A Comprehensive Evaluation Method for Grounding Grid Safety Characteristics Based on Multi-Dimensional Joint Analysis
By constructing a multi-dimensional joint analysis method for grounding grids, the problems of single evaluation dimensions and low detection efficiency in existing technologies are solved. This enables a comprehensive evaluation of the safety characteristics of grounding grids and defect identification, provides real-time handling guidance, and improves the practicality and safety of the evaluation.
Patent Information
- Application Number
- CN202611586558.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-10-14
- Publication Date
- 2026-11-17
AI Technical Summary
Existing grounding grid safety assessment technologies have limited assessment dimensions, narrow frequency range, static time scale, and incomplete physical quantity coverage, making it impossible to fully reflect the true safety status of the grounding grid. Furthermore, traditional detection methods are inefficient, costly, and unable to achieve real-time tracking and trend prediction.
A multi-dimensional joint analysis method is adopted to construct a three-dimensional coupled distribution map of the spatial location, soil electrical properties, and electrical parameters of the entire grounding grid, obtain a broadband impedance characteristic spectrum, construct a time-space-frequency-electrical four-dimensional tensor data array, use graph attention mechanism and neural differential equations for health diagnosis, and adopt a multi-task learning architecture to output safety margin, thereby realizing a multi-dimensional safety assessment of the grounding grid.
It enables a multi-dimensional comprehensive assessment of the safety characteristics of the grounding grid, identifies defect types, provides clear handling guidelines, improves the comprehensiveness and practicality of the assessment, and reduces the risk of equipment damage and personal safety.
Smart Images

Figure CN123116531A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of power system grounding technology, and more specifically, to a comprehensive evaluation method for the safety characteristics of grounding grids based on multi-dimensional joint analysis. Background Technology
[0002] A grounding grid is a mesh-like grounding system buried at a certain depth underground, composed of multiple metal grounding electrodes and connecting conductors. It is an indispensable safety protection facility in power facilities such as power substations, converter stations, and transmission line towers. The grounding grid protects equipment and personnel, and provides shielding and voltage equalization. In the event of a short-circuit fault or transient event such as a lightning strike, the grounding grid provides a low-impedance path for the fault current to enter the ground, thereby limiting ground potential rise, contact voltage, and step voltage, ensuring the safety of equipment and personnel. With the rapid development of power grids towards higher voltage, larger capacity, and smarter systems, the electromagnetic environment of the grounding grid is becoming increasingly complex, placing higher demands on the assessment of its safety characteristics.
[0003] Practical operational experience and research show that the power frequency grounding resistance value alone is insufficient to fully reflect the true safety status of the grounding grid, mainly in the following four aspects:
[0004] (1) As the grounding grid area increases, the grounding impedance, contact potential difference and step potential difference at different spatial locations vary significantly, and a single global index cannot characterize the local safety risks of each area;
[0005] (2) The current excitation borne by the grounding grid covers multiple frequency components such as power frequency, switching operation, lightning impulse and ultra-fast transient overvoltage (VFTO). The inductive effect, skin effect and soil ionization characteristics of the grounding body change significantly under different frequencies. Existing regulations lack systematic consideration of this.
[0006] (3) During long-term operation, the grounding grid continuously faces problems such as soil corrosion, stray current erosion and unbalanced current inflow. Its conductor performance shows a dynamic deterioration trend. Traditional periodic sampling excavation inspection is not only inefficient and blind, but also cannot achieve real-time tracking and trend prediction of the health status of the grounding grid.
[0007] (4) The safety assessment of the grounding grid involves multiple physical dimensions such as electrical, thermal, and electrochemical corrosion. The existing single-parameter, single-frequency, and single-moment assessment models are no longer suitable for the actual needs of high-reliability power systems.
[0008] Furthermore, existing grounding resistance measurement methods, such as the two-point method, three-point method, potential drop method, and four-electrode method, are all limited by inherent defects such as the inability to place the voltage electrode at infinity and the distortion of the ground current field caused by the current electrode, resulting in significant measurement errors. Grounding grid corrosion and breakpoint detection mainly rely on power outage sampling excavation, selecting several locations for inspection based on empirical data of soil corrosion rates. This method is labor-intensive, costly, and time-consuming, and the selection of excavation points is often arbitrary, failing to accurately locate corrosion or breakpoints. In summary, existing grounding grid safety assessment technologies suffer from prominent problems such as a single assessment dimension, limited frequency domain range, static time scale, and incomplete coverage of physical quantities. There is an urgent need for a grounding grid safety characteristic assessment method that can comprehensively analyze multiple spatial dimensions, multiple frequency dimensions, multiple time dimensions, and multiple physical quantities. Summary of the Invention
[0009] To address the aforementioned technical problems, this invention proposes a comprehensive evaluation method for the safety characteristics of grounding grids based on multi-dimensional joint analysis. This method solves the technical problems of existing grounding grid safety evaluation methods, such as having a single evaluation dimension, being unable to distinguish defect types, being unable to handle non-uniform sampling data, and lacking specific handling guidelines. This improves the comprehensiveness and practicality of grounding grid safety evaluation.
[0010] This invention discloses a comprehensive evaluation method for the safety characteristics of grounding grids based on multi-dimensional joint analysis, comprising the following steps:
[0011] Obtain the electrical parameters and soil electrical parameters of each measuring point in the grounding grid, and construct a three-dimensional coupled distribution map of the spatial location, soil electrical parameters, and electrical parameters of the entire grounding grid.
[0012] Inject current signals into the grounding grid respectively, and obtain the broadband impedance characteristic spectrum containing steady-state and transient components;
[0013] The ternary coupling distribution spectrum and broadband impedance characteristic spectrum are repeatedly acquired at non-uniform time intervals to construct a time-space-frequency-electrical four-dimensional tensor data array;
[0014] The four-dimensional tensor data array is input into the health diagnosis model. The graph attention mechanism is used to extract the coupling features in the spatial and frequency dimensions. The time evolution trajectory of the coupling features is modeled using neural differential equations, and the impedance deviation and degradation mechanism type at each spatial location are output.
[0015] The impedance deviation and degradation mechanism type are input into the safety margin comprehensive evaluation model, and the spatial safety margin, frequency safety margin and time safety margin of each spatial location are output using a multi-task learning architecture.
[0016] A comprehensive security assessment is made based on the spatial security margin, the frequency security margin, and the time security margin.
[0017] This scheme constructs a three-dimensional coupled distribution map of the spatial location, soil electrical properties, and electrical parameters of the entire grounding grid, specifically including:
[0018] A three-dimensional spatial measurement point matrix is laid out over the entire grounding grid. In the horizontal direction, the measurement point locations are the conductor intersection nodes, the midpoints of adjacent nodes, and the outer edge expansion areas. In the vertical direction, the layered acquisition depths are the ground surface, the burial depth of the grounding grid at preset multiples above and below the burial depth, and the burial depth itself.
