A distribution network topology graph actual difference diagnosis method based on a graph neural network

Abstract

The application belongs to the technical field of power distribution network operation and maintenance, and relates to a power distribution network topology graph actual difference diagnosis method based on a graph neural network, which comprises the following steps: injecting an initial detection signal sequence into a power distribution network, collecting and integrating responses generated by collaborative sensing terminals, generating a graph node original feature set, performing cross verification, generating an initial topology connection feature graph with increased confidence, inputting the initial topology connection feature graph into a preset graph neural network model, and generating a topology diagnosis probability graph; according to the topology diagnosis probability graph, a secondary detection signal sequence is generated, the secondary detection signal sequence is injected into the power distribution network, a focused response data set is generated in an abnormal area, the abnormal area is iteratively diagnosed based on the focused response data set, and a final topology difference positioning report is generated; the application solves the problem that modern power grids cannot support large-scale and normalized topology verification requirements under the requirements of real-time, accuracy and economic operation and maintenance.

Description

Technical Field

[0001] This invention belongs to the technical field of distribution network operation and maintenance, and relates to a method for diagnosing differences between distribution network topology and reality based on graph neural networks. Background Technology

[0002] The accuracy of the distribution network topology is the foundation for ensuring the safe and stable operation of the power grid and realizing efficient operation and maintenance decisions. However, in actual operation, due to various reasons such as construction changes, equipment aging, and routine inspections, the actual physical connection state of the power grid often differs from the archived theoretical topology diagram. This difference causes certain interference to key operations such as fault location, load transfer, and line loss calculation.

[0003] Currently, the industry commonly employs passive data analysis and traditional manual inspection. Passive data analysis methods typically utilize daily operational data collected by systems such as distribution automation systems and advanced metering architectures, inferring the connection relationships between nodes by analyzing the statistical correlations of parameters such as voltage, current, and power. Manual inspection, on the other hand, identifies topological differences through manual comparison.

[0004] However, methods relying on passive data analysis are highly dependent on the operating conditions of the power grid. When the network load is stable or does not change significantly, the signal characteristics are weak, making it difficult to accurately determine the connection relationship and to proactively initiate detection. Although manual on-site verification yields reliable results, it is inefficient and cannot meet the modern power grid's requirements for real-time, accurate, and economical operation and maintenance, nor can it support the large-scale, routine topology verification needs. Summary of the Invention

[0005] To address the aforementioned problems, this invention provides a method for diagnosing discrepancies in distribution network topology based on graph neural networks.

[0006] A method for diagnosing discrepancies between distribution network topology and reality based on graph neural networks includes the following steps:

[0007] S1. Inject an initial detection signal sequence, including timing coding and specific frequency components, into the distribution network to stimulate multiple collaborative sensing terminals distributed in the distribution network to generate responses;

[0008] S2. Collect and integrate the responses generated by the collaborative sensing terminals to generate a structured set of original feature data for the map nodes;

[0009] S3. Perform cross-verification based on the original feature set of structured graph nodes to generate an enhanced initial topology connection feature map for the theoretical topology connection of the distribution network;

[0010] S4. Input the initial topological connectivity feature map into the preset graph neural network model to identify and generate a topological diagnostic probability map that marks abnormal regions;

[0011] S5. Generate a secondary detection signal sequence based on the abnormal regions marked in the topological diagnostic probability map;

[0012] S6. Inject the secondary detection signal sequence into the distribution network to generate a focused response dataset in the abnormal area;

[0013] S7. Based on the focused response dataset, perform iterative diagnosis of abnormal regions and generate the final topological difference localization report.

[0014] A further aspect of the present invention, step S1, includes the following steps:

[0015] The initial detection signal sequence is injected into the power grid through the injection unit set at the beginning of the distribution network feeder or key connection point;

[0016] The initial detection signal sequence propagates along the physical path of the power grid, stimulating each cooperative sensing terminal along the line to synchronously acquire the transient electrical parameters and physical environment parameters of the response.

[0017] A further aspect of this invention involves generating the original feature set of the graph nodes, comprising the following steps:

[0018] The collaborative sensing terminal synchronously collects transient electrical responses and environmental IoT data at its location;

[0019] The collected transient electrical responses are timestamped with the environmental IoT data to generate data units;

[0020] All data units generated by the collaborative sensing terminals are aggregated to form a structured set of original feature features for the graph nodes;

[0021] The transient electrical response includes instantaneous waveforms and phase changes of voltage and current; the environmental IoT data includes infrared temperature readings on the equipment surface and micro-vibration waveforms of the switch box.

