Photovoltaic buried cable fault positioning system based on graph neural network

Abstract

The invention discloses a photovoltaic buried cable fault positioning system based on a graph neural network, and the system comprises a multi-source signal collection module which is used for collecting electromagnetic response characteristics, and generating a multi-source signal sample set; the topological relation construction module is used for forming weighted directed topological relation data; the frequency domain mapping and fusion module is used for forming frequency domain-space joint feature input; the frequency domain self-evolution graph neural model module is used for executing feature propagation and aggregation of nodes and edges and outputting fault probability and uncertainty results; the space-time active detection module is used for selecting a cable section in an area with the maximum potential energy as an active detection target to form a self-tuning active detection closed loop; and the multi-source constraint correction module is used for outputting the type, the position and the confidence degree result of the buried cable fault. According to the invention, through the frequency domain self-evolution graph nervous system, accurate identification and positioning of the photovoltaic buried cable fault are realized.

Description

Technical Field

[0001] This invention relates to the field of intelligent detection and fault diagnosis technology for power systems, and in particular to a fault location system for photovoltaic underground cables based on graph neural networks. Background Technology

[0002] In photovoltaic (PV) power generation systems, underground cables serve as a crucial carrier for power transmission, and their operational safety directly impacts the overall power generation efficiency and stability of the PV power plant. With the expansion of PV power plant scale and the increase in cable laying length, the operating environment of underground cables becomes complex, frequently affected by factors such as soil moisture, temperature changes, mechanical stress, and aging, leading to diverse and often concealed fault types. Current PV cable inspection methods primarily rely on manual inspection, low-frequency excitation testing, and fixed-point ranging. These methods are inefficient when dealing with large-scale underground lines, are easily affected by external interference, and cannot achieve multi-point synchronous detection and intelligent positioning.

[0003] While current multi-frequency signal detection technologies can acquire certain cable response characteristics, electromagnetic signal attenuation and distortion are severe under complex soil conditions, and the response varies significantly across different frequencies. Traditional algorithms often employ fixed-frequency scanning or single-feature analysis, leading to decreased detection accuracy. Existing methods typically rely solely on single-source measurement data for fault diagnosis, lacking comprehensive analysis of current distribution, grounding resistance, and direction information, resulting in proneness to bias and misjudgment in location results. Furthermore, existing intelligent algorithms are mostly shallow network models, failing to fully represent the topological relationships between cable nodes and struggling to adapt to the multidimensional nonlinear characteristics of buried environments.

[0004] To address the aforementioned issues, traditional detection technologies suffer from limitations such as limited signal feature extraction, reliance on manual frequency selection, insufficient data fusion capabilities, and static model structures, making it difficult to meet the high-precision fault location requirements in complex scenarios involving buried photovoltaic cables.

[0005] Therefore, how to provide a fault location system for photovoltaic underground cables based on graph neural networks is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a fault location system for buried photovoltaic cables based on graph neural networks. This invention fully utilizes multi-frequency electromagnetic signal detection technology, frequency domain feature mapping mechanisms, and graph neural network intelligent reasoning models. Through multi-source signal acquisition, topological relationship construction, frequency domain mapping fusion, graph structure reasoning, spatiotemporal active detection, and multi-source constraint correction, it achieves intelligent identification and precise location of faults in buried photovoltaic cables. By introducing a frequency domain self-evolving graph neural network model, this invention performs self-learning reconstruction of multi-frequency signal features, dynamically adjusting the excitation frequency and fusing multi-source information such as grounding resistance, direction, and current distribution. This enables high-precision judgment of the location and type of cable faults in complex soil environments. This invention has the advantages of high detection accuracy, strong anti-interference capability, good adaptability, and high automation, significantly improving the fault location efficiency and intelligent operation and maintenance level of buried photovoltaic cables.

[0007] A fault location system for underground photovoltaic cables based on graph neural networks according to an embodiment of the present invention includes the following modules: The multi-source signal acquisition module is used to apply multi-frequency excitation signals to the buried cable using a multi-frequency output transmitter, and to collect electromagnetic response characteristics through a differential coil receiver to generate a multi-source signal sample set. The topology construction module is used to establish the topology structure of buried cables based on multi-source signal sample sets and cable layout diagrams, forming weighted directed topology data. The frequency domain mapping and fusion module is used to perform frequency mapping, frequency band selection and feature reconstruction on multi-frequency signal features based on weighted directed topological relationship data, forming a joint frequency domain-space feature input. The frequency domain self-evolutionary graph neural model module is used to receive frequency domain-space joint feature input, perform feature propagation and aggregation of nodes and edges, and output the fault probability and uncertainty results of nodes and cable segments. The spatiotemporal active detection module is used to calculate the spatiotemporal evolution potential energy distribution based on the uncertainty results, select the cable segment with the maximum potential energy as the active detection target, and form a self-tuning active detection closed loop. The multi-source constraint correction module is used to input the grounding resistance, grounding direction and node current distribution as multi-source constraints into the frequency domain self-evolution graph neural model, and output the type, location and confidence results of the buried cable fault.

[0008] Optionally, modules can be integrated using the following methods: In photovoltaic power plants, a multi-frequency output transmitter is used to apply multi-frequency excitation signals to buried cables. The electromagnetic response characteristics of each frequency point are collected by a differential coil receiver, and multi-frequency signal characteristics are extracted and formed. At the same time, the grounding resistance, grounding direction and node current distribution are measured to generate a multi-source signal sample set. Based on the multi-source signal sample set and the cable layout diagram of the photovoltaic power station, the topological relationship structure of the underground cable is established. The junction box, junction box, branch point and grounding point are defined as nodes, and the cable segments between adjacent nodes are defined as edges, forming the weighted directed topological relationship data of the underground cable. Based on weighted directed topological relation data, a frequency domain feature mapping and fusion unit is constructed to perform frequency mapping, frequency band selection and feature reconstruction on multi-frequency signal features, generate virtual frequency point features, and fuse the virtual frequency domain features with the weighted directed topological relation data to form a frequency domain-space joint feature input; The frequency domain-space joint features are input into the frequency domain self-evolutionary graph neural model. The feature propagation and aggregation of nodes and edges are performed through the physical guided message passing mechanism, and the fault probability, uncertainty matrix and frequency response estimation results of nodes and cable segments are output. Based on the uncertainty matrix, the spatiotemporal evolution potential energy distribution is calculated, and the cable segment corresponding to the region with the maximum potential energy is selected as the active detection target. The excitation frequency of the multi-frequency output transmitter and the position of the measurement point are adjusted to form a spatiotemporally coupled self-tuning active detection closed loop. The grounding resistance, grounding direction, and node current distribution are used as multi-source constraint inputs to the frequency domain self-evolutionary graph neural model. The node state is dynamically corrected according to the constraint confidence weights, and the type, location coordinates, and confidence results of the underground cable fault are output.