[0019] Grounding impedance, contact potential difference, and step potential difference were measured at each spatial measurement point. At the same measurement point, the apparent resistivity under each electrode spacing was measured by gradually increasing the electrode spacing using the Wenner quadrupole method. The apparent resistivity test curve was fitted to the true resistivity and layer thickness of each soil layer through nonlinear stratification inversion, and a three-dimensional soil electrical stratification model was established at each measurement point.
[0020] Match electrical parameters at the same spatial coordinates with soil electrical stratification parameters, so that each spatial measuring point is simultaneously associated with both electrical parameter vectors and soil electrical parameter vectors;
[0021] First, radial basis function spatial interpolation is performed on the soil layer interface depth at each measuring point to construct a continuous three-dimensional soil electrical spatial model for the entire site. The surface soil resistivity in the model is used as a prior covariate, and the electrical parameters of each measuring point are spatially estimated using the co-kriging interpolation method.
[0022] The three-dimensional soil electrical spatial model and the spatial interpolation results of electrical parameters are superimposed in the same coordinate system, so that each coordinate point in space is simultaneously associated with the soil electrical parameters and electrical parameters at its location, generating a ternary coupled distribution map.
[0023] In this scheme, a broadband impedance characteristic spectrum containing both steady-state and transient components is obtained, specifically including:
[0024] A signal injection device is connected between any two nodes of the grounding grid to sequentially inject DC current signal, power frequency current signal, multi-frequency sinusoidal sweep signal, standard lightning current impulse waveform and VFTO damped oscillation waveform into the grounding grid.
[0025] The steady-state injection point voltage and current at each frequency point are collected, the steady-state impedance amplitude and steady-state impedance phase angle at each frequency point are calculated, and the steady-state broadband impedance spectrum is formed by arranging them in frequency order.
[0026] Fourier transforms are performed on the time-domain voltage response and time-domain current response of the lightning current impulse waveform and the VFTO damped oscillation waveform. The ratio of the voltage spectrum to the current spectrum at each frequency point is used as the transient impedance amplitude, and the phase difference is used as the transient impedance phase angle to construct the transient impedance spectrum.
[0027] Align the steady-state broadband impedance spectrum with the transient impedance spectrum on the same frequency coordinate, calculate the offset rate of the transient impedance amplitude relative to the steady-state impedance amplitude at each frequency point in the overlapping frequency band as the dynamic response factor, take the frequency point where the dynamic response factor undergoes a step jump as the soil ionization characteristic frequency, and take the turning point where the steady-state impedance phase angle changes from positive to negative as the inductance-dominant frequency.
[0028] The steady-state impedance amplitude, steady-state impedance phase angle, transient impedance amplitude, transient impedance phase angle, dynamic response factor, and the relative position markers of the frequency points with respect to the soil ionization characteristic frequency and the inductance-dominant frequency together constitute the data entries of the broadband impedance characteristic spectrum, which are arranged in frequency order to form the broadband impedance characteristic spectrum.
[0029] This scheme constructs a four-dimensional tensor data array of time, space, frequency, and electrical properties, specifically including:
[0030] The soil corrosivity level is comprehensively assessed based on the soil resistivity, pH value, redox potential and stray current density at the location of the grounding grid, and the time interval between two adjacent data collections is determined by the soil corrosivity level.
[0031] During each acquisition, spatial coordinates, frequency coordinates, and time coordinates are recorded to ensure that the ternary coupling distribution spectrum and broadband impedance characteristic spectrum acquired in each acquisition maintain the same index reference in both spatial and frequency dimensions.
[0032] Each collected data is used as a three-dimensional data volume. The three-dimensional data volume uses horizontal spatial coordinates and frequency coordinates as index dimensions to store electrical parameters and soil electrical parameters at the corresponding location and frequency point.
[0033] The three-dimensional data volumes collected in each time are stacked in chronological order, and time is used as the fourth dimension index to form a four-dimensional tensor data array.
[0034] The time and space consistency of the data points in the four-dimensional tensor data array are checked, and the missing data is compensated and filled by superimposing the last collected data at the current position with the expected degradation rate, so as to obtain the preprocessed time-space-frequency-electrical four-dimensional tensor data array.
[0035] In this scheme, the health diagnosis model employs a graph attention mechanism to extract coupled features along the spatial and frequency dimensions, specifically including:
[0036] Each spatial measuring point in the grounding grid is used as a node in the graph network. The edges between the nodes are established according to the spatial distance between the measuring points to construct the topology of the grounding grid.
[0037] An initial feature vector is constructed for each node. The initial feature vector is composed of soil stratification electrical parameters at the corresponding location of the node, steady-state impedance amplitude and phase angle at each frequency point, transient impedance amplitude and phase angle, dynamic response factor, contact potential difference and step potential difference.
[0038] The initial feature vectors of each node are mapped to the latent space through a shared linear transformation matrix to obtain the latent features of each node.
[0039] For each pair of adjacent nodes, the dynamic time warping algorithm is used to compare the frequency response curves formed by arranging the steady-state impedance amplitudes of each node in frequency order to obtain the frequency response curve similarity factor. The vectors formed by arranging the soil resistivity of each layer of each node in depth order are compared to obtain the soil electrical similarity factor. The spatial proximity factor, frequency response curve similarity factor and soil electrical similarity factor are weighted and fused and then normalized to obtain the attention weight between adjacent nodes.
[0040] The attention weights are used to weight and aggregate the hidden features of each neighboring node, and then concatenated with the hidden features of the node itself to obtain the enhanced features of each node.
[0041] The enhanced features output from the previous layer are used as the input to the next layer. The linear transformation, attention weight calculation, and weighted aggregation operations are repeatedly performed until the preset number of layers is reached.
[0042] The enhanced features of each node output from the last layer are rearranged according to the spatial measurement point position and frequency point corresponding to each node to obtain the spatial-frequency coupled feature map under the time slice.
[0043] In this scheme, the health diagnosis model uses neural differential equations to model the temporal evolution trajectory of the coupling features, outputting the impedance deviation and degradation mechanism type at each spatial location, specifically including:
[0044] Arrange the spatial-frequency coupled feature maps of each time slice in chronological order, and flatten the feature map of each time slice into a one-dimensional vector as the global latent vector.
[0045] Using the global latent vector at the first acquisition moment as the initial state, a neural differential equation with time as the independent variable is constructed. The derivative of the neural differential equation is output by the differential neural network and is used to describe the continuous evolution of the grounding grid health status in the time dimension.
[0046] Starting from the initial moment, the neural differential equation is numerically integrated to obtain the global latent vector prediction value at the current acquisition moment;
[0047] The predicted global latent vector value is compared with the actual global latent vector at the current acquisition time, and the impedance deviation is calculated according to each spatial location and each frequency point.
[0048] The impedance deviation at each spatial location and frequency point is input into a graph neural network classifier. Based on the diffusion pattern of the deviation in the spatial neighborhood, the distribution pattern in the frequency dimension, and the evolution trend in the time dimension, the degradation mechanism type at each spatial location is simultaneously determined.
[0049] In this scheme, the comprehensive safety margin assessment model employs a multi-task learning architecture to output the spatial safety margin, frequency safety margin, and temporal safety margin for each spatial location, specifically including:
[0050] The impedance deviation and degradation mechanism type at each spatial location, along with the data slices at the corresponding locations and times in the four-dimensional tensor data array, are used as input features for the comprehensive safety margin assessment model.
