonal implicit equation is as follows:

[0120] In the formula, σ is the point loss parameter of the medium. Δt is the time step, E x E y E z H represents the electric field components in the x, y, z directions. x H y H z The magnetic field components in the x, y, z directions, with the superscript indicating the time step, for example... Ex represents time step n. Let Ex represent the electric field at the 1 / 4 position between time steps n and n+1, where [I] is the sixth-order identity matrix.

[0121] For n+1 / 4 → n+2 / 4, the auxiliary vector Substituting the expression and expanding it, we get:

[0122] For n+2 / 4 → n+3 / 4, the auxiliary vector Substituting the expression and expanding it, we get:

[0123] For n+3 / 4 → n+1, the auxiliary vector Substituting the expression and expanding it, we get:

[0124] The components of the electric and magnetic fields in the x, y, and z directions at each step can be obtained from the above equations.

[0125] This embodiment significantly improves the accuracy and stability of the numerical solution by introducing higher-order time and space integration schemes. Traditional FDTD methods may be affected by numerical dissipation and numerical dispersion, but this embodiment reduces these problems through more accurate integration methods, particularly effective in handling long-term simulations or high-frequency electromagnetic wave propagation. For electromagnetic field simulations of complex structures, such as those involving media, non-uniform media, or non-uniform structures, this embodiment better handles these complexities. Its accurate integration and numerical methods can effectively simulate electromagnetic field propagation and interaction between different media.

[0126] It should be noted that this embodiment can obtain simulation results through calculation. The results show the spatial coordinates of each grid node and the distribution data of the electromagnetic field at that point. Since the coarseness of the simulation grid is directly proportional to the calculation accuracy and inversely proportional to the calculation time, in order to resolve the contradiction between accuracy and time, a real-time grid data enhancement simulation method is designed. Based on graph machine learning, the electromagnetic field distribution results of the coarse grid simulation are enhanced to achieve the same result as the finer grid simulation. Further, in step S40: a node interpolation enhancement strategy is used to generate new grid nodes and feature values ​​of the new grid nodes between adjacent grid nodes. Specifically, this is used to create synthetic nodes by combining existing nodes through Node Interpolation. This is a generative enhancement strategy that generates new samples by interpolating between minority class samples and their nearest neighbor samples. The training distribution is as follows:

[0127]

[0128]

[0129] In the formula, (x i y i , z j ) and (x j y j , z j ) are two adjacent grid nodes, both originating from V p , For the new grid node, λ∈[0,1], that is, through the point (x i ,y i ,z j ) and point (x) j ,y j ,z j Generate a new node between ) The new node position is determined by the value of λ.

[0130] As a further preferred technical solution, the field strength data of the new grid node is: E=(1-Δ)·E i +Δ·E j

[0131] H=(1-Δ)·H i +Δ·H j

[0132] In the formula, (E i H i ) and (E j H j(E, H) represents the field strength data of two adjacent grid nodes, (E, H) represents the field strength data of the new grid node, and Δ is an undetermined coefficient. Δ is determined by λ and the field strength attribute relationship between the two nodes. The specific value needs to be confirmed in conjunction with the corresponding field strength because the gradient of distance change is different from that of field strength change.

[0133] This allows new nodes to be generated between adjacent nodes. The new node feature values ​​also include spatial and field strength data, thus enhancing the electromagnetic field distribution results of the coarse-grid simulation to achieve the same results as fine-grid simulations. It should be noted that this embodiment uses an algorithm based on node proliferation between adjacent nodes, involving the spatial and electromagnetic field data of existing nodes. For grid data in GIS simulations, which have obvious topological structure characteristics, this embodiment can effectively capture the relationships between grids and preserve the topological structure information of the grid data. Furthermore, the algorithm in this embodiment generates new node representations based on the characteristics of adjacent nodes, allowing the algorithm to consider the local features of nodes. This embodiment is a graph-based algorithm for enhancing coarse-grid data obtained from simulations, adaptable to grid data of different scales, including small and large-scale grids, and compatible with GIS simulation data from different devices and voltage levels, achieving GIS simulation data enhancement and improving the quality of grid data.

[0134] Based on the content disclosed in Example 1, as shown in Figure 7, Example 3 provides a detailed description of the process by which the simulator in Example 1 enhances the simulation of GIS electromagnetic field distribution through finite-difference time-domain calculus. The simulator includes: a simulation module, used to use the coarse location result of the partial discharge source obtained based on the full-time-domain waveform as the initial injection point of the 3D simulation model of the GIS equipment to simulate the partial discharge phenomenon of the GIS equipment and perform simulation calculations to obtain coarse-grid simulation results; and an enhancement module, used to enhance the coarse-grid simulation results using an enhancement model to obtain enhanced simulation data.

[0135] Specifically, the simulation module is used to take the coarse location result of the partial discharge power source obtained based on the full-time waveform diagram as the initial injection point of the 3D simulation model of the GIS equipment to simulate the partial discharge phenomenon of the GIS equipment. The simulation calculation is performed using the time-domain finite element method to obtain the coarse mesh simulation result. Specifically, according to the structural parameters of the GIS equipment, a 3D data simulation model can be established using simulation software such as COMSOL Multiphysics. The 3D data model is discretized into multiple tetrahedral elements by the finite element method. Each tetrahedral element includes several nodes. The simulation calculation is performed to obtain the coarse mesh simulation data that needs to be processed.

[0136] Specifically, the enhancement module is used to enhance coarse-grid simulation data based on the GraphSAGE algorithm, including: a network abstraction unit, used to abstract the simulation model of the GIS device into a network topology structure G = (V p E s And consider any two adjacent free tetrahedrons in the network topology as a slice unit, where V p Denotes the set of nodes, E s The first aggregation unit represents the set of edges, with the lines connecting the grid vertices as edges; the second aggregation unit is used to aggregate the neighboring nodes of each node in each patch unit using the simulation feature data of the node as the initial baseline feature, and the GraphSAGE algorithm is used to generate the initial baseline feature of the new node; the third aggregation unit is used to aggregate the neighboring nodes of the new node using the GraphSAGE algorithm to generate the feature vector representation of the new node.

[0137] It should be noted that for any newly added node generated by the overlapping surfaces of two adjacent free tetrahedrons, each node in the sheet element formed by these two free tetrahedrons can be considered as the neighbor node of the newly added node. The GraphSAGE algorithm is used to perform secondary aggregation processing on the neighbor nodes of the newly added node to generate the feature vector representation of the newly added node, that is, to obtain the three-dimensional spatial coordinate data and electric field intensity of the newly added node. This allows more new node data to be generated on the basis of the original mesh node data, thus completing the enhancement and refinement of the mesh node data generated by the simulation calculation.

[0138] In detail, the GraphSAGE algorithm typically uses random sampling of neighboring nodes to update the representation of the target node. However, in GIS simulation, the selection of neighboring nodes needs to consider more factors, such as the physical distance between nodes and their similarity. Therefore, this embodiment treats two free tetrahedral elements as a single slice element, allowing the neighbors of newly added nodes to cluster within that slice element. This increases the similarity and correlation between nodes. Furthermore, considering the various types of node features in GIS simulation data, using the spatial coordinates and electric field intensity of nodes as node features better aligns with the data relationships in GIS simulation. Additionally, this embodiment adjusts the model depth of the GraphSAGE algorithm, i.e., the number of aggregation layers, to suit the complexity and scale of GIS simulation data, improving the algorithm's ability to represent data and achieving better performance.

[0139] As a further preferred technical solution, the primary aggregation unit is used to perform the following steps: (1) using the simulation feature data of each node as the initial reference feature of each node; (2) traversing each node i in each of the aforementioned slice units, and using all adjacent nodes of each node as sampling points to obtain the neighbor set of node i. Then, the initial baseline features of all nodes in the neighbor set corresponding to node i are aggregated using an aggregation function to obtain the first aggregated neighbor features. Specifically, each node in the network topology The feature vectors, i.e., the simulation feature data, serve as the baseline features for each node. For each slice unit, traverse each node i in the slice unit during the first aggregation. For node i, first obtain its neighbor set. The feature vector of node i is used as its initial baseline feature.

[0140] Specifically, the initial baseline features of all nodes in the neighbor set corresponding to node i are aggregated by an aggregator using an aggregation function, thus obtaining the first aggregated neighbor features. for:

[0141] In the formula, AGGREGATE1() represents an aggregate function.

[0142] (3) Connect the initial baseline feature and the aggregated neighbor feature. After obtaining the first joint feature, the first weight matrix is ​​multiplied with the first joint feature, and then a nonlinear transformation is performed to obtain the node features after one traversal. Specifically, in this embodiment, the features of each node after one traversal are... This serves as the new feature vector representation for this node, and as the baseline feature for each node during the second traversal, where:

[0143] In the formula, CONCAT is the concatenation function, and W 1 σ is the first weight matrix, and σ is the sigmoid function.

[0144] Furthermore, in this embodiment, the aggregation function used is the average aggregation function MEANAGGREGATE, i.e.

[0145] (4) Use an aggregation function to aggregate the initial datum features of the common nodes of the two free tetrahedrons in each slice element to obtain the initial datum features of the newly added nodes.