[0009] Optionally, the multi-frequency excitation signal includes a combination of low-frequency, medium-frequency, and high-frequency excitation signals.

[0010] Optionally, the electromagnetic response characteristics include electromagnetic response amplitude, phase, direction, and burial depth information.

[0011] Optionally, the weighted directed topology data forming the underground cable includes: Based on the multi-source signal sample set and the photovoltaic power station cable layout diagram, the type and number of topology nodes are determined. Combiner boxes, junction boxes, branch points and grounding points are taken as nodes. The plane coordinates and burial depth information of each node are recorded using a unified coordinate system, and a unique identifier is assigned to each node. Based on the actual physical connection of the cable and the direction of current transmission, determine the directed connection relationship between adjacent nodes. For each actual cable segment, establish a directed connection entry from the upstream node to the downstream node, recording the start point, end point, direction, number of parallel roots and branch positions. Do not establish a connection between nodes that do not have a physical connection. For each directed connection, the length, burial depth, and soil conductivity category of the recorded cable segment are normalized with unified dimensions and value ranges. Then, they are combined into a single edge weight according to the preset non-negative weight coefficients. The sum of each item of the non-negative weight coefficient is equal to one. At the same time, the length, burial depth, soil conductivity category and multi-frequency amplitude phase intensity index are saved as edge attribute entries. For each node, record the node voltage, node current, grounding resistance, and grounding direction identifier. Construct node attribute entries according to a unified field and maintain a one-to-one correspondence between node attributes and unique node identifiers. The node list, coordinates, node attribute entries, directed connection list, edge attribute entries, and corresponding edge weights are summarized to form the weighted directed topology data of the underground cable.

[0012] Optionally, the process of forming the frequency domain-spatial joint feature input includes: Read weighted directed topology data, obtain node list, node attribute entries, node coordinates, directed connection list, edge attribute entries and edge weights, and extract multi-frequency signal features corresponding to each directed connection from the multi-source signal sample set; For each directed connection, frequency mapping is performed. A set of virtual frequency points is generated based on the cable segment length, burial depth, soil conductivity type, and multi-frequency signal characteristics of the directed connection. At the same time, two types of reference anchor frequency points are set to correspond to the low-attenuation frequency band and the high-resolution frequency band, respectively. The range of values ​​of the virtual frequency points is constrained in the same way as the available frequency bands of the transmitter, and the correspondence between the original frequency points and the virtual frequency points is recorded. For each directed connection, frequency band selection is performed. Based on the information gain score, signal-to-noise ratio score, and score consistent with the direction of the directed connection, a selection indicator is calculated. A virtual frequency band is selected as the effective frequency band. Cross-correlation coefficient suppression is performed on the virtual frequency bands of the parallel cables to remove redundant frequency bands, forming a list of effective virtual frequency bands for the directed connection. For each directed connection, feature reconstruction is performed within the effective virtual frequency band, feature components are generated according to virtual frequency points, and combined in order of frequency points to form the virtual frequency domain feature vector of the directed connection. The virtual frequency domain feature vector is cascaded and fused with edge attribute entries and edge weights to form edge-level joint features. The node attribute entries and node coordinates are cascaded and fused to form node-level joint features. The output is a frequency domain-space joint feature input consisting of node-level joint features, edge-level joint features, and weighted directed topological relationship data.

[0013] Optionally, the failure probability, uncertainty matrix, and frequency response estimation results of the output node and cable segment include: Node-level joint features are used as the initial state of nodes in the frequency domain self-evolving graph neural model, edge-level joint features are used as the initial state of edges in the frequency domain self-evolving graph neural model, and weighted directed topological relation data are used as the input of directed connection relations in the frequency domain self-evolving graph neural model. In the frequency domain self-evolutionary graph neural model, a three-layer structure is set up, consisting of a frequency domain adaptive attention layer, a physical guidance graph message passing layer, and a fault assessment and uncertainty estimation layer connected in series. In the frequency domain adaptive attention layer, the importance of each virtual frequency band of each directed connection is scored based on the virtual frequency point features, multi-frequency signal features, and edge weights in the edge-level joint features. The attention coefficients of the directed connections on different virtual frequency bands are calculated, and the edge-level joint features are weighted and aggregated in the frequency domain dimension to generate the edge-level updated features after frequency domain adaptive weighting. In the physical guidance graph message passing layer, physical guidance coefficients are generated for each directed connection based on cable segment length, burial depth, soil conductivity type, and directed connection direction information. Combined with edge-level update features, each node is weighted and aggregated according to the edge-level update features and physical guidance coefficients of the incident directed connection to obtain the node intermediate state. Each directed connection is updated according to the intermediate states of the two end nodes, edge-level update features, and physical guidance coefficients to obtain the edge intermediate state. Several rounds of node intermediate state update and edge intermediate state update are repeated until the node intermediate state and edge intermediate state converge. In the fault assessment and uncertainty estimation layer, node fault assessment output is generated based on the converged intermediate state of the node. The node fault assessment output includes the fault probability, uncertainty index and frequency response estimation results at the node level. Cable segment fault assessment output is generated based on the converged intermediate state of the edge. The node fault assessment output is then summarized into a node fault assessment result set, and the cable segment fault assessment output is summarized into a cable segment fault assessment result set.