[0051] The comprehensive safety margin assessment model adopts a multi-task learning architecture, including a bottom-level shared feature extraction network and three parallel task branches;
[0052] The underlying shared feature extraction network uses each spatial measurement point as a graph node, establishes connections based on the spatial adjacency relationship between nodes and the similarity relationship of soil electrical properties, and propagates and aggregates the feature information of each node in the spatial neighborhood through graph convolution operation, outputting the shared feature vector of each spatial location.
[0053] The three parallel task branches respectively receive the shared feature vector and output spatial safety margin prediction value, frequency safety margin prediction value and time safety margin prediction value;
[0054] The predicted space safety margin is output by the first task branch, which takes the shared feature vector as input and outputs the space safety margin of the corresponding location in the three sub-items of contact potential difference, step potential difference and ground potential rise.
[0055] The frequency safety margin prediction value is output by the second task branch. The second task branch takes the shared feature vector as input, divides the frequency axis into three feature frequency bands with the soil ionization feature frequency and the inductance-dominant frequency as the dividing points, and predicts the frequency band aggregate safety margin of each frequency band as the frequency safety margin of the corresponding position.
[0056] The predicted time safety margin is output by the third task branch. The third task branch takes the shared feature vector and the degradation mechanism type as input. For gradual corrosion mechanism, it uses a trend extrapolation path to predict the remaining effective service life as the time safety margin. For sudden fracture, it uses a threshold determination path to output the current safety status as the time safety margin.
[0057] In this scheme, a comprehensive security determination is made based on the spatial security margin, the frequency security margin, and the time security margin, specifically including:
[0058] A three-level comprehensive safety assessment is performed based on the predicted values of spatial safety margin, frequency safety margin, and time safety margin. When all three values are greater than the first preset threshold, the safety is deemed acceptable.
[0059] If either value is between the first and second preset thresholds, a potential risk is identified and maintenance recommendations are output. If either value is less than the second preset threshold, a serious safety hazard is identified and an alarm signal and maintenance strategy are output.
[0060] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0061] This invention constructs a ternary coupled map of spatial location, soil electrical properties, and electrical parameters, along with a four-dimensional tensor data array of time-space-frequency-electrical properties. It cascades and matches soil stratified electrical and electrical parameters in a spatial coordinate system, enabling coordinate points to simultaneously associate information across four dimensions: soil electrical properties, electrical parameters, broadband impedance response, and temporal evolution trajectory. Based on this, a graph attention mechanism is employed to extract spatial-frequency coupling features, using spatial proximity, frequency response curve similarity, and soil electrical similarity as triple factors to determine attention weights, achieving selective aggregation. Neural differential equations are used to model the continuous dynamic system of non-uniform time-interval acquisition data, overcoming the dependence of traditional discrete-time models on equal-interval sampling. A multi-task learning architecture is employed to simultaneously output predicted values for three independent physical dimensions: spatial safety margin, frequency safety margin, and temporal safety margin. This invention overcomes the limitations of existing technologies that use a single power frequency grounding resistance value as a safety criterion, cannot distinguish defect types, and are difficult to adapt to non-uniform sampling data. It achieves multi-dimensional joint assessment of grounding grid safety characteristics and diagnostic identification of defect types.
[0062] This invention provides clear guidelines for the operation and maintenance of grounding grids, avoiding the shortcomings of existing technologies that only provide alarms for exceeding limits but cannot guide specific handling measures. It effectively improves the comprehensiveness and practicality of grounding grid safety assessment and reduces equipment damage and personal safety risks caused by grounding system accidents. Attached Figure Description
[0063] To more clearly illustrate the technical solutions in the embodiments or examples of the present invention, the accompanying drawings used in the embodiments or examples will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained according to these drawings without creative effort.
[0064] Figure 1 A flowchart of a comprehensive evaluation method for the safety characteristics of grounding grids based on multi-dimensional joint analysis is shown.
[0065] Figure 2 The flowchart for constructing a time-space-frequency-electrical four-dimensional tensor data array is shown;
[0066] Figure 3 This illustrates a flowchart of a health assessment model extracting coupled features in the spatial and frequency dimensions.
[0067] Figure 4 The flowchart shows the impedance deviation and degradation mechanism type of each spatial location output by the health diagnostic model. Detailed Implementation
[0068] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.
[0069] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.
[0070] like Figure 1 As shown, this embodiment provides a comprehensive evaluation method for the safety characteristics of grounding grids based on multi-dimensional joint analysis, including:
[0071] Obtain the electrical parameters and soil electrical parameters of each measuring point in the grounding grid, and construct a three-dimensional coupled distribution map of the spatial location, soil electrical parameters, and electrical parameters of the entire grounding grid.
[0072] Inject current signals into the grounding grid respectively, and obtain the broadband impedance characteristic spectrum containing steady-state and transient components;
[0073] The ternary coupling distribution spectrum and broadband impedance characteristic spectrum are repeatedly acquired at non-uniform time intervals to construct a time-space-frequency-electrical four-dimensional tensor data array;
[0074] The four-dimensional tensor data array is input into the health diagnosis model. The graph attention mechanism is used to extract the coupling features in the spatial and frequency dimensions. The time evolution trajectory of the coupling features is modeled using neural differential equations, and the impedance deviation and degradation mechanism type at each spatial location are output.
[0075] The impedance deviation and degradation mechanism type are input into the safety margin comprehensive evaluation model, and the spatial safety margin, frequency safety margin and time safety margin of each spatial location are output using a multi-task learning architecture.
[0076] A comprehensive security assessment is made based on the spatial security margin, the frequency security margin, and the time security margin.
[0077] In practice, a three-dimensional spatial measurement point matrix is deployed in the area where the grounding grid is located. Horizontally, the conductor intersections of the grounding grid are used as primary measurement points, the midpoints of the lines connecting adjacent nodes are used as supplementary measurement points, and the area extending outwards from the outermost conductor of the grounding grid at a predetermined distance is used as the measurement points for the edge expansion region. Vertically, the ground surface is used as the first layer of sampling depth, the grounding grid burial depth as the third layer of sampling depth, a predetermined multiple of the grounding grid burial depth as the second layer of sampling depth, and a predetermined multiple of the grounding grid burial depth as the fourth layer of sampling depth. The intersections of each horizontal measurement point location with each vertical layer of sampling depth constitute the specific location of each spatial measurement point.
[0078] Electrical parameters were measured at each spatial measuring point. The grounding impedance at each measuring point was measured using the potential drop method. Current electrodes were placed at a distance from the edge of the grounding grid that was a preset multiple of the maximum diagonal length of the grounding grid. Voltage electrodes were placed at multiple different distances along the line connecting the edge of the grounding grid and the current electrodes. The voltage variation curve with distance was recorded, and the voltage value corresponding to the flat section of the curve was obtained to calculate the grounding impedance. The contact potential difference and step potential difference at each measuring point were measured using the field simulated fault current injection method. A simulated short-circuit current was injected between any two nodes of the grounding grid, and the surface potential gradient at each measuring point was recorded simultaneously. The contact potential difference and step potential difference under the specified fault current level were obtained through proportional conversion.