[0146] Specifically, since the newly added nodes generated from the common face of two adjacent free tetrahedrons do not have initial datum features, this embodiment uses an aggregation function to aggregate the initial datum features of the common nodes of two adjacent free tetrahedrons to obtain the initial datum features of the newly added node P. The formula is expressed as:

[0147] In the formula, For the initial reference features of the common nodes of two adjacent free tetrahedrons, and further, in this embodiment, the aggregation function adopts the average aggregation function, then we have:

[0148] As a further preferred technical solution, the secondary aggregation unit is specifically used to perform the following steps: (1) converting the node features after one traversal (2) Using each node i in each of the aforementioned slice units as a neighbor sampling point of the newly added node, and using an aggregation function to aggregate the new benchmark features of all nodes in each of the aforementioned slice units, to obtain the secondary aggregated neighbor features. Specifically, in this embodiment, all nodes i in the slice unit are used as the neighbor set P(i) of the newly generated nodes in that slice unit. A second aggregator is then used to aggregate the new baseline features of all nodes in the neighbor set using an aggregation function. Obtain secondary aggregated neighbor features The formula is expressed as:

[0149] In the formula, AGGREGATE2() is an aggregate function. This is a new reference feature for each node in the slice unit.

[0150] (3) Connect the initial baseline features of the newly added node with the secondary aggregated neighbor features to obtain the second joint features. Perform matrix multiplication on the second weight matrix and the second joint features and then perform a nonlinear transformation to obtain the feature vector representation of the newly added node.

[0151] Specifically, the feature vector of the newly added node obtained in this embodiment is represented as follows:

[0152] In the formula, CONCAT is the concatenation function, and W 2 σ is the second weight matrix, and σ is the sigmoid function.

[0153] Furthermore, since the neighbors of a node do not have a natural order, the aggregator AGGREGATE in this embodiment... k Using the average aggregation function, we have Thus, a new node feature vector is generated between any two slice elements, namely the three-dimensional spatial coordinates and electric field intensity of the new node, thereby enhancing the simulation mesh node data of the GIS finite element method.

[0154] As a further preferred technical solution, the complexity of the GraphSAGE algorithm used in this embodiment is:

[0155] Here, K represents the number of aggregators, the number of weight matrices, and the number of layers in the network. This embodiment uses K aggregators. Used to aggregate node neighbor information, and K weight matrices. It is used to propagate information between different layers. To improve the sampling aggregation effect, K=2 is set, and aggregation is performed twice.

[0156] As a further preferred technical solution, the sampling of nodes in this embodiment is of fixed length, S n The number of sampled neighbors in the nth layer is represented by the predefined number of sampled neighbors S. The complexity is stabilized by using either repeated sampling with replacement or negative sampling. When the defined number of sampled neighbors is greater than the actual number of neighbors, repeated sampling is used to make the actual number of sampled neighbors reach S. When the defined number of sampled neighbors is less than the actual number of neighbors, negative sampling is used to reduce the actual number of sampled neighbors to S.

[0157] Example 4: Based on the content disclosed in Example 1, this example specifically uses a differential deep learning network model to enhance the coarse mesh simulation results, obtaining enhanced simulation data. The specific working principle of the differential deep learning network model is shown in Figure 8: The differential deep learning network model includes an enhancement network and a structural similarity network connected in sequence. The enhancement network includes a self-adjusting module and a differential convolution module. The coarse mesh simulation results include coarse mesh structure data and coarse mesh field strength data. The self-adjusting module and the differential convolution module are used to calculate the coarse mesh structure data and the coarse mesh field strength data, respectively, to obtain fine mesh enhanced mesh structure features and fine mesh enhanced field strength features. The structural similarity network is used to match the fine mesh enhanced mesh structure features and the fine mesh enhanced field strength features to calculate similarity features. Based on the similarity features, the enhanced simulation data is calculated.

[0158] This embodiment designs and constructs a differential deep learning network model, which uses convolution combined with differential algorithms to process coarse grid field strength data, enabling accurate simulation of electromagnetic field distribution in GIS, especially when dealing with complex geometric structures and boundary conditions. Furthermore, it utilizes deep learning networks to optimize the simulation process, improve simulation accuracy, and accelerate the simulation process, effectively shortening the design and evaluation cycle, thus balancing the accuracy and efficiency of GIS electromagnetic simulation.

[0159] As a further preferred technical solution, as shown in Figure 9, the self-adjusting module includes a first convolutional layer Conv1 and a second convolutional layer Conv2 connected in sequence. The outputs of the first convolutional layer Conv1 and the second convolutional layer Conv2 are connected through a first addition operation and then output to the activation function layer. The coarse grid structure data is used as the input of the first convolutional layer Conv1, the grid size supplementary information of the coarse grid structure data is used as the input of the second convolutional layer Conv2, the deviation vector of the coarse grid structure data is used as the input of the first addition operation, and the coarse grid structure data and the grid size supplementary information of the coarse grid structure data are output to the first addition operation through a residual connection Residual.

[0160] It should be noted that this embodiment achieves fine enhancement of coarse grid structure data by designing a self-adjusting module. The self-adjusting module combines convolution operations, residual connections, and the sigmoid activation function to achieve fine enhancement of coarse grid data of electromagnetic fields of GIS equipment. This method not only improves the detail representation of grid data, but also, through the self-adjusting mechanism, can adaptively adjust the enhancement strategy according to the original grid data, ensuring the accuracy and adaptability of the enhanced grid structure.

[0161] As a further preferred technical solution, the self-adjusting module functions to refine the mesh structure. In this embodiment, the coarse mesh data U of the k-th GIS equipment electromagnetic field is used. k The self-adjusting module in the input enhancement module is used to obtain the fine mesh enhanced mesh structure features. The formula is expressed as:

[0162] In the formula, To enhance the mesh structure features for finer meshes, Conv1() represents the mesh enhancement convolution operation, and Conv2() represents the fine-tuning convolution operation. k For the coarse grid structure data, h k To supplement the grid size information of the coarse grid structure data, Let be the deviation vector of the coarse grid structure data, Residual() be the residual link, and sigmoid be the activation function.

[0163] As a further preferred technical solution, as shown in Figure 9, the differential convolution module includes sequentially connected differential convolutional layers (Conv). npt Conv Dimension Integration Layer size The system consists of a self-attention mechanism layer (SelfAttention) and a third convolutional layer (Conv3), with a ReLU activation function following the Conv3 layer. The coarse-grid field strength data serves as the basis for the differential convolutional layer (Conv3). nptThe input of the differential convolutional layer Conv npt Used to calculate field strength features of different sizes in the coarse-grid field strength data using a differential algorithm; the size integration layer Conv size The field strength features of different sizes calculated by the differential convolutional layer are summed to obtain the reshaped field strength features. The reshaped field strength features are then passed through the SelfAttention layer and the third convolutional layer Conv3 to output fine-grid enhanced field strength features.

[0164] It should be noted that this embodiment employs a differential convolution module for accurate extraction of electromagnetic field strength data. This module, through specially designed differential convolution operations, can precisely extract electromagnetic field strength features, effectively simulating the physical distribution of the electromagnetic field. Furthermore, combined with a self-attention mechanism, this module can highlight important field strength features, improving the accuracy of field strength data extraction and the model's focus. Moreover, by extracting and summing features using convolution kernels of different sizes, multi-scale features of the electromagnetic field data can be captured. This is particularly important for understanding and simulating complex changes in the electromagnetic field. This method enhances the model's generalization ability, enabling it to better adapt to electromagnetic field variations at different scales.

[0165] As a further preferred technical solution, the convolution kernel of the differential convolutional layer adopts any one of the following: five-point differential convolution kernel, weighted differential convolution kernel, multi-scale differential convolution kernel, directional differential convolution kernel, nine-point differential convolution kernel, and hybrid mode differential convolution.

[0166] It should be noted that those skilled in the art can also select other difference algorithms to combine with deep learning networks according to actual application needs, and this embodiment does not make specific limitations.

[0167] Furthermore, taking a five-point difference as an example, as shown in Figure 10, the differential convolutional layer Conv... npt The design is as follows: Based on the five-point difference algorithm, the basic convolutional kernel K is designed. basic :

[0168] Considering the enhancement of electromagnetic field simulation data for 3D GIS, a dimensional transformation kernel K is used. transform Obtain the three-dimensional differential convolution kernel K 3D :

[0169] K 3D =K basic ☉K transform

[0170] Here, ⊙ represents element-wise multiplication.

[0171] Using 3D differential convolution kernel K 3DExtraction of field strength gradient information from coarse-grid data (extracting the electromagnetic field gradient changes from the coarse-grid to guide the generation of fine-grid data): G = I k *K 3D

[0172] Where K represents the convolution operation, and G is the application K. 3D The gradient information data was obtained subsequently; Upsampling was used to enhance the coarse-grid field strength data to obtain upsampled fine-grid field strength data U (obtaining refined coarse-grid data): U = I k *K upsample

[0173] Among them, K upsample The upsampled convolutional kernels are obtained during network training; preliminary fine-grid field strength data are obtained by combining the coarse-grid field strength gradient information G.

[0174]

[0175] This embodiment directly utilizes the differential algorithm to define the convolution kernel, ensuring that physical principles are accurately mapped to the convolution operation. This is achieved by extracting the core calculation formula of the differential algorithm and transforming it into the form of a convolution kernel, thereby guaranteeing the correct application of continuous physical laws to discrete data. Furthermore, this embodiment considers scale transformation when designing the differential convolution kernel, allowing for the analysis of electromagnetic fields at different resolutions. By adjusting the size and shape of the convolution kernel, different physical resolutions can be simulated, thereby capturing multi-scale features.

[0176] Extending two-dimensional convolution kernels to three-dimensional space through a certain transformation is a key step in applying the traditional difference method to three-dimensional GIS electromagnetic field simulation. This dimensional transformation considers the physical effects in each direction in three-dimensional space. By designing a symmetrical three-dimensional convolution kernel, the rotation invariance problem can be solved. A symmetrical kernel will produce a consistent response in all directions, thus solving the isotropic problem in electromagnetic field simulation.