[0014] Optionally, adjusting the excitation frequency and measurement point position of the multi-frequency output transmitter includes: The uncertainty index, fault probability and frequency response estimation results of each cable segment are read from the node fault assessment result set and the cable segment fault assessment result set. Combined with the weighted directed topology data, the position, length and relationship of each cable segment in the topology are determined. Establish a time-series inspection record for each cable segment, and store the changes in uncertainty indicators, fault probability, and frequency response estimation of the cable segment in the current inspection round and in historical inspection rounds to form the time-series response change information of the cable segment. Based on the temporal response change information, uncertainty index, and position in the weighted directed topology of each cable segment, the spatiotemporal evolution potential energy score of the cable segment is calculated. The spatiotemporal evolution potential energy score comprehensively characterizes the instability of the cable segment in the time dimension and its importance in the spatial dimension. All cable segments are sorted according to the spatiotemporal evolution potential energy score. The target cable segment or target cable segment area with the highest spatiotemporal evolution potential energy score is selected as the active detection target area. Combining virtual frequency point characteristics and multi-frequency signal characteristics, the next round of priority excitation frequency combination and candidate excitation frequency band corresponding to the active detection target area are determined. Based on the active detection target area and the corresponding excitation frequency combination, the travel path and measurement point position of the differential coil receiver are planned under the unified coordinate system of the photovoltaic power station. An active detection configuration command containing the target cable segment area, excitation frequency combination and measurement point position is generated. The active detection configuration command is sent to the multi-frequency output transmitter and differential coil receiver for the next round of multi-source signal sample set acquisition, forming a spatiotemporally coupled self-tuning active detection closed loop.

[0015] Optionally, the output of the type, location coordinates, and confidence level of the underground cable fault includes: The node fault probability, cable segment fault probability and corresponding uncertainty index are read from the node fault assessment result set and the cable segment fault assessment result set, and the grounding resistance, grounding direction and node current distribution information are read simultaneously to construct a multi-source constraint input set. In the frequency domain self-evolution graph neural model, a multi-source constraint input interface is set up, and the grounding resistance, grounding direction and node current distribution are mapped as constraint vectors respectively, forming grounding constraint vector, direction constraint vector and current consistency constraint vector, and an independent reliable weight coefficient is assigned to each type of constraint vector. For each node and its adjacent cable segments, a node correction factor is calculated based on the node fault probability, uncertainty index, three types of constraint vectors and their corresponding confidence weights. The node correction factor is then weighted and fused with the current node state to generate a node correction state. For each cable segment, edge state updates are performed based on the correction states of the two adjacent nodes to form the corrected edge state. The corrected node and edge states are input into the constraint feedback layer of the frequency domain self-evolutionary graph neural model. The correction results are iteratively updated in the layer until the node and edge states converge. After convergence, the final state of the node and the final state of the cable segment are output. Based on the final state of the node and the final state of the cable segment, the type of underground cable fault, the location coordinates, and the confidence result are output.

[0016] The beneficial effects of this invention are: This invention overcomes the limitations of traditional photovoltaic buried cable fault detection, which relies on single-frequency signals and fixed algorithm structures, by introducing a frequency-domain self-evolving graph neural network model. It achieves adaptive extraction of multi-frequency signal features and topological structured modeling. Utilizing a multi-source signal acquisition module and a topology relationship construction module, it can automatically establish weighted directed topological relationship data between cable nodes and cable segments, providing an accurate structured input foundation for fault location in complex scenarios. Through a frequency-domain mapping and fusion module, multi-frequency signal features are mapped to virtual frequency points and fused, enabling dynamic learning and reconstruction of frequency distribution, and improving the distinguishability and stability of multi-frequency signals under different soil conditions.

[0017] This invention constructs a three-layer graph neural model structure with a frequency-domain adaptive attention layer, a physical guidance message passing layer, and a fault assessment layer. By combining physical properties and signal characteristics for deep reasoning, it achieves multi-dimensional representation and dynamic aggregation of node and cable segment states. The uncertainty matrix output by the model guides the active detection process. Through spatiotemporal evolution potential energy distribution calculation and a dynamic frequency tuning mechanism, it can automatically adjust the excitation frequency and measurement point location, forming a closed-loop self-learning detection system. This enables accurate identification and predictive detection of potential cable fault areas.

[0018] This invention introduces a multi-source constraint correction mechanism, incorporating grounding resistance, grounding direction, and node current distribution information into a unified dynamic constraint framework. This effectively eliminates redundant interference and signal deviations in complex electrical environments, improving the accuracy and reliability of fault location. Compared with existing technologies, this invention represents a leap from passive detection to active intelligent detection, possessing high precision, high stability, and strong adaptability, significantly enhancing the intelligent operation and maintenance and safety assurance capabilities of buried photovoltaic cables. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0020] Figure 1 This is a schematic diagram of the structure of a photovoltaic underground cable fault location system based on graph neural network proposed in this invention; Figure 2 This is a flowchart illustrating a method for fault location of buried photovoltaic cables based on graph neural networks proposed in this invention. Detailed Implementation

[0021] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0022] refer to Figure 1 A fault location system for underground photovoltaic cables based on graph neural networks includes the following modules: The multi-source signal acquisition module is used to apply multi-frequency excitation signals to the buried cable using a multi-frequency output transmitter, and to collect electromagnetic response characteristics through a differential coil receiver to generate a multi-source signal sample set. The topology construction module is used to establish the topology structure of buried cables based on multi-source signal sample sets and cable layout diagrams, forming weighted directed topology data. The frequency domain mapping and fusion module is used to perform frequency mapping, frequency band selection and feature reconstruction on multi-frequency signal features based on weighted directed topological relationship data, forming a joint frequency domain-space feature input. The frequency domain self-evolutionary graph neural model module is used to receive frequency domain-space joint feature input, perform feature propagation and aggregation of nodes and edges, and output the fault probability and uncertainty results of nodes and cable segments. The spatiotemporal active detection module is used to calculate the spatiotemporal evolution potential energy distribution based on the uncertainty results, select the cable segment with the maximum potential energy as the active detection target, and form a self-tuning active detection closed loop. The multi-source constraint correction module is used to input the grounding resistance, grounding direction and node current distribution as multi-source constraints into the frequency domain self-evolution graph neural model, and output the type, location and confidence results of the buried cable fault.