[0079] Soil electrical stratification parameters were measured at various spatial measuring points. Four electrodes were arranged using the Wenner four-electrode method at the same measuring point, with the spacing between adjacent electrodes increasing sequentially. The apparent resistivity at each electrode spacing was measured, obtaining a test curve showing the apparent resistivity changing with the electrode spacing. Using the actual resistivity and layer thickness of each subsurface layer as the unknowns and the apparent resistivity test curve as the knowns, a nonlinear inversion objective function was constructed. The inversion objective function is the sum of squared errors between the measured apparent resistivity curve and the curve calculated by the inversion model. An optimization iterative algorithm was used to minimize the fitting error between the inversion curve and the test curve, outputting the actual resistivity and layer thickness of each soil layer from top to bottom at each measuring point, establishing a three-dimensional soil electrical stratification model for each measuring point. Electrical parameters (grounding impedance, contact potential difference, step potential difference) at the same spatial coordinates were spatially matched with soil electrical stratification parameters (actual resistivity and layer thickness of each layer), so that each spatial measuring point simultaneously associates both the electrical parameter vector and the soil electrical parameter vector.
[0080] After data matching at each discrete measuring point, radial basis function spatial interpolation is performed on the soil stratification interface depth at each measuring point. Using the stratification interface depth at each spatial measuring point as the reference quantity for the vertical coordinate, interpolation is performed on the stratification interface depth of each soil layer on the horizontal plane, resulting in a continuous stratification interface surface at all depths across the entire site. This ensures that each soil layer is continuously distributed within the spatial region enclosed by its stratification interface, thus extending the soil stratification inversion results from discrete measuring points into a continuous three-dimensional soil electrical spatial model for the entire site. Using the surface soil resistivity at each location in the three-dimensional soil electrical spatial model as a prior covariate, co-kriging interpolation is used to spatially estimate the contact potential difference and step potential difference at each measuring point. Measuring points with higher surface soil resistivity among neighboring measuring points at non-measuring point locations are assigned higher interpolation weights.
[0081] The three-dimensional soil electrical spatial model and the spatial interpolation results of electrical parameters are superimposed in the same three-dimensional spatial coordinate system, so that each coordinate point in space is simultaneously associated with the soil electrical parameters and electrical parameters at the corresponding depth, forming a ternary coupled distribution map of spatial coordinates, soil electrical parameters, and electrical parameters.
[0082] In practice, a signal injection device is connected between any two nodes of the grounding grid. This device is coupled to the grounding grid under test via an impedance isolation transformer to prevent the injected signal from interfering with the power supply system in the opposite direction. DC current signal, power frequency current signal, multi-frequency sinusoidal sweep signal, standard lightning current impulse waveform, and VFTO damped oscillation waveform are sequentially injected into the grounding grid.
[0083] For the responses to DC current signals, power frequency current signals, and multi-frequency sinusoidal sweep signals, after the response signals stabilize, the effective values of voltage and current and their phase difference at the injection point are synchronously acquired. The ratio of voltage to current is taken as the impedance amplitude at that frequency point, and the phase difference between voltage and current is taken as the impedance phase angle. The impedance amplitudes and phase angles of all sweep frequency points are arranged in logarithmic frequency order to form a steady-state broadband impedance spectrum.
[0084] For the response to the standard lightning current impulse waveform and the VFTO damped oscillation waveform, the time-domain voltage waveform and time-domain current waveform at the injection point are simultaneously acquired at a preset sampling rate during transient signal injection. The acquisition window covers the entire process from the start of the transient signal to the complete decay of the response to zero. Fourier transforms are performed on the acquired time-domain voltage and current waveforms to convert them into frequency-domain voltage and current spectra. The ratio of the voltage spectrum to the current spectrum at each frequency point is used as the transient impedance amplitude, and the phase difference between the voltage spectrum and the current spectrum at each frequency point is used as the transient impedance phase angle, thus constructing the transient impedance spectrum.
[0085] Align the steady-state broadband impedance spectrum and the transient impedance spectrum on the same logarithmic frequency coordinate system. For each frequency point within the overlapping frequency band, calculate the offset rate of the transient impedance amplitude relative to the steady-state impedance amplitude as the dynamic response factor. Traverse all swept frequency points, and take the frequency point where the dynamic response factor first shows a step change as the soil ionization characteristic frequency. Calculate the rate of change of the steady-state impedance phase angle with frequency at each frequency point, and take the inflection point where the steady-state impedance phase angle changes from positive to negative as the inductance-dominant frequency.
[0086] The steady-state impedance amplitude, steady-state impedance phase angle, transient impedance amplitude, transient impedance phase angle, dynamic response factor, and the relative position of the frequency point with the soil ionization characteristic frequency and the inductance-dominant frequency together constitute the data entries of the broadband impedance characteristic spectrum, which are arranged in frequency order to form the broadband impedance characteristic spectrum.
[0087] In practice, the soil corrosivity level is comprehensively assessed based on the soil resistivity, pH value, redox potential, and stray current density at the location of the grounding grid. Measured values of soil resistivity, pH value, redox potential, and stray current density are obtained at each measuring point. The measured value of each indicator is compared with its corresponding grading threshold, and a grading score is assigned to each indicator. The weighted sum of the grading scores of the four indicators yields a comprehensive corrosion score. The soil corrosivity level of the area is determined based on the range of values within which the comprehensive corrosion score falls. Soil corrosivity levels are divided into high corrosion, medium corrosion, and low corrosion levels. High corrosion level corresponds to the shortest sampling time interval, low corrosion level corresponds to the longest sampling time interval, and medium corrosion level corresponds to a sampling time interval in between. When the area where the grounding grid is located spans multiple soil corrosion levels, the highest corrosion level is used to determine a uniform sampling time interval for the entire area. When extreme weather events occur in the area where the grounding grid is located, or when the grounding grid experiences a near-zone short-circuit fault, a non-periodic temporary sampling is immediately triggered.
[0088] In another preferred embodiment, the time series of soil electrical parameters at each measuring point in the ternary coupled distribution map collected in each previous collection are extracted. The soil electrical parameters include soil resistivity, pH value, redox potential, and stray current density. The first-order difference value of each soil electrical parameter is calculated according to the time series. When the redox potential shows a unidirectional monotonically decreasing trend in multiple consecutive collection cycles and the cumulative decrease exceeds the first warning threshold, it is determined that there is a trend of deterioration of the corrosion environment at the current location, an advanced warning signal is issued, and the collection time interval at the current location is shortened. When the stray current density shows a unidirectional monotonically increasing trend in multiple consecutive collection cycles and the cumulative increase exceeds the second warning threshold, it is determined that there is a risk of accelerated electrochemical corrosion at the current location, an advanced warning signal is issued, and a maintenance suggestion to add a drainage device is output.