[0177] It should be noted that the special differential convolution operation in this implementation is based on methods for solving numerical partial differential equations, such as the five-point difference method. This method is often used in numerical analysis to approximate the solutions to partial differential equations. In deep learning, traditional convolutional layers are typically used to extract features from input data, while special differential convolution kernels can be designed to better suit the simulation of specific physical phenomena, such as electromagnetic field simulations. The design of such convolution kernels is usually based on the discrete approximation of the second derivative of the function space in the difference algorithm, enabling the network to capture the physical characteristics of the data while learning it.

[0178] Compared to traditional convolutional layers, differential convolution can maintain high-quality data while ensuring that the data conforms to specific physical laws, making the data generated by the network more physically reliable. Furthermore, this combination makes the model perform better when dealing with problems requiring precise physical modeling, such as electromagnetic field simulations and climate change simulations.

[0179] Furthermore, the differential convolution module in this embodiment is configured to acquire accurate fine-grid field strength data, and to use the differential convolution module to process the coarse-grid field strength data I of the k-th GIS device electromagnetic field. k Processing is performed to obtain fine-mesh enhanced field strength characteristics. The formula is expressed as:

[0180]

[0181]

[0182] In the formula, For the field strength features generated by differential convolution, Conv npt () represents the differential convolution operation, I k For the coarse-grid field strength data, For the field strength characteristics after normalization, Conv size This indicates the application of convolution kernels of different sizes. To enhance the field strength features of the fine mesh, SelfAttention() is a self-attention mechanism operation, Conv() is a convolution operation, and ReLU is an activation function.

[0183] Furthermore, this embodiment optimizes the computational efficiency and accuracy of simulation by combining a self-adjusting module and a differential convolution module. The self-adjusting module reduces unnecessary calculations through intelligent enhancement strategies, while the differential convolution module ensures high-precision extraction of field strength data. Together, they achieve efficient and accurate electromagnetic field simulation. In summary, the design and implementation of the self-adjusting module and the differential convolution module, through intelligent enhancement, accurate feature extraction, and efficient calculation methods, greatly improve the accuracy, efficiency, and applicability of GIS electromagnetic field simulation.

[0184] As a further preferred technical solution, as shown in Figure 11, the structural similarity network includes a first branch network, a second branch network, a second addition operation, and a first multilayer perceptron (MLP). The outputs of the first branch network and the second branch network are both connected to the second addition operation, and the output of the second addition operation is connected to the first MLP. final The multilayer perceptron (MLP) final It is followed by the activation function ReLU.

[0185] Specifically, the first branch network includes a convolutional neural network layer (CNN) and a batch regularization operation (BatchNorm) connected in sequence, with the BatchNorm followed by an activation function (ReLU); the second branch network includes a second multilayer perceptron (MLP), with the second MLP followed by an activation function (ReLU).

[0186] Furthermore, the function of the structural similarity network is to match the enhanced grid data with the field strength data, and to enhance the grid structure features of the coarse grid number of the electromagnetic field of the k-th GIS device. Enhanced field strength characteristics with fine mesh Field strength and structural characteristics:

[0187]

[0188]

[0189] f k,final =ReLU(MLP) final (fk,combined))

[0190] In the formula, Information is extracted from the enhanced grid structure features corresponding to the coarse grid number of the electromagnetic field of the kth GIS device. This provides information on the enhanced field strength features of the finer grid corresponding to the coarse grid number of the electromagnetic field of the k-th GIS device; α and β represent the information on the enhanced grid structure features extracted by the finer grid. Information extracted by enhancing field strength features with fine mesh The modulation weights; fk, combined is a comprehensive feature extracted from the fine-mesh enhancement of mesh structure features and fine-mesh enhancement of field strength features; f k，final Similarity features; BatchNorm() is a batch regularization operation, CNN() is a convolution operation, and MLP is a similarity feature. final () is a multilayer perceptron for structural similarity modules, and fk,combined is the combined field strength and structural features.

[0191] It should be noted that the structural similarity network designed in this embodiment serves the following purposes: (1) Preserving physical structure: By focusing on the structural attributes of the data, the structural similarity module ensures that the enhanced data retains its physical accuracy while maintaining structural integrity and coherence. This is particularly important in the field of simulation, as the physical behavior of electromagnetic fields strongly depends on their spatial structural characteristics. (2) Improving data enhancement quality: This module helps generate higher quality data, especially in restoration and super-resolution tasks, ensuring that the details of the generated data have a high structural similarity to the original coarse grid data. (3) Enhancing feature representation: By fusing features extracted by CNN and MLP, the structural similarity module provides an effective mechanism to enhance and significantly represent key features in electromagnetic field data, especially in terms of field strength and grid structure.

[0192] As a further preferred technical solution, the calculation of GIS fine grid enhanced data based on the similarity features specifically includes the following steps: (1) Calculating the normalized coarse grid structure data, normalized coarse grid field strength data, normalized fine grid enhanced grid structure features, and normalized fine grid enhanced field strength features based on the coarse grid structure data, the coarse grid field strength data, the fine grid enhanced grid structure features, and the fine grid enhanced field strength features, respectively; It should be noted that in this embodiment, a field strength consistency module is specifically used to process the coarse grid structure data, the coarse grid field strength data, the fine grid enhanced grid structure features, the fine grid enhanced field strength features, and the similarity features, the purpose of which is to converge the differences between the fine grid enhanced data and the coarse grid data.

[0193] The electromagnetic field structure characteristics of the k-th GIS device are calculated using the following formula. and fine mesh enhance field strength characteristics With the coarse grid data of the electromagnetic field of the kth GIS device U k And coarse grid field strength data I k Comprehensive compensation of fine-grid enhanced data

[0194]

[0195]

[0196]

[0197] In the formula, and These represent the regularized electromagnetic field fine mesh enhanced mesh structure characteristics, regularized fine mesh enhanced field strength characteristics, regularized electromagnetic field coarse mesh data, and regularized coarse mesh field strength data, respectively; ||·||2 is the matrix 2-norm.

[0198] (2) Based on the regularized coarse mesh structure data and the regularized fine mesh enhanced mesh structure features, calculate the mesh structure consistency; specifically, the formula for calculating mesh structure consistency is:

[0199] In the formula, D S To ensure grid structure consistency.

[0200] (3) Based on the normalized coarse-grid field strength data and the normalized fine-grid enhanced field strength characteristics, calculate the grid field strength consistency; specifically, the formula for calculating grid field strength consistency is:

[0201] In the formula, D E To ensure grid field strength consistency.

[0202] (4) Based on the mesh structure consistency and the mesh field strength consistency, calculate the mesh field strength loss compensation; specifically, the calculation formula for mesh field strength loss compensation is: C k =λ·D S +(1-λ)·D E

[0203] In the formula, C k This is for compensation of field strength loss in the grid.

[0204] (5) Based on the fine mesh enhancement structure characteristics, the mesh field strength loss compensation, and the similarity characteristics, calculate the fine mesh enhancement structure data; specifically, the calculation formula for the fine mesh enhancement structure data is:

[0205] In the formula, Enhance structural data with finer meshes.

[0206] (6) Based on the fine-mesh enhanced field strength characteristics, the mesh field strength loss compensation, and the similarity characteristics, calculate the fine-mesh enhanced field strength data; specifically, the calculation formula for the fine-mesh enhanced field strength data is:

[0207] In the formula, The field strength data is enhanced by a finer grid; λ and γ are scaling factors.

[0208] As a further preferred technical solution, the differential deep learning network model used in this embodiment is a pre-trained model that can be used to calculate the similarity between the grid structure and the grid field strength. The training process includes the following steps: using a GIS electromagnetic field coarse grid historical simulation dataset, and training the differential deep learning network model using the historical simulation dataset until the total loss function of the model reaches its minimum value. The total loss function includes a constraint loss function based on Maxwell's equations and a field strength and structure consistency loss function, expressed as: L total =ηL consistency +ζL Maxwell

[0209] In the formula, L total Let L be the total loss function. consistency Let L be the loss function for field strength and structural consistency. Maxwell Let η be the Maxwell-constrained loss function, and ζ be the weight parameters.

[0210] It should be noted that this embodiment employs a multi-dimensional loss function design. By combining structural similarity loss, field strength consistency loss, and physical constraint loss, it comprehensively considers multiple important characteristics of the simulation data. This multi-dimensional loss function design is beneficial in ensuring simulation accuracy while also allowing for the adjustment of the weights between different losses to adapt to specific simulation requirements and optimization objectives.

[0211] Furthermore, the formula for the field strength and structural consistency loss function is expressed as follows:

[0212]

[0213] L consistency =σL struct +δL field

[0214] In the formula, L represents the gradient obtained by applying the finite difference method. struct For structural similarity loss, L field The loss is the uniformity of field strength, where σ and δ are scaling factors. To enhance the mesh structure features with finer meshes, U k Data with a coarse grid structure. To enhance the field strength characteristics with finer mesh, I k For coarse-grid field strength data, || || 2 It is the 2-norm of the vector.

[0215] It should be noted that this embodiment ensures that the enhanced electromagnetic field data maintains a high degree of consistency with the original data in terms of structure and field strength through structural similarity loss and field strength consistency loss. This consistency not only promotes the convergence of simulation data, but also ensures the physical authenticity of simulation results and improves the reliability of simulation.

[0216] Furthermore, by introducing a scaling factor, a flexible loss function adjustment mechanism is provided. This mechanism can not only adjust the importance of different parts of the loss function according to the specific task, but also improve the model's generalization ability and adaptability when processing different types of electromagnetic field data.