[0023] refer to Figure 2 A fault location method for underground photovoltaic cables based on graph neural networks includes: In photovoltaic power plants, a multi-frequency output transmitter is used to apply multi-frequency excitation signals to buried cables. The electromagnetic response characteristics of each frequency point are collected by a differential coil receiver, and multi-frequency signal characteristics are extracted and formed. At the same time, the grounding resistance, grounding direction and node current distribution are measured to generate a multi-source signal sample set. Based on the multi-source signal sample set and the cable layout diagram of the photovoltaic power station, the topological relationship structure of the underground cable is established. The junction box, junction box, branch point and grounding point are defined as nodes, and the cable segments between adjacent nodes are defined as edges, forming the weighted directed topological relationship data of the underground cable. Based on weighted directed topological relation data, a frequency domain feature mapping and fusion unit is constructed to perform frequency mapping, frequency band selection and feature reconstruction on multi-frequency signal features, generate virtual frequency point features, and fuse the virtual frequency domain features with the weighted directed topological relation data to form a frequency domain-space joint feature input; The frequency domain-space joint features are input into the frequency domain self-evolutionary graph neural model. The feature propagation and aggregation of nodes and edges are performed through the physical guided message passing mechanism, and the fault probability, uncertainty matrix and frequency response estimation results of nodes and cable segments are output. Based on the uncertainty matrix, the spatiotemporal evolution potential energy distribution is calculated, and the cable segment corresponding to the region with the maximum potential energy is selected as the active detection target. The excitation frequency of the multi-frequency output transmitter and the position of the measurement point are adjusted to form a spatiotemporally coupled self-tuning active detection closed loop. The grounding resistance, grounding direction, and node current distribution are used as multi-source constraint inputs to the frequency domain self-evolutionary graph neural model. The node state is dynamically corrected according to the constraint confidence weights, and the type, location coordinates, and confidence results of the underground cable fault are output.

[0024] In this embodiment, the multi-frequency excitation signal includes a combination of low-frequency, medium-frequency, and high-frequency excitation signals.

[0025] In this embodiment, the electromagnetic response characteristics include electromagnetic response amplitude, phase, direction, and burial depth information.

[0026] In this embodiment, the weighted directed topology data forming the buried cable includes: Based on the multi-source signal sample set and the photovoltaic power station cable layout diagram, the type and number of topology nodes are determined. Combiner boxes, junction boxes, branch points and grounding points are taken as nodes. The plane coordinates and burial depth information of each node are recorded using a unified coordinate system, and a unique identifier is assigned to each node. Based on the actual physical connection of the cable and the direction of current transmission, determine the directed connection relationship between adjacent nodes. For each actual cable segment, establish a directed connection entry from the upstream node to the downstream node, recording the start point, end point, direction, number of parallel roots and branch positions. Do not establish a connection between nodes that do not have a physical connection. For each directed connection, the length, burial depth, and soil conductivity category of the recorded cable segment are normalized with unified dimensions and value ranges. Then, they are combined into a single edge weight according to the preset non-negative weight coefficients. The sum of each item of the non-negative weight coefficient is equal to one. At the same time, the length, burial depth, soil conductivity category and multi-frequency amplitude phase intensity index are saved as edge attribute entries. For each node, record the node voltage, node current, grounding resistance, and grounding direction identifier. Construct node attribute entries according to a unified field and maintain a one-to-one correspondence between node attributes and unique node identifiers. The node list, coordinates, node attribute entries, directed connection list, edge attribute entries, and corresponding edge weights are summarized to form the weighted directed topology data of the underground cable.

[0027] In this embodiment, the formation of frequency domain-spatial joint feature input includes: Read weighted directed topology data, obtain node list, node attribute entries, node coordinates, directed connection list, edge attribute entries and edge weights, and extract multi-frequency signal features corresponding to each directed connection from the multi-source signal sample set; For each directed connection, frequency mapping is performed. A set of virtual frequency points is generated based on the cable segment length, burial depth, soil conductivity type, and multi-frequency signal characteristics of the directed connection. At the same time, two types of reference anchor frequency points are set to correspond to the low-attenuation frequency band and the high-resolution frequency band, respectively. The range of values ​​of the virtual frequency points is constrained in the same way as the available frequency bands of the transmitter, and the correspondence between the original frequency points and the virtual frequency points is recorded. For each directed connection, frequency band selection is performed. A selection indicator is calculated based on information gain score, signal-to-noise ratio score, and score consistent with the direction of the directed connection, considering the multi-frequency signal characteristics. A virtual frequency band is selected as the effective frequency band. Cross-correlation coefficient suppression is performed on the virtual frequency bands of the parallel cables to remove redundant frequency bands, forming a list of effective virtual frequency bands for the directed connections, where: Information gain score is an index used to measure the amount of information contained in each frequency band when distinguishing cable fault characteristics. It reflects the degree of effective information provided by the frequency band in fault identification by comparing the difference between the signal distribution under a specific frequency band and the overall fault category distribution. Signal-to-noise ratio (SNR) is an indicator used to measure the ratio of useful signal energy to background noise energy in a specific frequency band, representing the relative level of signal quality in that band. The directed connection alignment score is an index used to describe the degree of matching between the direction of frequency band signal propagation and the actual topology connection direction of the cable. It is determined by calculating the consistency of the angle between the signal phase propagation direction and the geometric direction of the directed connection. For each directed connection, feature reconstruction is performed within the effective virtual frequency band, feature components are generated according to virtual frequency points, and combined in order of frequency points to form the virtual frequency domain feature vector of the directed connection. The virtual frequency domain feature vector is cascaded and fused with edge attribute entries and edge weights to form edge-level joint features. The node attribute entries and node coordinates are cascaded and fused to form node-level joint features. The output is a frequency domain-space joint feature input consisting of node-level joint features, edge-level joint features, and weighted directed topological relationship data.