[0089] like Figure 2 As shown, each acquisition operation records coordinate information in three dimensions: spatial coordinates, which record the three-dimensional spatial position coordinates of each measurement point in this acquisition, and the spatial coordinates remain consistent between each acquisition; frequency coordinates, which record the frequency values of each frequency point in the broadband impedance characteristic spectrum of this acquisition, and the frequency coordinates remain consistent between each acquisition; and time coordinates, which record the absolute timestamp of this acquisition and the relative time offset relative to the first acquisition.
[0090] The ternary coupling distribution map and broadband impedance characteristic spectrum obtained from each acquisition are used as three-dimensional data volumes. Using horizontal spatial coordinates and frequency coordinates as index dimensions, electrical parameters and soil electrical parameters at corresponding locations and frequencies are stored. The three-dimensional data volumes from each acquisition are stacked chronologically, with time as the fourth-dimensional index, ensuring each acquisition occupies an independent slice position in the time dimension, forming a time-space-frequency-electrical four-dimensional tensor data array. Temporal and spatial consistency checks are performed on the electrical parameters at each spatial location within each time slice in the four-dimensional tensor data array. For any spatial location, the current electrical parameter value at a certain frequency point in the current time slice is obtained, along with historical electrical parameter values from multiple consecutive acquisitions at the same frequency point. The average change of these historical electrical parameter values is calculated. The current value is compared with the value from the previous acquisition at the same frequency point to obtain the current change. When the current change exceeds a preset multiple of the average change, the data point is preliminarily identified as a suspected outlier. Suspected outliers are removed and replaced with the linear interpolation result of two adjacent acquisitions at the location. In cases where all data for a specific area is missing due to measurement equipment malfunction or signal interference, compensation is performed using data from the previous data collection period for that area, combined with the expected degradation rate corresponding to the soil corrosivity level. After outlier detection and preprocessing, a four-dimensional tensor data array suitable for input to the health diagnostic model is obtained.
[0091] like Figure 3 As shown, a four-dimensional tensor data array is input into the health diagnosis model, and a graph attention mechanism is used to extract coupling features in the spatial and frequency dimensions. Each spatial measurement point in the grounding grid is treated as a node in the graph network. When the spatial Euclidean distance between two measurement points is less than a preset distance threshold, an edge is established between the two nodes; when the spatial distance is greater than or equal to the preset distance threshold, no edge is established. The distance threshold is determined based on the maximum diagonal length of the grounding grid, ensuring that each node's neighborhood contains a reasonable number of adjacent nodes. All nodes and edges together constitute the topology of the grounding grid.
[0092] An initial feature vector is constructed for each node. This initial feature vector is composed of the soil resistivity and thickness of each layer at the node's location, the steady-state impedance amplitude and phase angle at each frequency point, the transient impedance amplitude and phase angle at each frequency point, the dynamic response factor at each frequency point, and the measured values of contact potential difference and step potential difference, arranged in a predetermined dimensional order. In broadband impedance feature spectrum extraction, the dynamic response factor characterizes the degree of shift of transient impedance relative to steady-state impedance, expressed as: , For frequency The transient impedance amplitude under the following conditions For frequency steady-state impedance amplitude, This indicates that the transient current dissipation capability is lower than that in the steady state, and the grounding electrode exhibits a significant inductive effect; This indicates that the transient current dissipation capability is higher than that in the steady state, suggesting that soil ionization occurs near the frequency point, reducing the impedance. The initial feature vectors of each node are projected onto the latent space through a shared linear transformation matrix, mapping the original feature dimensions to a high-dimensional latent space, thus obtaining the latent layer features of each node.
[0093] For each pair of adjacent nodes, the spatial proximity factor, frequency response curve similarity factor, and soil electrical similarity factor are calculated. The spatial proximity factor is determined by the Euclidean distance between the spatial measurement points corresponding to the two nodes; the closer the distance, the higher the spatial proximity factor. The frequency response curve similarity factor is determined by the similarity in shape of the frequency response curves corresponding to the two nodes. The frequency response curves are formed by arranging the steady-state impedance amplitudes of each node in frequency order. The shape similarity of the two curves is compared using a dynamic time warping algorithm; the smaller the sum of the distances, the more similar the curve shapes. The soil electrical similarity factor is determined by the similarity in the electrical parameters of the soil layers at the locations corresponding to the two nodes. The resistivity of each soil layer is arranged into a vector in depth order, and the cosine similarity between the two vectors is calculated. (Original attention weights) The weighted fusion of spatial proximity factor, frequency response curve similarity factor, and soil electrical property similarity factor is expressed as: ,in Spatial proximity factor This is the frequency response curve similarity factor. Soil electrical similarity factor, , , These are the preset weight coefficients. The softmax function is used to normalize the original attention weights of all adjacent nodes, resulting in normalized attention weight coefficients between adjacent nodes.
[0094] The hidden features of all neighboring nodes are weighted and summed using normalized attention weight coefficients to obtain the aggregated neighborhood feature vector of the node. This aggregated neighborhood feature vector is then concatenated with the node's own hidden features to obtain the node's enhanced feature vector. In this enhanced feature vector, spatially adjacent neighboring nodes with similar frequency response curve shapes and consistent soil electrical profiles contribute significantly more to the central node's features than other neighboring nodes. Spatially adjacent neighboring nodes with significantly different frequency response curve shapes or soil electrical properties are suppressed in the attention weight allocation, thus preserving the node's differentiated features. The enhanced features output from the previous layer are used as input to the next layer, and the linear transformation, attention weight calculation, and node feature aggregation and update operations are repeatedly performed until a preset number of layers is reached. The enhanced features of each node output from the last layer are rearranged according to the spatial measurement point location and frequency point corresponding to each node to obtain the spatial-frequency coupled feature map under the time slice.
[0095] like Figure 4 As shown, neural differential equations are used to model the temporal evolution trajectory of coupled features. Output the impedance deviation and degradation mechanism type at each spatial location, among which For a moment The system state vector, It is a differential neural network. These are the trainable parameters for the network. The spatial-frequency coupled feature maps of each time slice are arranged chronologically. The feature map of each time slice is flattened into a one-dimensional vector. This one-dimensional vector contains the coupled feature information of all spatial locations and all frequency points at the current time, and is denoted as the global latent vector of the current time slice. The global latent vectors of all time slices are arranged chronologically to form a latent vector sequence, with non-uniform time intervals between adjacent latent vectors.
[0096] The degradation process of the grounding grid's health status is considered as a continuous-time dynamic system governed by physical laws. The state at any given time is represented by a state vector, the dimension of which is the same as the dimension of the global latent vector. The evolution of this continuous-time dynamic system is modeled using a differential neural network. The input is the current state vector and the current timestamp, and the output is the derivative of the state vector with respect to time. The parameters of the differential neural network are determined by learning the state evolution patterns from historical data collected by the grounding grid during training.