[0217] Furthermore, the formula for the constraint loss function based on Maxwell's equations is expressed as follows:

[0218]

[0219]

[0220]

[0221]

[0222] L Maxwell =w1·L Gauss +w2·L Gauss-Mag +w3·L Faraday +w4·L Ampere

[0223] In the formula, L is the magnetic field component. Gauss For Gaussian law loss, L Gauss-Mag Loss due to Gauss's law of magnetism, L Faraday Loss due to Faraday's law of electromagnetic induction, L Ampere For the loss due to Ampere's law, κ is the proportionality constant, and w1, w2, w3, and w4 are proportionality factors. Indicates curl, ||·|| 2 Let be the vector 2-norm; μ0 is the permeability of vacuum, and ε0 is the permittivity of vacuum. It is fine-grid enhanced field strength data.

[0224] It should be noted that the physical constraint loss function ensures that the enhanced electromagnetic field data strictly follows the basic laws of electromagnetism through the key components of Maxwell's equations—Gauss's law, Gauss's magnetic law, Faraday's law of electromagnetic induction, and Ampere's law. This constraint makes the electromagnetic field data output by the model not only numerically accurate but also physically rigorous, increasing the scientific rigor and practicality of the model.

[0225] It should be noted that when combining differential methods with generative adversarial networks (GANs) to form differentially enhanced GANs, the difference-based convolutional kernels are directly integrated into the generator architecture, thereby introducing differential enhancement of spatial features. The adversarial aspect here is mainly reflected in the model's loss; that is, when using GANs, the model's total loss function includes not only the physical constraint loss function and the field strength and structural consistency loss function, but also the adversarial loss component.

[0226] In summary, the loss function designed in this embodiment combines the advantages of traditional physical models and deep learning. Through precise mathematical expression and physical constraints, it helps to optimize the stability and convergence speed of the algorithm, thereby improving training efficiency. The overall loss function, by combining electromagnetic field theory and deep learning techniques, not only enhances the model's ability to process electromagnetic field simulation data but also ensures the accuracy and physical realism of the simulation results, providing an effective optimization strategy for electromagnetic field simulation.

[0227] Example 5: Based on the content disclosed in Example 1, this example specifically uses a generative adversarial network (GAN) model to enhance the coarse mesh simulation results, obtaining enhanced fine mesh data. The specific implementation principle is shown in Figure 12: The enhancement model uses a GAN model, which includes a generator and a discriminator. The generator extracts features from the coarse mesh simulation results and performs cyclic enhancement on the extracted features to obtain enhanced fine mesh data. The discriminator includes a structural similarity module, a field strength consistency module, and an accuracy discrimination module, wherein: the structural similarity module calculates a first confidence value for the structural information contained in the enhanced fine mesh data and the simulated real fine mesh data; the field strength consistency module calculates a second confidence value for the field strength information contained in the enhanced fine mesh data and the simulated real fine mesh data; and the accuracy discrimination module narrows the distance between the distributions of the enhanced fine mesh data and the simulated real fine mesh data.

[0228] It should be noted that in this embodiment, real coarse-grid data and real fine-grid data are pre-input into the generative adversarial network for training until the overall optimization loss function is minimized. The trained generator is then used as a data augmentation model to augment the real coarse-grid data generated in real time. The generator network is responsible for learning the latent distribution characteristics of real samples and synthesizing new samples; the discriminator network is responsible for distinguishing between real samples and newly generated samples. The generator and discriminator achieve an adversarial effect through alternating training, continuously improving their respective generation and discrimination capabilities, ultimately leading to a Nash equilibrium between the generator and discriminator networks.

[0229] In this embodiment, the discriminator is designed to determine whether the input data meets the requirements. The discriminator consists of a structural similarity module, a field strength consistency module, and an accuracy discrimination module. It divides the grid data into structural data and field strength data for comparative analysis, which can effectively improve the discrimination accuracy.

[0230] Since GIS equipment has a certain physical structure, the data points for electromagnetic field simulation cannot exceed a certain range. In the PU-GAN (Projection Upsampling Generative Adversarial Network) network designed in this embodiment of the invention, the coarse grid data points are amplified by upsampling. The generator uses cyclic amplification to enhance the electromagnetic field between coarse grid data points, which to a certain extent ensures that it does not exceed the scope of the physical structure of GIS and improves the accuracy of GIS electromagnetic field simulation data enhancement.

[0231] As a further preferred technical solution, as shown in Figure 13, the generator includes several cascaded feature extraction and expansion networks, each of which is followed by a multilayer perceptron. Each feature extraction and expansion network includes a feature extraction module (Feature Extraction) and a feature expansion module connected in sequence. The output of the Feature Extraction module in the previous level of the feature extraction and expansion network is connected to the output of the Feature Extraction module in the next level of the feature extraction and expansion network. As shown in Figure 14, the Feature Extraction module includes a first MLPs layer, a KNN layer, a second MLPs layer, a maxpooling layer, and a compression layer connected in sequence. The coarse-grid simulation result serves as the input to the first MLPs layer. The outputs of the first MLPs layer and the second MLPs layer are concatenated and output to the maxpooling layer. The coarse-grid simulation result is concatenated with the output of the maxpooling layer and serves as the input to the compression layer. The output of the compression layer is connected to the feature expansion module. The feature expansion module is used to add a vector encoded as 1 or -1 to each feature output by the feature extraction module to generate positional perturbation.

[0232] It should be noted that the generator designed in this embodiment is used to enhance coarse grid data in GIS and generate the required fine grid data. The generator is designed with cyclic enhancement, which can adjust the enhancement factor of the coarse grid according to the needs. The enhancement factor can better preserve the characteristics of the coarse grid data.

[0233] Specifically, the feature extraction module in this embodiment works as follows: First, the N×4 dimensional information composed of the three-dimensional spatial coordinates (x, y, z) and field strength value v of each input sampling point is extracted into fixed-scale features C' through the first MLPs layer. Then, the features are grouped into N×K groups using the KNN layer, and the group features are further optimized using the second MLPs layer to obtain G-channel features. The features output by the second MLPs layer are merged with the features output by the first MLPs layer to obtain a new G-channel feature (a total of N×K×(2G+C')). Finally, the grouped information is condensed through maxpooling to obtain the final N×(2G+C'), while adding the N×4 dimensional features of the initially output sampling points. This feature, as a multi-scale feature aggregation of the module, will be fed into subsequent modules for reuse to improve reconstruction accuracy and parameter efficiency while reducing model size. Under the compression effect of the last data compression layer, N×C features are obtained. After copying the obtained N×C features, a vector encoded as 1 or -1 is added after each feature to generate positional perturbation. Then, the 2N×(C+1) features are regressed into 2N×4 data using the subsequent multilayer perceptron (MLP).

[0234] It should be noted that after multiple iterations of feature extraction and expansion, a 2 can be obtained. r The N×C features were finally regressed to the four-dimensional data using a multilayer perceptron (MLP). r N×4.

[0235] It should be noted that the value of r depends on the difference in accuracy between the input simulation coarse and fine meshes, and this embodiment does not impose specific limitations.

[0236] Specifically, different modules target different data precisions. If it is necessary to pass the features from the previous layer to the next layer, a good interpolation technique is needed to match the features at different levels. In this embodiment, when copying the features output by the feature extraction module in the previous level feature extraction and expansion network to the output of the feature extraction module in the next level feature extraction and expansion network, an interpolation technique is used to match the features at different levels.

[0237] As a further preferred technical solution, this embodiment uses bilateral interpolation technology, which utilizes the features around the target point to perform interpolation, and interpolates point p. i The feature value interpolation value output from the previous feature extraction module is:

[0238] Where, p i p represents the coordinates and electric field strength of the i-th point. i′ These are the coordinates and electric field strength of the i′-th point, f i f is the eigenvalue corresponding to the i-th point.i′ It is the feature value corresponding to the i′-th point. This is the set of data points to be inserted, where the joint weighting function is represented as follows:

[0239]

[0240] Here, the width parameters r and h are the average distances from a point to its nearest neighbor, and || represents the magnitude of the vector.

[0241] This embodiment is for coarse grid data augmentation. Using bilateral interpolation can effectively avoid the overlap between the generated grid data and the source grid data, thus avoiding waste.

[0242] As a further preferred technical solution, as shown in Figure 15, the structural similarity module includes a first structural branch network, a second structural branch network, and a first regression network. The enhanced structural information contained in the fine mesh generation data and the real fine structural information contained in the real fine mesh data are respectively used as inputs to the first structural branch network and the second structural branch network. The outputs of the first structural branch network and the second structural branch network are multiplied by a matrix to output a structural feature matrix. The structural feature matrix is ​​used as input to the first regression network to obtain a first confidence value between the enhanced structural information and the real fine structural information.

[0243] Specifically, as shown in Figure 15, both the first and second structural branch networks include a multilayer perceptron and a max-pooling layer connected in sequence. The outputs of the first and second structural branch networks, after matrix multiplication, output a structural feature matrix as follows:

[0244]

[0245]

[0246] In the formula, W s The structural feature matrix, The structural features of the data generated for the fine mesh, The structural features of the real fine-grid data are α and β respectively. The modulation weights are defined by MLPs(), which represents the multilayer perceptron operation, and maxpool(), which represents the max pooling operation. To generate structural information for the i-th point of the data in the fine mesh, This represents the structural information of the i-th point in the real fine-grid data.

[0247] Specifically, as shown in Figure 15, the first regression network includes a first self-attention unit, a multilayer perceptron, and a fully connected layer connected in sequence. The first self-attention unit is used to generate structural attention weights based on the structural feature matrix. The structural attention weights are passed through the multilayer perceptron and the fully connected layer to output a first confidence value.