[0028] In this embodiment, the failure probability, uncertainty matrix, and frequency response estimation results of the output node and cable segment include: Node-level joint features are used as the initial state of nodes in the frequency domain self-evolving graph neural model, edge-level joint features are used as the initial state of edges in the frequency domain self-evolving graph neural model, and weighted directed topological relation data are used as the input of directed connection relations in the frequency domain self-evolving graph neural model. In the frequency domain self-evolutionary graph neural model, a three-layer structure is set up, consisting of a frequency domain adaptive attention layer, a physical guidance graph message passing layer, and a fault assessment and uncertainty estimation layer connected in series. In the frequency domain adaptive attention layer, the importance of each virtual frequency band of each directed connection is scored based on the virtual frequency point features, multi-frequency signal features, and edge weights in the edge-level joint features. The attention coefficients of the directed connections on different virtual frequency bands are calculated, and the edge-level joint features are weighted and aggregated in the frequency domain dimension to generate the edge-level updated features after frequency domain adaptive weighting. In the physical guidance graph message passing layer, physical guidance coefficients are generated for each directed connection based on cable segment length, burial depth, soil conductivity type, and directed connection direction information. Combined with edge-level update features, each node is weighted and aggregated according to the edge-level update features and physical guidance coefficients of the incident directed connection to obtain the node intermediate state. Each directed connection is updated according to the intermediate states of the two end nodes, edge-level update features, and physical guidance coefficients to obtain the edge intermediate state. Several rounds of node intermediate state update and edge intermediate state update are repeated until the node intermediate state and edge intermediate state converge. In the fault assessment and uncertainty estimation layer, node fault assessment output is generated based on the converged intermediate state of the node. The node fault assessment output includes the fault probability, uncertainty index and frequency response estimation results at the node level. Cable segment fault assessment output is generated based on the converged intermediate state of the edge. The node fault assessment output is then summarized into a node fault assessment result set, and the cable segment fault assessment output is summarized into a cable segment fault assessment result set.

[0029] In this embodiment, adjusting the excitation frequency and measurement point position of the multi-frequency output transmitter includes: The uncertainty index, fault probability and frequency response estimation results of each cable segment are read from the node fault assessment result set and the cable segment fault assessment result set. Combined with the weighted directed topology data, the position, length and relationship of each cable segment in the topology are determined. Establish a time-series inspection record for each cable segment, and store the changes in uncertainty indicators, fault probability, and frequency response estimation of the cable segment in the current inspection round and in historical inspection rounds to form the time-series response change information of the cable segment. Based on the temporal response change information, uncertainty index, and position in the weighted directed topology of each cable segment, the spatiotemporal evolution potential energy score of the cable segment is calculated. This spatiotemporal evolution potential energy score comprehensively characterizes the instability of the cable segment in the time dimension and its importance in the spatial dimension. Specifically, the calculation of the spatiotemporal evolution potential energy score of the cable segment involves: Time series analysis was performed on the historical electromagnetic response characteristics of each cable segment under multi-frequency excitation conditions. The response change amplitude and spectral drift between adjacent time points were calculated, and the response volatility, slope change and signal abrupt change were extracted. The overall response instability score of the cable segment within the detection period was obtained. Using the node uncertainty matrix output by the frequency domain self-evolution graph neural model, the uncertainty values ​​of the nodes connecting the two ends of the cable segment are weighted and summed. At the same time, the path centrality and criticality indices in the weighted directed topology are combined to evaluate the sensitivity of the cable segment in the overall structure and form a spatial structure importance score. After normalizing the instability score in the time dimension and the structural importance score in the spatial dimension, a weighted fusion is performed to obtain the spatiotemporal evolution potential energy score of the cable segment. All cable segments are sorted according to the spatiotemporal evolution potential energy score. The target cable segment or target cable segment area with the highest spatiotemporal evolution potential energy score is selected as the active detection target area. Combining virtual frequency point characteristics and multi-frequency signal characteristics, the next round of priority excitation frequency combination and candidate excitation frequency band corresponding to the active detection target area are determined. Based on the active detection target area and the corresponding excitation frequency combination, the travel path and measurement point position of the differential coil receiver are planned under the unified coordinate system of the photovoltaic power station. An active detection configuration command containing the target cable segment area, excitation frequency combination and measurement point position is generated. The active detection configuration command is sent to the multi-frequency output transmitter and differential coil receiver for the next round of multi-source signal sample set acquisition, forming a spatiotemporally coupled self-tuning active detection closed loop.

[0030] In this embodiment, the output of the type, location coordinates, and confidence level of the underground cable fault includes: The node fault probability, cable segment fault probability and corresponding uncertainty index are read from the node fault assessment result set and the cable segment fault assessment result set, and the grounding resistance, grounding direction and node current distribution information are read simultaneously to construct a multi-source constraint input set. In the frequency domain self-evolution graph neural model, a multi-source constraint input interface is set up, and the grounding resistance, grounding direction and node current distribution are mapped as constraint vectors respectively, forming grounding constraint vector, direction constraint vector and current consistency constraint vector, and an independent reliable weight coefficient is assigned to each type of constraint vector. For each node and its adjacent cable segments, a node correction factor is calculated based on the node fault probability, uncertainty index, three types of constraint vectors and their corresponding confidence weights. The node correction factor is then weighted and fused with the current node state to generate a node correction state. For each cable segment, edge state updates are performed based on the correction states of the two adjacent nodes to form the corrected edge state. The corrected node and edge states are input into the constraint feedback layer of the frequency domain self-evolutionary graph neural model. The correction results are iteratively updated in the layer until the node and edge states converge. After convergence, the final state of the node and the final state of the cable segment are output. Based on the final state of the node and the final state of the cable segment, the type of underground cable fault, the location coordinates, and the confidence result are output.