[0097] The global latent vector at the initial acquisition time is used as the initial value of the state vector. For the latest acquisition time, starting from the initial time, the neural differential equation is used as the initial value problem, and a numerical integration method is used to progressively solve it. At each step, the state value of the next time step is calculated based on the derivative at the current time step and the step size. This process is continued until the current acquisition time, obtaining the continuous evolution trajectory of the state vector within the time interval and the target time. State vector prediction value , represented as , The moment of first data collection The global latent vector, This is the time integration variable. The predicted state vector at the current moment is compared with the global latent vector obtained from the actual acquired data at the current moment. For each spatial location and each frequency point, the impedance deviation at the current location and frequency point is obtained by dividing the absolute value of the difference between the predicted and actual impedance values by the impedance value in the initial state during the first acquisition. , represented as: ,in Predicted by neural differential equations The impedance value at that location and frequency at that time. For actual collection The impedance value at that location and frequency at that time. This is the initial impedance value at this location and frequency point when it is first acquired.
[0098] The calculated impedance deviations at each spatial location and frequency point are input into a graph neural network classifier. The classifier propagates and aggregates the deviations in the spatial dimension, extracting the diffusion pattern features within the spatial neighborhood; it sorts and analyzes the deviations at each frequency point in the frequency dimension, extracting the distribution pattern features of deviations changing with frequency; and it performs trend analysis on the historical sequence of deviations obtained from previous acquisitions in the time dimension, extracting the evolution trend features of deviations over time. The classifier internally consists of three parallel criterion channels: a spatial diffusion pattern criterion, which distinguishes between local conductor breaks and electrochemical corrosion thinning based on whether the deviation is a local or regional anomaly in the spatial neighborhood; a frequency distribution pattern criterion, which further verifies the defect type based on whether the deviation is a local distortion or a full-band drift in the frequency domain; and a temporal evolution trend criterion, which distinguishes between conductor mechanical fracture and electrochemical corrosion thinning based on whether the deviation changes abruptly or gradually over time. After the three criteria channels output preliminary discrimination results, they are integrated by the top-level fusion layer of the classifier. When the three channels point to the same degradation mechanism type, that type is output directly. When they point to different mechanism types, the output result of the time evolution trend criterion is given priority. Finally, the impedance deviation value, degradation mechanism type (normal, conductor mechanical fracture, electrochemical corrosion thinning, or connection loosening) and confidence score of each type are output for each spatial location.
[0099] When identifying the degradation mechanism type, the mechanical fracture of the conductor is further subdivided into three subtypes: complete fracture, partial fracture, and loose connection. A complete fracture subtype is defined as follows: when the impedance deviation at a certain location tends to the preset maximum deviation threshold at all frequency points, and the impedance deviation at the current location exhibits a concentric circle-like outward decay distribution centered on the current location; a partial fracture subtype is defined as follows: when the impedance deviation at a certain location shows an isolated peak at the inductor-dominant frequency while the deviation in other frequency bands is relatively small, and there are two or more paths connecting the current location to the current location within the spatial neighborhood; the physical state corresponding to the partial fracture subtype is that some strands of the conductor are broken, but the remaining strands remain conductive; and a loose connection subtype is defined as follows: when the impedance deviation at a certain location decreases with the increase of the injected current amplitude, and the current location exhibits irregular random fluctuations in the time dimension; the physical state corresponding to the loose connection subtype is that the conductor connection point has poor contact.
[0100] In practical implementation, impedance deviation and degradation mechanism type are input into the comprehensive safety margin assessment model. A multi-task learning architecture is used to output the spatial safety margin, frequency safety margin, and time safety margin for each spatial location. The impedance deviation and degradation mechanism type of each spatial location output by the health diagnosis model, along with data slices of the corresponding location and time in the four-dimensional tensor data array, are used as input features of the comprehensive safety margin assessment model. For each spatial location, the input feature vector includes soil electrical parameters, current values of electrical parameters, broadband impedance characteristic parameters, impedance deviation values at each frequency point, degradation mechanism type codes, and confidence scores. The above information is concatenated into a fixed-length one-dimensional feature vector according to a preset dimensional order.
[0101] The comprehensive safety margin assessment model employs a multi-task learning architecture, including a bottom-level shared feature extraction network and three parallel task branches. The bottom-level shared feature extraction network uses a graph neural network architecture, treating each spatial measurement point as a graph node, with node features being the concatenated input feature vectors at each location. The rule for establishing edge relationships is as follows: when the spatial Euclidean distance between two measurement points is less than a preset spatial neighborhood radius threshold, and the cosine similarity between the soil resistivity vectors of each layer of the two points is greater than a preset similarity threshold, an edge is established between the two points; if any condition is not met, no edge is established. The bottom-level shared network propagates and aggregates the feature information of each node within the spatial neighborhood through multi-layer graph convolution operations, outputting a high-dimensional shared feature vector for each spatial location.
[0102] Three parallel task branches are set up above the underlying shared feature extraction network, with no parameters shared between the branches. The first task branch takes the shared feature vector as input, passes through an independent fully connected mapping layer, and outputs the predicted spatial safety margin value for the location in three sub-items: contact potential difference, step potential difference, and ground potential rise. The loss function of the first task branch is expressed as:
[0103]
[0104] in This is a predicted value for the contact voltage safety margin. This serves as the benchmark value for the measured calculation of the contact voltage safety margin. This is a predicted value for the step voltage safety margin. This serves as the measured benchmark value for calculating the step voltage safety margin. This is the predicted value for the safety margin of ground potential rise. This serves as the measured baseline value for calculating the safety margin of ground potential rise. The weighting factor for the contact voltage safety margin loss term. The weighting factor for the step voltage safety margin loss term. This is the weighting coefficient for the ground potential rise safety margin loss term.
[0105] The second task branch takes the shared feature vector as input, passes through an independent fully connected mapping layer, and divides the frequency axis into three feature frequency bands with the soil ionization characteristic frequency and the inductance-dominant frequency as the dividing points. It first predicts the impedance safety margin at each frequency point, and then weights and aggregates the safety margins of each frequency point falling within the three frequency bands to obtain the predicted frequency safety margin value for the target location. The loss function of the second task branch is expressed as:
[0106]
[0107] in This is the index number for the characteristic frequency band. These correspond to the low-frequency band (below the characteristic frequency of soil ionization), the mid-frequency band (between the two characteristic frequencies), and the high-frequency band (above the dominant frequency of inductance), respectively. For the first The predicted value of the frequency band aggregation security margin for each characteristic frequency band. For the first The measured benchmark value for the frequency band aggregation safety margin of each characteristic frequency band. For the first The weighting coefficients of the loss term for each characteristic frequency band.
[0108] The third task branch takes the shared feature vector and degradation mechanism type as input, and internally sets up two sub-paths: For gradual corrosion mechanisms, a trend extrapolation path is used, predicting the cumulative degradation within a preset time window using the historical impedance deviation sequence as input and converting it into a time safety margin characterized by remaining effective service life; for sudden fracture or loosening mechanisms, a threshold determination path is used, directly determining the current safety status and outputting the time safety margin based on whether the impedance deviation at the current moment exceeds a preset threshold. The loss function of the third task branch is expressed as: ,in This is a predicted value for the time safety margin. This serves as the benchmark value for the actual calculation of the time safety margin.