[0248] As a further preferred technical solution, as shown in Figure 16, the field strength consistency module includes a first enhancement branch network, a second enhancement branch network, and a second regression network. The enhanced field strength information contained in the fine mesh generated data and the real fine field strength information contained in the real fine mesh data are respectively used as inputs to the first enhancement branch network and the second enhancement branch network. The outputs of the first enhancement branch network and the second enhancement branch network are multiplied by a matrix to output a field strength feature matrix. The field strength feature matrix is ​​used as input to the second regression network to obtain a second confidence value between the enhanced field strength information and the real fine field strength information.

[0249] Specifically, both the first and second enhancement branch networks include a multilayer perceptron and an activation function connected in sequence. The outputs of the first and second enhancement branch networks, after matrix multiplication, output a field strength feature matrix as follows:

[0250]

[0251]

[0252] In the formula, W E The field strength characteristic matrix, The field strength characteristics of the data generated for the fine mesh are determined. The field strength characteristics of the real fine-grid data are α′ and β′ respectively. Modulation weights, The field strength information is generated for the i-th point of the data in the fine mesh. This represents the field strength information at the i-th point of the real fine-grid data.

[0253] Specifically, the second regression network includes a second self-attention unit, a multilayer perceptron, and a fully connected layer connected in sequence. The second self-attention unit is used to generate field strength attention weights based on the field strength feature matrix. The field strength attention weights are then passed through the multilayer perceptron and the fully connected layer to output a second confidence value.

[0254] As a further preferred technical solution, as shown in Figure 17, the self-attention unit used in this embodiment is used for enhanced feature integration. Specifically, the input features are converted into G, H, and F through three independent MLPs, and then attention weights are generated from G and H: M = F(f softmax (G T H)) T

[0255] Among them, f softmax This represents the softmax function.

[0256] Then, the weighted features M are obtained, and finally, the output features are generated. The output features are the sum of the input features and the weighted features.

[0257] As a further preferred technical solution, in this embodiment, the outputs of the structural similarity module and the field strength consistency module are both connected to the accuracy discrimination module. The accuracy discrimination module is used to control the generation of grid structure and field strength data and distinguish the accuracy of the generation results. The accuracy discrimination module can be defined as a region-level fully convolutional binary classifier, which can reduce the distance between the distribution of generated fine grid simulation data and numerical method fine grid simulation data, while using least squares adversarial loss to ensure the stability of the training process.

[0258] The adversarial loss between fine-mesh generated data and real fine-mesh data is defined as follows: L gan =(D(f) real )-1) 2 +(D(f enh )) 2

[0259] In the formula, D(I) real ) and D(I enh The numbers f and f represent the fine-mesh generation data f generated by the discriminator D from the generator output. enh Compared with real fine-grid data f real The confidence value predicted in the middle.

[0260] As a further preferred technical solution, the formula for the overall optimization loss function is expressed as: L adv =λ1L ss +λ2L col +L gan

[0261] In the formula, L adv Let L be the overall optimization loss function. ss L is a first confidence value for the structural information contained in the generated fine mesh data and the actual fine mesh data. col L is a second confidence value for the field strength information contained in the generated fine-mesh data and the actual fine-mesh data. ganThe adversarial loss between the generated data and the actual fine-mesh data is defined as λ1 and λ2, which are weighting parameters.

[0262] The overall optimized loss function set in this embodiment comprehensively considers the structure and field strength confidence of the generated fine mesh and the real fine mesh. Its advantages are: (1) It promotes the generator to generate realistic data: By optimizing the loss function, the generator is forced to generate samples that are as close as possible to the distribution of real data. (2) It improves the accuracy of the discriminator: The discriminator improves its ability to distinguish between real data and generated data by optimizing the loss function. This includes enabling the discriminator to more accurately identify the data generated by the generator, and to better distinguish between generated data and real data. (3) It achieves a balance between the generator and the discriminator: Optimizing the loss function ensures that there is a balance between the generator and the discriminator, avoiding over-optimization of one of them and resulting in training instability. This balance helps the generator generate more realistic data, while ensuring that the discriminator can effectively distinguish between real data and generated data.

[0263] Example 7: Based on the content disclosed in Example 1, this example describes the specific implementation of the location searcher in Example 1 as follows: The simulation enhancement data includes enhanced discharge location and enhanced discharge intensity. The location searcher includes: a state transition model construction module, which, based on the enhanced discharge location and enhanced discharge intensity at each time step, uses a Markov decision model to reconstruct the state transition model corresponding to the process of exploring the actual discharge location of the local discharge source using the coarse localization result of the local discharge source as the center of the ant colony algorithm; a heuristic spatial parameterization module, which uses a heuristic learner based on a graph neural network to generate a heuristic metric, converting the state transition model into a location exploration model that is affected by the heuristic metric and requires graph traversal; and an iterative search module, which uses a neural-guided perturbation interleaved local search algorithm to iteratively search the location exploration model to obtain the actual discharge location of the local discharge source.

[0264] As shown in Figure 18, this embodiment uses a data-driven approach to automatically design and enhance the heuristic rules in the ant colony algorithm, thereby more effectively solving the rapid iterative search for the optimal matching position of the discharge source. The deep ant colony algorithm uses graph neural networks to generate heuristic metrics, reducing the need for expert knowledge, and combines probabilistic local search to achieve better performance and ensure efficient solution.

[0265] As a further preferred technical solution, the state transition model construction module is specifically used to: construct the state space corresponding to the actual discharge location of the partial discharge source in the simulator, based on the simulated discharge location and simulated discharge intensity at each moment.

[0266]

[0267] In the formula: p represents the state corresponding to the local discharge power node i discovered at time t in the simulator. t E represents the simulated discharge position at time t. t This represents the simulated discharge intensity at time t. The initial state is defined as the initial discharge intensity E and initial discharge position P of the partial discharge source obtained from the full-time-domain waveform diagram of partial discharge based on GIS, and T represents the exploration period, t = 1, 2, ..., T; The action space corresponding to the exploration of the actual discharge position of the partial discharge source in the simulator is constructed.

[0268] In the formula: This represents the action at time t, where the action is selected using an ε-greedy strategy. The state transition model corresponding to the process of the ant colony algorithm exploring the actual discharge location of the partial discharge source with the initial discharge location as the center is as follows:

[0269] In the formula: Indicates the action The probability of transitioning from discharge node i to discharge node j. τ represents the state corresponding to node j of the partial discharge source discovered at time t+1 in the simulator. ij η represents the pheromone concentration corresponding to the transition from node i to node j. ij τ represents the heuristic function for the transition from node i to node j. im This indicates a transition from node i to the node included in the allowed list. S The pheromone concentration corresponding to node m in the diagram, η im This indicates a transition from node i to the node included in the allowed list. S The heuristic function corresponding to node m in the diagram, where α and β represent the pheromone concentration and the weights of the heuristic function, respectively. S This represents the set of non-uniform mesh local discharge power nodes that can be selected in the next time step.

[0270] Furthermore, the reward function for the transition from node i to node j is: E′ t E represents the measured discharge intensity obtained by processing the full time-domain waveform acquired by the sensor at time t. t This represents the partial discharge intensity obtained from the simulation at time t.

[0271] It should be noted that this embodiment utilizes Markov models to formalize and model the state transition process in the ant colony algorithm. This modeling can help understand and analyze the ant colony algorithm's search process in the solution space, including key steps such as ant path selection and pheromone updates.

[0272] As a further preferred technical solution, the heuristic spatial parameterization module is used to perform the following steps: Extracting edge features from the l-th layer of a graph neural network, consisting of the connection between the i-th node and the j-th node, using a multilayer perceptron. And mapped to the heuristic metric η θ The state transition model is then converted into a location exploration model that is influenced by heuristic metrics and requires T-step graph traversal:

[0273] In the formula: This represents a location exploration model. This indicates that the state corresponding to the local discharge power node i is discovered at time t in the simulator. This represents the state corresponding to the local discharge power node i at time t+1 in the simulator, where T represents the exploration period, t = 1, 2, ..., T.

[0274] It should be noted that this embodiment constructs a heuristic learner based on graph neural networks to heuristically parameterize the exploration of local discharge sources in GIS. By using graph neural networks to generate heuristic metrics and combining them with probabilistic local search, better performance is achieved to ensure efficient solution.

[0275] Furthermore, the propagation characteristics of the l-th layer of the graph neural network are as follows:

[0276]

[0277] In the formula: This represents the feature of the i-th node in the l-th layer. This represents the feature of the j-th node in the l-th layer. This represents the feature of the i-th node in the (l+1)-th layer. This represents the edge feature of the l-th layer, which connects the i-th node and the j-th node. U represents the edge feature of the (l+1)th layer, which is the connection between the i-th node and the j-th node. l V l P l Q l and R l This represents the learnable parameters of the l-th layer. This represents the activation function, and BN represents batch normalization. This indicates that the aggregation calculation is performed on the neighborhood of the i-th node. Let represent the neighborhood of the i-th node, σ represent the Sigmoid function, ⊙ represent the Hadamard product, and l = 1, 2, ..., L.

[0278] It should be noted that when the ant colony algorithm explores the location of local discharges, it constructs a graph by traversing the discharge nodes, and the nodes of the graph neural network correspond to the discharge nodes.