[0031] Example 1: To verify the feasibility of this invention in practice, it was applied to a photovoltaic power station with an installed capacity of 100MW. The station has 62 buried DC combiner cables with a total length of approximately 22.6 kilometers and an average burial depth of 0.9 meters. The surface soil is a mixture of cohesive and sandy layers, and is subject to long-term seasonal humidity changes, resulting in large fluctuations in grounding resistance and frequent degradation of the insulation performance of the DC branches. In July 2025, during routine maintenance inspections, the power station discovered intermittent grounding alarm signals in some branches. Traditional low-frequency detection methods lacked sufficient accuracy, with an error of approximately 4-6 meters, and the average time to detect one cable was about 40 minutes. To address this issue, the photovoltaic buried cable fault location method based on graph neural networks of this invention was used for field verification at this power station.

[0032] During implementation, a multi-source signal acquisition module was first used to generate a combined excitation signal ranging from 1kHz to 45kHz via a multi-frequency output transmitter. This signal was then used to segmentally excite the target cable, and a differential coil receiver was used to acquire the electromagnetic response characteristics. The signals acquired on-site included electromagnetic response amplitude, phase, and direction data. Simultaneously, a DC grounding detection device recorded the grounding resistance (3.2Ω~6.8Ω) and direction information of each node, and a wireless current acquisition device obtained the node current distribution. Subsequently, the system automatically generated a multi-source signal sample set and established weighted topology data containing 248 nodes and 356 directed connections based on the cable layout diagram. The system then performed frequency mapping, frequency band selection, and virtual frequency point reconstruction through a frequency domain feature mapping and fusion module, ultimately forming 22 virtual frequency point feature vectors and generating a frequency domain-space joint feature input.

[0033] During the model inference phase, the frequency domain self-evolutionary graph neural model operates in a three-layer structure: the first layer is a frequency domain adaptive attention layer, used to extract the weight distribution of multi-frequency features in different signal channels; the second layer is a physical-guided message passing layer, which dynamically propagates node features based on cable length, burial depth, and soil conductivity; the third layer is a fault assessment and uncertainty estimation layer, outputting the fault probability, uncertainty matrix, and frequency response estimation results for each node and cable segment. The average model computation time is 8.5 seconds. Based on the output results, the system calculates the spatiotemporal evolution potential energy distribution and automatically identifies the cable segment region with the highest potential energy. The multi-frequency excitation frequencies are automatically tuned to 17.6kHz and 33.2kHz for active detection, and the differential coil receiver automatically adjusts the measurement point spacing from the original 2 meters to 0.8 meters, achieving fine scanning. The final detection results form a closed-loop update in the system.

[0034] During the testing period, a total of 62 buried cables were inspected, among which 6 grounding faults were found (including 2 high-resistance grounding faults, 3 low-resistance grounding faults, and 1 latent short-circuit fault). The system's positioning error was controlled within 0.7 meters, and on-site excavation verified that all fault points were accurately located. Compared with traditional methods, the average inspection time was reduced to 13 minutes per cable, improving inspection efficiency by approximately 67%, and the multi-frequency signal recognition stability rate was improved to 97.8%. Under different soil moisture contents (16%–32%), the system's frequency response signal-to-noise ratio remained above 24dB, an improvement of approximately 30% compared to manual judgment, verifying the system's anti-interference capability and algorithm stability in complex electromagnetic environments.

[0035] On-site technicians reported that the system achieves an automated closed loop for detection, analysis, and location, and can adjust the detection frequency and measurement point spacing in real time without requiring manual frequency band selection, thus reducing operational complexity. By dynamically correcting the multi-source constraint input model using grounding resistance, direction, and current distribution, the final output fault location confidence level is higher than 0.92, and the results are visualized through the photovoltaic operation and maintenance monitoring platform. Compared with manual inspection, this invention improves detection speed and location accuracy while reducing manual intervention by approximately 40% and the number of high-support climbing operations by 60%, thereby improving the safety and operation and maintenance efficiency of photovoltaic power plants.

[0036] Table 1 Comparison of Fault Detection Results for Underground Cables in Photovoltaic Power Stations

[0037] As shown in Table 1, the method of this invention has significant advantages over traditional methods. The traditional low-frequency detection method takes an average of 40.2 minutes to detect a single buried cable, while the multi-frequency manual analysis method requires 27.5 minutes. In contrast, the method of this invention reduces the detection time for a single cable to 13.1 minutes, a reduction of more than half. This is due to the combined use of unified acquisition of multi-source signals, frequency domain mapping, and batch inference using graph neural networks. This reduces the time spent on repeated manual frequency switching, cable routing, and reverse positioning, allowing for comprehensive detection of more circuits within the same working period and improving on-site maintenance efficiency.

[0038] In terms of detection accuracy and stability, the method of this invention also has significant advantages. Traditional low-frequency detection methods have an average positioning error of 5.4 meters, and multi-frequency manual analysis methods have an error of 3.2 meters, while the method of this invention controls the error to the order of 0.7 meters, meeting the precise requirements for fault location in underground cable excavation. Regarding fault identification success rates, traditional methods achieve 85.3% and 90.6% respectively, while the method of this invention reaches 98.1%, indicating that in complex soil environments and under conditions of multiple cables operating in parallel, the frequency domain self-evolutionary graph neural model can more reliably distinguish the target cable from interference signals. The method of this invention achieves a higher average signal-to-noise ratio (24.3 dB) within the same moisture content range, and the uncertainty converges within 3-4 iterations, demonstrating that the algorithm's convergence efficiency and result stability are superior to traditional empirical analysis.

[0039] From the perspectives of intelligence level and operation and maintenance cost, the advantages of the method of this invention are also quite prominent. Traditional methods lack automatic frequency adjustment capabilities or can only roughly adjust the frequency. The automatic frequency adjustment accuracy deviation of the method of this invention is controlled within ±0.3kHz, indicating that the system can finely optimize the excitation frequency based on uncertainty and spatiotemporal potential energy, making each detection more targeted. The number of high-altitude operations on high supports is reduced from 22 times per day in the traditional low-frequency detection method to 6 times, significantly reducing high-risk operations for maintenance personnel and significantly improving both safety and labor intensity. In terms of power consumption, the method of this invention, while achieving higher computing power and automation functions, still controls the average detection power consumption at 49W, which is slightly lower than that of traditional methods, demonstrating the feasibility and economy of the overall system in engineering applications. In summary, this invention has achieved comprehensive improvements in detection efficiency, positioning accuracy, intelligence level, and operation and maintenance safety.