[0109] The three task branches employ a joint training strategy, with the total loss function being the weighted sum of the loss functions of the three branches. Backpropagation simultaneously optimizes the parameters of the underlying shared network and the three branch networks. The spatial safety margin prediction values, frequency safety margin prediction values, and time safety margin prediction values output by the three task branches are compared with preset first and second thresholds, where the second threshold is less than the first threshold. When all three are greater than the first threshold, the current location is deemed to be in a safe state, a normal status indicator is output, and no warning is triggered. When any one of the three is between the first and second thresholds, the current location is deemed to have a potential risk. Based on the task branch from which the current indicator originates, targeted maintenance recommendations are output: if the spatial safety margin is low, it is recommended to add a local equipotential bonding strip or densify the grounding conductor; if the frequency safety margin is low, it is recommended to add grounding electrodes or adjust the grounding arrangement in the sensitive area of the corresponding frequency band; if the time safety margin is low, it is recommended to develop a maintenance plan based on the degradation mechanism type, with corrosion-induced thinning types included in periodic maintenance and fracture-induced loosening types included in near-term maintenance. When any one of the three is less than the second threshold, the current location is deemed to have a serious safety hazard. Based on the task branch to which the current indicator belongs and the degradation mechanism type output by the health diagnosis model, differentiated alarm signals and corresponding maintenance strategies are output: when the space safety margin is severely insufficient, an emergency alarm is triggered and an immediate power outage is recommended; when the frequency safety margin is severely insufficient, an important alarm is triggered and it is recommended to handle it within the next power outage window; when the time safety margin is severely insufficient, a general alarm is triggered, and a specific maintenance plan is output according to the degradation mechanism type: for corrosion thinning type, the conductor should be replaced; for breakage and loosening type, the conductor should be reconnected.
[0110] After comprehensively assessing safety and outputting alarm signals and maintenance strategies, the process also includes defect spatial location and maintenance plan generation steps: obtaining the degradation mechanism type and impedance deviation value of each spatial location, and determining the location coordinates of the defect in the three-dimensional spatial coordinate system; determining the soil layer depth where the defect is located based on the soil electrical stratification parameters at the location in the three-dimensional coupling distribution map of the location coordinates; determining the boundary contour of the defect's influence range with the defect location coordinates as the center and the boundary of the continuous area where the impedance deviation exceeds the preset threshold; and outputting a maintenance plan that includes the excavation center coordinates, excavation depth, excavation radius, and conductor replacement length by combining the defect type, soil layer depth, and influence range contour.
[0111] The above description is only a specific embodiment of the present invention. The preset threshold, preset multiple, preset distance, preset number of layers, preset time window and other parameters in the above embodiments can be set by those skilled in the art according to the actual engineering requirements and the specific scale of the grounding grid. However, the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the protection scope of the present invention.
Claims
1. A comprehensive evaluation method for the safety characteristics of grounding grids based on multi-dimensional joint analysis, characterized in that, Includes the following steps: Obtain the electrical parameters and soil electrical parameters of each measuring point in the grounding grid, and construct a three-dimensional coupled distribution map of the spatial location, soil electrical parameters, and electrical parameters of the entire grounding grid. Inject current signals into the grounding grid respectively, and obtain the broadband impedance characteristic spectrum containing steady-state and transient components; The ternary coupling distribution spectrum and broadband impedance characteristic spectrum are repeatedly acquired at non-uniform time intervals to construct a time-space-frequency-electrical four-dimensional tensor data array; The four-dimensional tensor data array is input into the health diagnosis model. The graph attention mechanism is used to extract the coupling features in the spatial and frequency dimensions. The time evolution trajectory of the coupling features is modeled using neural differential equations, and the impedance deviation and degradation mechanism type at each spatial location are output. The impedance deviation and degradation mechanism type are input into the safety margin comprehensive evaluation model, and the spatial safety margin, frequency safety margin and time safety margin of each spatial location are output using a multi-task learning architecture. A comprehensive security assessment is made based on the spatial security margin, the frequency security margin, and the time security margin.
2. The comprehensive evaluation method for grounding grid safety characteristics based on multi-dimensional joint analysis according to claim 1, characterized in that, Constructing a three-dimensional coupled distribution map of the spatial location, soil electrical properties, and electrical parameters of the entire grounding grid, specifically including: A three-dimensional spatial measurement point matrix is laid out over the entire grounding grid. In the horizontal direction, the measurement point locations are the conductor intersection nodes, the midpoints of adjacent nodes, and the outer edge expansion areas. In the vertical direction, the layered acquisition depths are the ground surface, the burial depth of the grounding grid at preset multiples above and below the burial depth, and the burial depth itself. Grounding impedance, contact potential difference, and step potential difference were measured at each spatial measurement point. At the same measurement point, the apparent resistivity under each electrode spacing was measured by gradually increasing the electrode spacing using the Wenner quadrupole method. The apparent resistivity test curve was fitted to the true resistivity and layer thickness of each soil layer through nonlinear stratification inversion, and a three-dimensional soil electrical stratification model was established at each measurement point. Match electrical parameters at the same spatial coordinates with soil electrical stratification parameters, so that each spatial measuring point is simultaneously associated with both electrical parameter vectors and soil electrical parameter vectors; First, radial basis function spatial interpolation is performed on the soil layer interface depth at each measuring point to construct a continuous three-dimensional soil electrical spatial model for the entire site. The surface soil resistivity in the model is used as a prior covariate, and the electrical parameters of each measuring point are spatially estimated using the co-kriging interpolation method. The three-dimensional soil electrical spatial model and the spatial interpolation results of electrical parameters are superimposed in the same coordinate system, so that each coordinate point in space is simultaneously associated with the soil electrical parameters and electrical parameters at its location, generating a ternary coupled distribution map.
3. The comprehensive evaluation method for grounding grid safety characteristics based on multi-dimensional joint analysis according to claim 1, characterized in that, Obtaining a broadband impedance characteristic spectrum containing both steady-state and transient components, specifically including: A signal injection device is connected between any two nodes of the grounding grid to sequentially inject DC current signal, power frequency current signal, multi-frequency sinusoidal sweep signal, standard lightning current impulse waveform and VFTO damped oscillation waveform into the grounding grid. The steady-state injection point voltage and current at each frequency point are collected, the steady-state impedance amplitude and steady-state impedance phase angle at each frequency point are calculated, and the steady-state broadband impedance spectrum is formed by arranging them in frequency order. Fourier transforms are performed on the time-domain voltage response and time-domain current response of the lightning current impulse waveform and the VFTO damped oscillation waveform. The ratio of the voltage spectrum to the current spectrum at each frequency point is used as the transient impedance amplitude, and the phase difference is used as the transient impedance phase angle to construct the transient impedance spectrum. Align the steady-state broadband impedance spectrum with the transient impedance spectrum on the same frequency coordinate, calculate the offset rate of the transient impedance amplitude relative to the steady-state impedance amplitude at each frequency point in the overlapping frequency band as the dynamic response factor, take the frequency point where the dynamic response factor undergoes a step jump as the soil ionization characteristic frequency, and take the turning point where the steady-state impedance phase angle changes from positive to negative as the inductance-dominant frequency. The steady-state impedance amplitude, steady-state impedance phase angle, transient impedance amplitude, transient impedance phase angle, dynamic response factor, and the relative position markers of the frequency points with respect to the soil ionization characteristic frequency and the inductance-dominant frequency together constitute the data entries of the broadband impedance characteristic spectrum, which are arranged in frequency order to form the broadband impedance characteristic spectrum.