[0279] As a further preferred technical solution, the location searcher also includes a training module for performing the following steps: training a graph neural network-based heuristic learner using a gradient strategy, wherein the objective function used during the training process is... for:

[0280] In the formula: This represents the objective function used to explore the actual discharge location of a partial discharge source using a neurally guided perturbation-interleaved local search algorithm, where W represents the equilibrium... and The parameters, In the heuristic metric η θ The expected value of the actual discharge location of the partial discharge source under the influence of [the following factors] Let f(·) represent the state space corresponding to the actual discharge location of the partial discharge source, and let f(·) represent the objective function; where the objective function is... The gradient is The formula is expressed as:

[0281] In the formula: This represents the average target value for directly exploring the actual discharge location of the partial discharge source. This represents the average target value used in a local search algorithm with neurally guided perturbation interleaving to explore the actual discharge location of the local discharge source. In the heuristic metric η θ The gradient of the actual discharge location of the partial discharge source is explored under the influence of the gradient; when the maximum number of iterations is reached... Or the maximum number of iterations has not been reached. But the objective function Training ends when θ represents the minimum threshold corresponding to the objective function.

[0282] As a further preferred technical solution, the iterative search module is used to perform the following steps: Iteratively search the location exploration model using a local search algorithm to obtain a locally optimal solution; specifically, the locally optimal solution... The formula is expressed as:

[0283]

[0284] In the formula: f(·) represents the objective function, and LS represents the local search operator. Let E' represent the state space corresponding to the actual discharge location of the partial discharge source. LS() represents the local search for the discharge location under the local search operator. t E represents the measured discharge intensity obtained by processing the full time-domain waveform acquired by the sensor at time t. t Let represent the partial discharge intensity obtained from the simulation at time t, and T represent the exploration period, t = 1, 2, ..., T.

[0285] The local optimum obtained in the current iteration is subjected to neural-guided perturbation. Through interleaved local searches using neural-guided perturbations, the optimal exploration scheme for the actual discharge location of the local discharge source in the current iteration is obtained. Specifically, the optimal exploration scheme for the actual discharge location of the local discharge source in the previous iteration is expressed by the following formula:

[0286] In the formula: This represents the optimal exploration scheme for the actual discharge position of the local discharge source in the current iteration, obtained by neural-guided perturbation of the local optimal solution. This is a locally optimal solution; T p η represents the number of moves in the disturbance. θ This represents a heuristic metric.

[0287] It should be noted that the optimal exploration scheme for the actual discharge location of the local discharge source obtained in the current iteration, based on the perturbation, is further subjected to a local search to obtain the optimal solution for the local discharge source. Local optimal solution obtained by local search

[0288] After T NLS The optimal exploration scheme for GIS local discharge sources was obtained through a series of iterations of neural-guided perturbation-interleaved local search.

[0289] In the formula: argmin() represents the set of parameters or variables that make the function achieve its minimum value.

[0290] It should be noted that when using a local search algorithm to iteratively search the location exploration model, local optima may occur. In this embodiment, a heuristic metric obtained from training a graph neural network is used to perturb the local optima. The heuristic metric can help the algorithm quickly locate and escape local optima, guide the search process toward the global optimum, and guide the search process to explore the solution space more effectively.

[0291] After all partial discharge power source nodes are selected, the pheromone concentration among each partial discharge power source node is updated; specifically, the pheromone concentration update formula is expressed as: τ ij (k+1)=ρτ ij (k)+Δτ ij (k, k+1)

[0292] In the formula: ρ represents the update coefficient, τ ij (T) represents the pheromone concentration during the k-th exploration cycle, τ ij (k+1) represents the pheromone concentration during the (k+1)th exploration cycle, Δτ ij (k,k+1) represents the change in pheromone concentration from the k-th exploration cycle to the (k+1)-th exploration cycle.

[0293] When the iterative search reaches the iterative convergence condition, the optimal exploration scheme for the actual discharge location of the partial discharge source is determined, and the actual discharge location of the partial discharge source is matched based on the optimal exploration scheme.

[0294] In this embodiment, the ant colony algorithm simulates the search process of ants in a graph to find possible locations of partial discharge sources. Graph neural networks can effectively propagate partial discharge signal information through message passing and feature aggregation, and combine the search results from the ant colony algorithm to infer and locate the source. The parallel computing power and efficient feature learning of graph neural networks enable them to quickly process large-scale graph data, making them suitable for real-time or near-real-time partial discharge localization tasks. As a heuristic optimization method, the ant colony algorithm, combined with the feature learning capabilities of graph neural networks, can intelligently guide the search process and quickly discover possible discharge source locations. This method effectively avoids getting trapped in local optima and improves global search capabilities, enabling fast, real-time, and high-precision localization.

[0295] It should be noted that, as shown in Figure 19, this embodiment uses the initial discharge location as the center and the radius of the spherical area for partial discharge location as the radius of the GIS cross-section to determine the optimal exploration scheme, and uses the path of the optimal exploration scheme to match the actual discharge location of the partial discharge source, thereby achieving rapid location of the partial discharge source.

[0296] Example 8: Based on the content disclosed in Example 1, this example provides a detailed description of the early warning device mentioned in Example 1 as follows: The early warning device is equipped with a fault classification module, which includes: a deep feature extraction network to obtain deep features corresponding to the full-time waveform, wherein the deep feature extraction network includes a feature extraction network and an output network, the feature extraction network is formed by stacking several feature extraction layers, each of the feature extraction layers including a channel attention module and a spatial attention module connected in sequence; a reinforcement learning unit to construct a dataset using the deep features, and to train a GIS partial discharge feature matching Markov model using a deep reinforcement learning algorithm to obtain the optimal GIS partial discharge feature matching result; and an early warning unit to generate early warning information based on the optimal GIS partial discharge feature matching result and the actual discharge location of the partial discharge source, and to provide fault early warning.

[0297] Specifically, as shown in Figure 20, the deep feature extraction network includes a feature extraction network and an output network. The feature extraction network is formed by stacking several feature extraction layers. Each feature extraction layer includes a channel attention module and a spatial attention module connected in sequence. This embodiment adopts a dual attention mechanism, that is, fusing spatial attention and channel attention to extract features from the full-time-domain waveform signal. This can not only improve the distinguishability of GIS partial discharge features and reduce redundancy, but also consider features from a spatial perspective. It can consider features from different angles, which is very helpful for the extraction of GIS partial discharge features. At the same time, by stacking multiple deep layers, sample information can be fully and effectively utilized. In addition, for the extraction of GIS partial discharge features, a Markov model for GIS partial discharge feature matching is trained based on a deep reinforcement learning algorithm to obtain the optimal GIS partial discharge feature matching strategy. When partial discharge occurs in GIS, feature matching of GIS partial discharge signals is realized to determine the GIS partial discharge type.

[0298] Example 9 This example proposes an implementation method corresponding to the GIS field-electricity fusion real-time state perception and early warning system in Example 1 above, as shown in Figure 21. The method includes the following steps: When monitoring the electromagnetic field signal inside the GIS containing partial discharge signals based on pre-selected strongly correlated features, a full-time domain waveform diagram containing the time-domain waveform diagram and frequency-domain waveform diagram of the partial discharge signal is obtained; Based on the full-time domain waveform diagram of the GIS partial discharge, the partial discharge source is coarsely located, and the coarse location result is used as the initial injection point for simulation enhancement calculation to obtain the simulation enhancement data of the partial discharge source; The measured discharge intensity calculated based on the full-time domain waveform diagram is iteratively searched with the simulation enhancement result to obtain the actual discharge location of the partial discharge source; An early warning is issued based on the full-time domain waveform and the early warning information containing the actual discharge location of the partial discharge source.

[0299] It should be noted that the method described in this invention is used to realize GIS fault early warning by utilizing the GIS field-electricity fusion real-time status perception and early warning system described in the above embodiments. Other embodiments can refer to the above embodiments, and will not be repeated here.

[0300] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0301] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0302] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A GIS field-electricity fusion real-time state perception and early warning system, characterized in that, The system includes an embedded device for synchronous signal acquisition and measurement and an industrial control computer, in which a simulator, a location searcher, and an early warning device are deployed; The embedded device for synchronous acquisition and measurement of signals is used to obtain a full-time-domain waveform diagram containing the time-domain waveform diagram and the frequency-domain waveform diagram when monitoring the electromagnetic field signal inside the GIS containing the partial discharge signal based on the pre-screened strong correlation features. The simulator is used to coarsely locate the partial discharge source based on the full-time-domain waveform diagram of GIS partial discharge, and use the coarse location result as the initial injection point to perform simulation enhancement calculation to obtain simulation enhancement data of the partial discharge source. The location searcher is used to iteratively search the measured discharge intensity calculated based on the full-time-domain waveform diagram and the simulation enhancement results to obtain the actual discharge location of the partial discharge source. The warning device is used to issue a warning based on the time-domain full waveform and warning information including the actual discharge location of the partial discharge power source.

2. The GIS field-electric fusion real-time condition sensing and early warning system of claim 1, wherein, The embedded device for synchronous signal acquisition and measurement includes: The signal acquisition location determination module is used to determine the installation location of each sensor in the signal acquisition module; The signal acquisition module is used to acquire electromagnetic field signals inside the GIS. The signal processing front end is used to perform signal conditioning on the electromagnetic field signal transmitted by the signal acquisition module, and the conditioned signal is transmitted to the analog-to-digital converter. An analog-to-digital converter is used to convert conditioned signals from analog signals to digital signals before transmitting them to a signal measurement module. The signal measurement module is used to monitor the digital signal based on the pre-selected strong correlation features. When it is determined that a partial discharge signal is detected, a full-time-domain waveform diagram including the time-domain waveform diagram and the frequency-domain waveform diagram of the partial discharge signal is obtained. The strong correlation features include time-domain strong correlation features and frequency-domain strong correlation features.