[0040] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A fault location system for underground photovoltaic cables based on graph neural networks, characterized in that, Includes the following modules: The multi-source signal acquisition module is used to apply multi-frequency excitation signals to the buried cable using a multi-frequency output transmitter, and to collect electromagnetic response characteristics through a differential coil receiver to generate a multi-source signal sample set. The topology construction module is used to establish the topology structure of buried cables based on multi-source signal sample sets and cable layout diagrams, forming weighted directed topology data. The frequency domain mapping and fusion module is used to perform frequency mapping, frequency band selection and feature reconstruction on multi-frequency signal features based on weighted directed topological relationship data, forming a joint frequency domain-space feature input. The frequency domain self-evolutionary graph neural model module is used to receive frequency domain-space joint feature input, perform feature propagation and aggregation of nodes and edges, and output the fault probability and uncertainty results of nodes and cable segments. The spatiotemporal active detection module is used to calculate the spatiotemporal evolution potential energy distribution based on the uncertainty results, select the cable segment with the maximum potential energy as the active detection target, and form a self-tuning active detection closed loop. The multi-source constraint correction module is used to input the grounding resistance, grounding direction and node current distribution as multi-source constraints into the frequency domain self-evolution graph neural model, and output the type, location and confidence results of the buried cable fault.

2. A method for fault location of underground photovoltaic cables based on graph neural networks, applied to the fault location system for underground photovoltaic cables based on graph neural networks as described in claim 1, characterized in that, include: In photovoltaic power plants, a multi-frequency output transmitter is used to apply multi-frequency excitation signals to buried cables. The electromagnetic response characteristics of each frequency point are collected by a differential coil receiver, and multi-frequency signal characteristics are extracted and formed. At the same time, the grounding resistance, grounding direction and node current distribution are measured to generate a multi-source signal sample set. Based on the multi-source signal sample set and the cable layout diagram of the photovoltaic power station, the topological relationship structure of the underground cable is established. The junction box, junction box, branch point and grounding point are defined as nodes, and the cable segments between adjacent nodes are defined as edges, forming the weighted directed topological relationship data of the underground cable. Based on weighted directed topological relation data, a frequency domain feature mapping and fusion unit is constructed to perform frequency mapping, frequency band selection and feature reconstruction on multi-frequency signal features, generate virtual frequency point features, and fuse the virtual frequency domain features with the weighted directed topological relation data to form a frequency domain-space joint feature input; The frequency domain-space joint features are input into the frequency domain self-evolutionary graph neural model. The feature propagation and aggregation of nodes and edges are performed through the physical guided message passing mechanism, and the fault probability, uncertainty matrix and frequency response estimation results of nodes and cable segments are output. Based on the uncertainty matrix, the spatiotemporal evolution potential energy distribution is calculated, and the cable segment corresponding to the region with the maximum potential energy is selected as the active detection target. The excitation frequency of the multi-frequency output transmitter and the position of the measurement point are adjusted to form a spatiotemporally coupled self-tuning active detection closed loop. The grounding resistance, grounding direction, and node current distribution are used as multi-source constraint inputs to the frequency domain self-evolutionary graph neural model. The node state is dynamically corrected according to the constraint confidence weights, and the type, location coordinates, and confidence results of the underground cable fault are output.

3. The method for fault location of buried photovoltaic cables based on graph neural networks according to claim 2, characterized in that, The multi-frequency excitation signal includes a combination of low-frequency, medium-frequency, and high-frequency excitation signals.

4. The method for fault location of buried photovoltaic cables based on graph neural networks according to claim 2, characterized in that, The electromagnetic response characteristics include electromagnetic response amplitude, phase, direction, and burial depth information.

5. The method for fault location of buried photovoltaic cables based on graph neural networks according to claim 2, characterized in that, The weighted directed topology data forming the underground cable includes: Based on the multi-source signal sample set and the photovoltaic power station cable layout diagram, the type and number of topology nodes are determined. Combiner boxes, junction boxes, branch points and grounding points are taken as nodes. The plane coordinates and burial depth information of each node are recorded using a unified coordinate system, and a unique identifier is assigned to each node. Based on the actual physical connection of the cable and the direction of current transmission, determine the directed connection relationship between adjacent nodes. For each actual cable segment, establish a directed connection entry from the upstream node to the downstream node, recording the start point, end point, direction, number of parallel roots and branch positions. Do not establish a connection between nodes that do not have a physical connection. For each directed connection, the length, burial depth, and soil conductivity category of the recorded cable segment are normalized with unified dimensions and value ranges. Then, they are combined into a single edge weight according to the preset non-negative weight coefficients. The sum of each item of the non-negative weight coefficient is equal to one. At the same time, the length, burial depth, soil conductivity category and multi-frequency amplitude phase intensity index are saved as edge attribute entries. For each node, record the node voltage, node current, grounding resistance, and grounding direction identifier. Construct node attribute entries according to a unified field and maintain a one-to-one correspondence between node attributes and unique node identifiers. The node list, coordinates, node attribute entries, directed connection list, edge attribute entries, and corresponding edge weights are summarized to form the weighted directed topology data of the underground cable.