4. The comprehensive evaluation method for grounding grid safety characteristics based on multi-dimensional joint analysis according to claim 1, characterized in that, Constructing a four-dimensional tensor data array of time, space, frequency, and electrical properties, specifically including: The soil corrosivity level is comprehensively assessed based on the soil resistivity, pH value, redox potential and stray current density at the location of the grounding grid, and the time interval between two adjacent data collections is determined by the soil corrosivity level. During each acquisition, spatial coordinates, frequency coordinates, and time coordinates are recorded to ensure that the ternary coupling distribution spectrum and broadband impedance characteristic spectrum acquired in each acquisition maintain the same index reference in both spatial and frequency dimensions. Each collected data is used as a three-dimensional data volume. The three-dimensional data volume uses horizontal spatial coordinates and frequency coordinates as index dimensions to store electrical parameters and soil electrical parameters at the corresponding location and frequency point. The three-dimensional data volumes collected from each time are stacked in chronological order, with time as the fourth dimension index, to form a four-dimensional tensor data array. The time and space consistency of the data points in the four-dimensional tensor data array are checked, and the missing data is compensated and filled by superimposing the last collected data at the current position with the expected degradation rate, so as to obtain the preprocessed time-space-frequency-electrical four-dimensional tensor data array.
5. The comprehensive evaluation method for grounding grid safety characteristics based on multi-dimensional joint analysis according to claim 1, characterized in that, In the health assessment model, a graph attention mechanism is used to extract coupled features in the spatial and frequency dimensions, specifically including: Each spatial measuring point in the grounding grid is used as a node in the graph network. The edges between the nodes are established according to the spatial distance between the measuring points to construct the topology of the grounding grid. An initial feature vector is constructed for each node. The initial feature vector is composed of soil stratification electrical parameters at the corresponding location of the node, steady-state impedance amplitude and phase angle at each frequency point, transient impedance amplitude and phase angle, dynamic response factor, contact potential difference and step potential difference. The initial feature vectors of each node are mapped to the latent space through a shared linear transformation matrix to obtain the latent features of each node. For each pair of adjacent nodes, the dynamic time warping algorithm is used to compare the frequency response curves formed by arranging the steady-state impedance amplitudes of each node in frequency order to obtain the frequency response curve similarity factor. The vectors formed by arranging the soil resistivity of each layer of each node in depth order are compared to obtain the soil electrical similarity factor. The spatial proximity factor, frequency response curve similarity factor and soil electrical similarity factor are weighted and fused and then normalized to obtain the attention weight between adjacent nodes. The attention weights are used to weight and aggregate the hidden features of each neighboring node, and then concatenated with the hidden features of the node itself to obtain the enhanced features of each node. The enhanced features output from the previous layer are used as the input to the next layer. The linear transformation, attention weight calculation, and weighted aggregation operations are repeatedly performed until the preset number of layers is reached. The enhanced features of each node output from the last layer are rearranged according to the spatial measurement point position and frequency point corresponding to each node to obtain the spatial-frequency coupled feature map under the time slice.
6. The comprehensive evaluation method for grounding grid safety characteristics based on multi-dimensional joint analysis according to claim 1, characterized in that, In the health assessment model, neural differential equations are used to model the temporal evolution trajectory of the coupling features, outputting the impedance deviation and degradation mechanism type at each spatial location, specifically including: Arrange the spatial-frequency coupled feature maps of each time slice in chronological order, and flatten the feature map of each time slice into a one-dimensional vector as the global latent vector. Using the global latent vector at the first acquisition moment as the initial state, a neural differential equation with time as the independent variable is constructed. The derivative of the neural differential equation is output by the differential neural network and is used to describe the continuous evolution of the grounding grid health status in the time dimension. Starting from the initial moment, the neural differential equation is numerically integrated to obtain the global latent vector prediction value at the current acquisition moment; The predicted global latent vector value is compared with the actual global latent vector at the current acquisition time, and the impedance deviation is calculated according to each spatial location and each frequency point. The impedance deviation at each spatial location and frequency point is input into a graph neural network classifier. Based on the diffusion pattern of the deviation in the spatial neighborhood, the distribution pattern in the frequency dimension, and the evolution trend in the time dimension, the degradation mechanism type at each spatial location is simultaneously determined.
7. The comprehensive evaluation method for grounding grid safety characteristics based on multi-dimensional joint analysis according to claim 1, characterized in that, In the comprehensive safety margin assessment model, a multi-task learning architecture is used to output the spatial safety margin, frequency safety margin, and temporal safety margin for each spatial location, specifically including: The impedance deviation and degradation mechanism type at each spatial location, along with the data slices at the corresponding locations and times in the four-dimensional tensor data array, are used as input features for the comprehensive safety margin assessment model. The comprehensive safety margin assessment model adopts a multi-task learning architecture, including a bottom-level shared feature extraction network and three parallel task branches; The underlying shared feature extraction network uses each spatial measurement point as a graph node, establishes connections based on the spatial adjacency relationship between nodes and the similarity relationship of soil electrical properties, and propagates and aggregates the feature information of each node in the spatial neighborhood through graph convolution operation, outputting the shared feature vector of each spatial location. The three parallel task branches respectively receive the shared feature vector and output spatial safety margin prediction value, frequency safety margin prediction value and time safety margin prediction value; The predicted space safety margin is output by the first task branch, which takes the shared feature vector as input and outputs the space safety margin of the corresponding location in the three sub-items of contact potential difference, step potential difference and ground potential rise. The frequency safety margin prediction value is output by the second task branch. The second task branch takes the shared feature vector as input, divides the frequency axis into three feature frequency bands with the soil ionization feature frequency and the inductance-dominant frequency as the dividing points, and predicts the frequency band aggregate safety margin of each frequency band as the frequency safety margin of the corresponding position. The predicted time safety margin is output by the third task branch. The third task branch takes the shared feature vector and the degradation mechanism type as input. For gradual corrosion mechanism, it uses a trend extrapolation path to predict the remaining effective service life as the time safety margin. For sudden fracture, it uses a threshold determination path to output the current safety status as the time safety margin.
8. The comprehensive evaluation method for grounding grid safety characteristics based on multi-dimensional joint analysis according to claim 1, characterized in that, A comprehensive security assessment is performed based on the aforementioned spatial security margin, frequency security margin, and time security margin, specifically including: A three-level comprehensive safety assessment is performed based on the predicted values of spatial safety margin, frequency safety margin, and time safety margin. When all three values are greater than the first preset threshold, the safety is deemed acceptable. If either value is between the first and second preset thresholds, a potential risk is identified and maintenance recommendations are output. If either value is less than the second preset threshold, a serious safety hazard is identified and an alarm signal and maintenance strategy are output.