3. The GIS field-electric fusion real-time condition sensing and early warning system of claim 2, wherein, The embedded device for synchronous signal acquisition and measurement also includes a feature filtering module, which uses the time-frequency domain features used to detect partial discharge signals as nodes of a graph neural network to filter the correlation strength, thereby obtaining the time-domain strong correlation features and the frequency-domain strong correlation features, wherein the lines between nodes represent the relationship between the features represented by the nodes.

4. The GIS field-electric fusion real-time condition sensing and early warning system of claim 3, wherein, The feature filtering module includes: The feature map construction unit is configured to construct a complete feature map G=(V, E) of the graph neural network, wherein V is a node set, E is an edge set, x v , x co[v] , h ne[v] , x ne[v] respectively represent a feature vector of a node v, a feature vector of an edge, a state vector of the node v and its peripheral nodes, and a feature vector of the peripheral nodes of the node v. The aggregation update unit is used to aggregate and update the information of input nodes and edges according to the local transition function of the updated node state, and output the node labels. The formula is expressed as: h v = f(x v ,x co[v] ,h ne[v] ,x ne[v] ) o v = g(h v , x v ) where f(·) is a local transition function that updates the node state, g(·) is a local output function, h v is the state vector learned through iterative learning of the graph neural network, o v is the node label; The iterative unit is used to calculate the correlation between a node and its neighboring nodes in the continuous iteration of the graph neural network, so as to obtain strong correlation features in the time domain and strong correlation features in the frequency domain.

5. The GIS field-electric fusion real-time condition sensing and early warning system of claim 4, wherein, The iterative unit is specifically used for: Let H, O, X and X N are the stacked vectors of all nodes' state vectors, all output labels, all nodes and edges' feature vectors and all nodes' feature vectors, respectively. The formula can be written in a more compact form as O = G(H, X N ) H t+1 = F(H t ,X) where H t denotes the t-th iteration of H t+1 = F(H t , X) denotes the state vector of all nodes of the t+1-th iteration obtained from the feature vector and the state vector of the t-th iteration by the global transition function, F(·) and G(·) are the global transition function and the global output function, respectively, which are stacks of the local transition functions f(·) and the local output functions g(·) of all nodes; During the iteration process, nodes with similar states, nodes with complementary states, and the degree of influence of each node on the overall graph neural network are determined. One node with similar states is selected, nodes with complementary states are merged into one node, and the node with the greatest influence on the overall graph neural network is selected as the strong correlation feature.

6. The GIS field-electric fusion real-time condition sensing and early warning system of claim 2, wherein, The signal measurement module includes: A buffer unit is used to buffer the digital signal; The first monitoring unit is used to extract the strong time-domain correlation features from the digital signal, determine the detected partial discharge signal based on the strong time-domain correlation features, and output the first time-domain waveform during discharge. The second monitoring unit is used to extract the strong correlation features in the frequency domain from the digital signal, determine the detected partial discharge signal based on the strong correlation features in the frequency domain, and output the frequency domain waveform during discharge. The waveform synthesis unit is used to obtain the full-time waveform diagram of the partial discharge signal based on the first time-domain waveform diagram and the frequency-domain waveform diagram.

7. The GIS field-electric fusion real-time condition sensing and early warning system of claim 6, wherein, The second monitoring unit includes: The Fourier transform subunit is used to perform at least two fast Fourier transforms on the electromagnetic field signal to obtain frequency domain data of at least two frequency bands. The normalization subunit is used to extract the strong correlation features of the frequency domain data of each frequency band, and after analyzing and matching the strong correlation features, it determines that the frequency domain data of each frequency band exceeds the second discharge threshold. Then, it normalizes the frequency domain waveforms corresponding to each frequency band to obtain the frequency domain waveform diagram during discharge.

8. The GIS field-fusion real-time condition sensing and warning system of claim 6, wherein, The waveform synthesis unit includes: A waveform conversion subunit is used to convert the frequency domain waveform into a second time domain waveform during discharge. The waveform synthesis subunit is used to synthesize the first time-domain waveform diagram and the second time-domain waveform diagram to obtain the full-time-domain waveform diagram of the partial discharge signal.

9. The GIS field-fusion real-time condition sensing and warning system of claim 1, wherein, The simulator includes: The simulation module is used to take the coarse location result of the partial discharge power source obtained based on the full-time waveform diagram as the initial injection point of the 3D simulation model of the GIS equipment to simulate the partial discharge phenomenon of the GIS equipment, and perform simulation calculations to obtain coarse mesh simulation results. The enhancement module is used to enhance the coarse mesh simulation results using an enhancement model to obtain enhanced simulation data.

10. The GIS field-fusion real-time condition sensing and warning system of claim 9, wherein, The simulation module includes: Spatial discrete unit is used to discretize the three-dimensional simulation model of GIS equipment into a spatial grid and determine the correspondence between the number of each grid node and the spatial coordinate data. The differential simulation unit is used to inject the coarse positioning results as initial injection points into the corresponding spatial grid to simulate the partial discharge phenomenon of GIS equipment. It converts Maxwell's equations, which describe the electromagnetic field of GIS, into a fourth-order matrix and uses the four-step HIE-FDTD algorithm to solve the fourth-order matrix, calculating the field strength data of each grid node.

11. The GIS field-fusion real-time condition sensing and warning system of claim 10, wherein, The differential simulation unit includes: A matrix reconstruction subunit is configured to convert Maxwell equations for describing a GIS electromagnetic field into a six-order matrix form: In the formulae, Let M be a vector consisting of the components of the electric and magnetic fields in a rectangular coordinate system, and let [M] be a sixth-order matrix. a matrix conversion subunit for converting a six-order matrix form into a four-order rectangular form: where [A H ] and [B H ] are six order matrices; The sub-element is used to solve the fourth-order matrix form using the four-step HIE-FDTD algorithm, and to calculate the field strength data for each grid node.

12. The GIS field-fusion real-time condition sensing and warning system of claim 11, wherein, The solution subunit is used to perform the following steps: The four-step HIE-FDTD algorithm is decomposed in the time domain into four sub-steps: n→n+1 / 4, n+1 / 4→n+2 / 4, n+2 / 4→n+3 / 4, and n+3 / 4→n+1. For each sub-step, a semi-implicit difference scheme is used to calculate the tridiagonal implicit difference of the field strength data of the grid nodes. The catch-up method is used to solve the implicit tridiagonal problem and obtain the field strength data of the grid nodes.

13. The GIS field-fusion real-time condition sensing and warning system of claim 9, wherein, The enhancement model employs a differential deep learning network model. The coarse mesh simulation results include coarse mesh structure data and coarse mesh field strength data. The differential deep learning network model includes an enhancement network and a structural similarity network connected in sequence. The enhancement network includes a self-adjusting module and a differential convolution module. The self-adjusting module and the differential convolution module are used to calculate the coarse grid structure data and the coarse grid field strength data, respectively, to obtain the fine grid enhanced grid structure features and the fine grid enhanced field strength features; The structural similarity network is used to match the structural features of the fine mesh enhanced mesh and the field strength features of the fine mesh enhanced mesh, and the similarity features are calculated. The simulation enhancement data is calculated based on the similarity features.

14. The GIS field-fusion real-time condition sensing and warning system of claim 13, wherein, The self-adjusting module includes a first convolutional layer and a second convolutional layer connected in sequence. The output features of the first convolutional layer and the output features of the second convolutional layer are output to the activation function layer after a first addition operation. The coarse grid structure data is used as the input to the first convolutional layer, the grid size supplementary information of the coarse grid structure data is used as the input to the second convolutional layer, the deviation vector of the coarse grid structure data is used as the input to the first addition operation, and the coarse grid structure data and the grid size supplementary information of the coarse grid structure data are output to the first addition operation via residual connection.

15. The GIS field-fusion real-time condition sensing and warning system of claim 13, wherein, The differential convolution module includes a differential convolution layer, a size integration layer, a self-attention mechanism layer, and a third convolution layer connected in sequence, with an activation function connected after the third convolution layer; The coarse grid field strength data is used as the input to the differential convolutional layer, which is used to calculate the field strength features of different sizes of the coarse grid field strength data using a differential algorithm. The size integration layer is used to sum the field strength features of different sizes calculated by the differential convolutional layer to obtain the renormalized field strength features.

16. The GIS field-coupled fusion real-time condition sensing and warning system of claim 15, wherein, The convolution kernel of the differential convolutional layer can be any one of the following: five-point differential convolution kernel, weighted differential convolution kernel, multi-scale differential convolution kernel, directional differential convolution kernel, nine-point differential convolution kernel, or hybrid mode differential convolution.

17. The GIS field-coupled fusion real-time condition sensing and warning system of claim 13, wherein, The structural similarity network includes a first branch network, a second branch network, a second addition operation, and a first multilayer perceptron. The outputs of the first branch network and the second branch network are both connected to the second addition operation, and the output of the second addition operation is connected to the first multilayer perceptron. The multilayer perceptron is followed by an activation function.

18. The GIS field-coupled fusion real-time condition sensing and warning system of claim 17, wherein, The first branch network includes a convolutional neural network (CNN) layer and a batch regularization operation connected in sequence, wherein an activation function is followed by the regularization operation; The second branch network includes a second multilayer perceptron, which is followed by an activation function.