6. The method for fault location of buried photovoltaic cables based on graph neural networks according to claim 2, characterized in that, The formation of the frequency domain-spatial joint feature input includes: Read weighted directed topology data, obtain node list, node attribute entries, node coordinates, directed connection list, edge attribute entries and edge weights, and extract multi-frequency signal features corresponding to each directed connection from the multi-source signal sample set; For each directed connection, frequency mapping is performed. A set of virtual frequency points is generated based on the cable segment length, burial depth, soil conductivity type, and multi-frequency signal characteristics of the directed connection. At the same time, two types of reference anchor frequency points are set to correspond to the low-attenuation frequency band and the high-resolution frequency band, respectively. The range of values ​​of the virtual frequency points is constrained in the same way as the available frequency bands of the transmitter, and the correspondence between the original frequency points and the virtual frequency points is recorded. For each directed connection, frequency band selection is performed. Based on the information gain score, signal-to-noise ratio score, and score consistent with the direction of the directed connection, a selection indicator is calculated. A virtual frequency band is selected as the effective frequency band. Cross-correlation coefficient suppression is performed on the virtual frequency bands of the parallel cables to remove redundant frequency bands, forming a list of effective virtual frequency bands for the directed connection. For each directed connection, feature reconstruction is performed within the effective virtual frequency band, feature components are generated according to virtual frequency points, and combined in order of frequency points to form the virtual frequency domain feature vector of the directed connection. The virtual frequency domain feature vector is cascaded and fused with edge attribute entries and edge weights to form edge-level joint features. The node attribute entries and node coordinates are cascaded and fused to form node-level joint features. The output is a frequency domain-space joint feature input consisting of node-level joint features, edge-level joint features, and weighted directed topological relationship data.

7. The method for fault location of buried photovoltaic cables based on graph neural networks according to claim 2, characterized in that, The failure probability, uncertainty matrix, and frequency response estimation results of the output node and cable segment include: Node-level joint features are used as the initial state of nodes in the frequency domain self-evolving graph neural model, edge-level joint features are used as the initial state of edges in the frequency domain self-evolving graph neural model, and weighted directed topological relation data are used as the input of directed connection relations in the frequency domain self-evolving graph neural model. In the frequency domain self-evolutionary graph neural model, a three-layer structure is set up, consisting of a frequency domain adaptive attention layer, a physical guidance graph message passing layer, and a fault assessment and uncertainty estimation layer connected in series. In the frequency domain adaptive attention layer, the importance of each virtual frequency band of each directed connection is scored based on the virtual frequency point features, multi-frequency signal features, and edge weights in the edge-level joint features. The attention coefficients of the directed connections on different virtual frequency bands are calculated, and the edge-level joint features are weighted and aggregated in the frequency domain dimension to generate the edge-level updated features after frequency domain adaptive weighting. In the physical guidance graph message passing layer, physical guidance coefficients are generated for each directed connection based on cable segment length, burial depth, soil conductivity type, and directed connection direction information. Combined with edge-level update features, each node is weighted and aggregated according to the edge-level update features and physical guidance coefficients of the incident directed connection to obtain the node intermediate state. Each directed connection is updated according to the intermediate states of the two end nodes, edge-level update features, and physical guidance coefficients to obtain the edge intermediate state. Several rounds of node intermediate state update and edge intermediate state update are repeated until the node intermediate state and edge intermediate state converge. In the fault assessment and uncertainty estimation layer, node fault assessment output is generated based on the converged intermediate state of the node. The node fault assessment output includes the fault probability, uncertainty index and frequency response estimation results at the node level. Cable segment fault assessment output is generated based on the converged intermediate state of the edge. The node fault assessment output is then summarized into a node fault assessment result set, and the cable segment fault assessment output is summarized into a cable segment fault assessment result set.

8. The method for fault location of buried photovoltaic cables based on graph neural networks according to claim 2, characterized in that, The adjustment of the excitation frequency and measurement point position of the multi-frequency output transmitter includes: The uncertainty index, fault probability and frequency response estimation results of each cable segment are read from the node fault assessment result set and the cable segment fault assessment result set. Combined with the weighted directed topology data, the position, length and relationship of each cable segment in the topology are determined. Establish a time-series inspection record for each cable segment, and store the changes in uncertainty indicators, fault probability, and frequency response estimation of the cable segment in the current inspection round and in historical inspection rounds to form the time-series response change information of the cable segment. Based on the temporal response change information, uncertainty index, and position in the weighted directed topology of each cable segment, the spatiotemporal evolution potential energy score of the cable segment is calculated. The spatiotemporal evolution potential energy score comprehensively characterizes the instability of the cable segment in the time dimension and its importance in the spatial dimension. All cable segments are sorted according to the spatiotemporal evolution potential energy score. The target cable segment or target cable segment area with the highest spatiotemporal evolution potential energy score is selected as the active detection target area. Combining virtual frequency point characteristics and multi-frequency signal characteristics, the next round of priority excitation frequency combination and candidate excitation frequency band corresponding to the active detection target area are determined. Based on the active detection target area and the corresponding excitation frequency combination, the travel path and measurement point position of the differential coil receiver are planned under the unified coordinate system of the photovoltaic power station. An active detection configuration command containing the target cable segment area, excitation frequency combination and measurement point position is generated. The active detection configuration command is sent to the multi-frequency output transmitter and differential coil receiver for the next round of multi-source signal sample set acquisition, forming a spatiotemporally coupled self-tuning active detection closed loop.

9. A method for fault location of buried photovoltaic cables based on graph neural networks according to claim 2, characterized in that, The output of the type, location coordinates, and confidence level of the underground cable fault includes: The node fault probability, cable segment fault probability and corresponding uncertainty index are read from the node fault assessment result set and the cable segment fault assessment result set, and the grounding resistance, grounding direction and node current distribution information are read simultaneously to construct a multi-source constraint input set. In the frequency domain self-evolution graph neural model, a multi-source constraint input interface is set up, and the grounding resistance, grounding direction and node current distribution are mapped as constraint vectors respectively, forming grounding constraint vector, direction constraint vector and current consistency constraint vector, and an independent reliable weight coefficient is assigned to each type of constraint vector. For each node and its adjacent cable segments, a node correction factor is calculated based on the node fault probability, uncertainty index, three types of constraint vectors and their corresponding confidence weights. The node correction factor is then weighted and fused with the current node state to generate a node correction state. For each cable segment, edge state updates are performed based on the correction states of the two adjacent nodes to form the corrected edge state. The corrected node and edge states are input into the constraint feedback layer of the frequency domain self-evolutionary graph neural model. The correction results are iteratively updated in the layer until the node and edge states converge. After convergence, the final state of the node and the final state of the cable segment are output. Based on the final state of the node and the final state of the cable segment, the type of underground cable fault, the location coordinates, and the confidence result are output.

Images