19. The GIS field-coupled fusion real-time condition sensing and warning system of claim 18, wherein, The process of matching the fine-mesh enhanced mesh structural features and the fine-mesh enhanced field strength features using the structural similarity network, and calculating similarity features, includes: extracting, by using the first branch network, structure feature information of structure features of the fine grid enhanced grid structure, and the formula is expressed as: In the formulae, a feature information of a feature of the k-th fine mesh reinforcing the mesh structure, To enhance the mesh structure features of the fine mesh, CNN() is a stacked convolution operation, BatchNorm() is a batch regularization operation, and ReLU is an activation function; The field strength feature information of the fine grid enhanced field strength features is extracted by using the second branch network, and is expressed by a formula as follows: In the formulae, a feature information of the fine mesh reinforcement field strength feature described in the kth row, MLP() is the operation performed by the second multilayer perceptron to enhance the field strength characteristics of the fine mesh. The field strength and structure combined feature is calculated by using the modulation weight corresponding to the feature information of the fine mesh enhanced grid structure feature and the feature information of the fine mesh enhanced field strength feature, and is expressed by a formula as follows: In the formula, fk, combined represents the field strength and structure combination feature, and α and β represent the feature information of the fine mesh enhanced grid structure feature and the modulation weights corresponding to the feature information of the fine mesh enhanced field strength feature, respectively. Based on the combined characteristics of the field strength and structure, the similarity feature is calculated using the following formula: f k,final = ReLU(MLP final (fk,combined)) wherein f k,final is the similarity feature, MLP final () is an operation performed by the first multi-layer perceptron.

20. The GIS field-coupled fusion real-time condition sensing and warning system of claim 13, wherein, The calculation of the simulation enhancement data based on the similarity features includes: Based on the coarse mesh structure data, the coarse mesh field strength data, the fine mesh enhanced mesh structure features, and the fine mesh enhanced field strength features, respectively, the normalized coarse mesh structure data, the normalized coarse mesh field strength data, the normalized fine mesh enhanced mesh structure features, and the normalized fine mesh enhanced field strength features are calculated. Based on the regularized coarse mesh structure data and the regularized fine mesh enhanced mesh structure features, the mesh structure consistency is calculated; Based on the regularized coarse grid field strength data and the regularized fine grid enhanced field strength characteristics, the grid field strength consistency is calculated; Based on the mesh structure consistency and the mesh field strength consistency, calculate the mesh field strength loss compensation; Based on the fine mesh reinforcement structure characteristics, the mesh field strength loss compensation, and the similarity characteristics, the fine mesh reinforcement structure data is calculated; The simulation enhancement data is calculated based on the fine mesh enhanced field strength characteristics, the mesh field strength loss compensation, and the similarity characteristics.

21. The GIS field-electric fusion real-time condition sensing and warning system of claim 1, wherein, The simulation enhancement data includes enhanced discharge location and enhanced discharge intensity, and the location searcher includes: The state transition model construction module reconstructs the state transition model corresponding to the process of exploring the actual discharge location of the local discharge source with the coarse localization result of the local discharge source as the center, based on the enhanced discharge location and enhanced discharge intensity at each time step using the Markov decision model. The heuristic space parameterization module uses a graph neural network-based heuristic learner to generate heuristic metrics, transforming the state transition model into a location exploration model that is influenced by heuristic metrics and requires graph traversal. The iterative search module is used to iteratively search the location exploration model using a neural-guided perturbation interleaved local search algorithm to obtain the actual discharge location of the local discharge source.

22. The GIS field-coupled fusion real-time condition sensing and warning system of claim 21, wherein, The state transition model construction module is used for: According to the simulation discharge position and the simulation discharge intensity at each time, a state space corresponding to the exploration of the actual discharge position of the partial discharge source in the simulator is constructed In the formulae: represents the state corresponding to the partial discharge source node i explored at the tth moment in the simulator, p t represents the simulation discharge position at the tth moment, E t represents the simulation discharge intensity at the tth moment, The initial state is represented by the initial discharge intensity E and initial discharge position P of the partial discharge source obtained from the full-time waveform diagram of partial discharge based on GIS, and T represents the exploration period, t=1,2,...,T; Constructing the action space corresponding to the exploration of the actual discharge location of the partial discharge source in the simulator In the formulae: atime t, the action is selected based on an e-greedy policy The state transition model corresponding to the process of reconstructing the ant colony algorithm to explore the actual discharge position of the partial discharge source with the initial discharge position of the partial discharge source as the center is: In the formulae: indicates in the action the probability of a transition from discharge node i to discharge node j, τ represents the state corresponding to node j of the partial discharge source discovered at time t+1 in the simulator. ij η represents the pheromone concentration corresponding to the transition from node i to node j. ij τ represents the heuristic function for the transition from node i to node j. im This indicates a transition from node i to the node included in the allowed list. S The pheromone concentration corresponding to node m in the diagram, η im This indicates a transition from node i to the node included in the allowed list. S The heuristic function corresponding to node m in the diagram, where α and β represent the pheromone concentration and the weights of the heuristic function, respectively. S This represents the set of non-uniform mesh local discharge power nodes that can be selected in the next time step.

23. The GIS field-coupled fusion real-time condition sensing and warning system of claim 22, wherein, The reward function transferred from node i to node j is E′ t E represents the measured discharge intensity obtained by time-domain full waveform processing through the sensor at the tth moment t E represents the simulated partial discharge intensity at the tth moment.

24. The GIS field-coupled fusion real-time condition sensing and warning system of claim 21, wherein, The heuristic space parameterization module is configured to map edge features of the graph neural network l th layer connected by the i th node and the j th node to a heuristic metric η θ ;​ Converting the state-transition model into a position exploration model that is influenced by heuristics and requires T-step graph traversal is: In the formulae: position exploration model, represents the state corresponding to the partial discharge source node i explored at the tth time in the simulator, This represents the state corresponding to the local discharge power node i at time t+1 in the simulator, where T represents the exploration period, t = 1, 2, ..., T.

25. The GIS field-coupled fusion real-time condition sensing and warning system of claim 21, wherein, The location searcher also includes a training module for: The gradient strategy is adopted to train the heuristic learner based on the graph neural network, and the target function adopted in the training process For: In the formulae: W represents the balance of the objective function corresponding to the actual discharge position of the partial discharge source explored by the local search algorithm using the neural-guided perturbation interleaving With parameters of the user, represents the heuristic metric η θ influences the expected value of the actual discharge location of the partial discharge source, Let f(·) represent the state space corresponding to the actual discharge location of the partial discharge source, and let f(·) represent the objective function. wherein the objective function the gradient of the function f is The formula is expressed as: In the formulae: representing an average target value for directly exploring the actual discharge location of the partial discharge source, representing the average target value of the actual discharge location of the partial discharge source explored by the local search algorithm using the neural-guided perturbation interleaving, represents the heuristic metric η θ a gradient that influences the exploration of the actual discharge location of the partial discharge source; when the maximum number of iterations is reached or the maximum number of iterations is not reached But the objective function At this time, the training is ended, wherein, This represents the minimum threshold corresponding to the objective function.

26. The GIS field-coupled fusion real-time condition sensing and warning system of claim 21, wherein, The iterative search module is used for: A local search unit is used to iteratively search the location exploration model using a local search algorithm to obtain a local optimal solution. The perturbation unit is used to perform neural-guided perturbation on the local optimal solution obtained in the current iteration search. The optimal exploration scheme for the actual discharge position of the local discharge source in the current iteration is obtained through the interleaved local search of neural-guided perturbation. The pheromone concentration update unit is used to update the pheromone concentration between each local discharge power source node after all local discharge power source nodes have been selected. The exploration unit is used to determine the optimal exploration scheme for the actual discharge location of the partial discharge source when the iterative search reaches the iterative convergence condition, and to achieve the matching of the actual discharge location of the partial discharge source based on the optimal exploration scheme.

27. The GIS field-coupled fusion real-time condition sensing and warning system of claim 26, wherein, The process of the disturbance unit exploring the optimal exploration scheme of the actual partial discharge source position in the current iteration is expressed as: In the formulae: represents an optimal exploration scheme of the current iteration local discharge source actual discharge position obtained by performing neural guided perturbation on the local optimal solution, is a local optimum; T p denotes the number of moves of the perturbation, η θ denotes the heuristic measure.

28. The GIS field-coupled fusion real-time condition sensing and warning system of claim 1, wherein, The early warning device is equipped with a fault classification module, which includes: A deep feature extraction network is used to obtain deep features corresponding to ultra-high frequency signals. The deep feature extraction network includes a feature extraction network and an output network. The feature extraction network is formed by stacking several feature extraction layers. Each feature extraction layer includes a channel attention module and a spatial attention module connected in sequence. The reinforcement learning unit is used to construct a dataset using the deep features, train a Markov model for GIS partial discharge feature matching using a deep reinforcement learning algorithm, and obtain the optimal GIS partial discharge feature matching result. The early warning unit generates early warning information based on the optimal GIS partial discharge feature matching result and the actual discharge location of the partial discharge source, and performs fault early warning.

29. A GIS field-fusion real-time status sensing and early warning method, characterized in that, The method for implementing GIS fault early warning using the GIS field-electricity fusion real-time status perception and early warning system as described in any one of claims 1-28 includes: When monitoring the electromagnetic field signals inside the GIS based on the pre-screened strong correlation features, which include partial discharge signals, a full-time-domain waveform diagram containing both time-domain and frequency-domain waveform diagrams of the partial discharge signals is obtained. The partial discharge source is coarsely located based on the full-time-domain waveform diagram of GIS partial discharge, and the coarse location result is used as the initial injection point for simulation enhancement calculation to obtain the simulation enhancement data of the partial discharge source. The actual discharge location of the partial discharge source is obtained by iteratively searching the measured discharge intensity calculated based on the full-time-domain waveform diagram and the simulation enhancement results. Warnings are issued based on the full time-domain waveform and warning information including the actual discharge location of the partial discharge source.

Images