[0022] A further aspect of the present invention involves generating an enhanced initial topology connectivity feature map, comprising the following steps:

[0023] For two connected nodes in the theoretical topology, the corresponding transient electrical response is extracted from the original feature set of the structured graph nodes, the attenuation and time delay during signal propagation are calculated, and a preliminary criterion for electrical connectivity is formed.

[0024] Simultaneously extract and compare the environmental IoT data of two related nodes to form auxiliary criteria for judging the physical connection status;

[0025] When the preliminary criteria and the auxiliary criteria corroborate each other, the connection is enhanced and assigned a higher value; otherwise, it is downweighted to generate an enhanced initial topological connection feature map.

[0026] A further aspect of the present invention involves forming an auxiliary criterion, comprising the following steps:

[0027] Extract the infrared temperature readings of two related nodes from the environmental IoT data and calculate the temperature difference between them;

[0028] The temperature difference is compared with a preset abnormal temperature difference threshold to determine whether an abnormal temperature difference exists, and the determination result is used as an auxiliary criterion for the physical connection status.

[0029] A further aspect of the present invention generates a topological diagnostic probability map, comprising the following steps:

[0030] The graph neural network model compares and analyzes the input enhanced initial topology connectivity feature map with a pre-set healthy topology feature library.

[0031] The graph neural network model infers and outputs the probability value of the graph-real difference at each connection in the network.

[0032] Visualize and render the probability values ​​of the graph-to-real-world differences of all connections to form a topology diagnostic probability map.

[0033] A further aspect of the present invention generates a secondary detection signal sequence, comprising the following steps:

[0034] Analyze the topological diagnostic probability map, identify connections whose map-to-real discrepancy probability values ​​exceed a preset probability threshold, and define them as target diagnostic regions.

[0035] Based on the electrical characteristics of the target diagnostic area, select signal frequency and timing combinations from a preset waveform library that match the potential fault types in that area;

[0036] The selected signal frequency and timing combination are used to construct a secondary detection signal sequence.

[0037] A further aspect of the present invention generates a focused response dataset, comprising the following steps:

[0038] The secondary detection signal sequence is then injected back into the distribution network;

[0039] The energy of the secondary detection signal sequence will be concentrated on the target diagnostic area, triggering a response on the collaborative sensing terminal in that area;

[0040] The response data of each terminal within the target diagnostic area is collected again to form a focused response dataset.

[0041] A further aspect of the present invention generates a final topology difference localization report, comprising the following steps:

[0042] The focused response dataset is repeatedly input into the cross-verification process of step S3 and the diagnostic process of step S4 to re-infer the topological state of the target diagnostic region and obtain the updated map-real difference probability value.

[0043] By comparing the changes in the probability values ​​of the difference between the initial diagnosis and the subsequent reasoning, we can obtain confidence conclusions about the fault points and abnormal states within the target diagnosis area.

[0044] Based on the confidence level conclusion, the root cause of the map-to-reals discrepancy is identified and output, forming the final topological discrepancy localization report.

[0045] A further aspect of the present invention is that the final topology difference location report includes: the fault location that caused the map-to-surface difference, the fault type, and a confidence assessment calculated based on the change in the probability value of the map-to-surface difference.

[0046] In summary, the present invention has the following beneficial technical effects:

[0047] 1. By combining active detection signals with secondary detection signals, a two-stage diagnostic mode is achieved, which improves diagnostic efficiency and accuracy. The initial detection can scan the entire network and identify abnormal areas using graph neural networks, avoiding indiscriminate inspection of the entire power grid and reducing the scope and time of diagnosis. Subsequently, secondary detection signals are injected in a targeted manner to conduct in-depth analysis of suspicious areas, thereby improving the reliability of the final diagnostic conclusion.

[0048] 2. Cross-verification through transient electrical response and environmental IoT data enhances the confidence of topology connection judgment. By synchronously collecting and analyzing electrical characteristics and physical characteristics such as equipment temperature rise and micro-vibration during signal propagation, the uncertainty brought by a single information source can be reduced through the mutual verification of multi-dimensional information, and the real physical connection interruption and temporary electrical isolation can be distinguished more accurately, thus improving the robustness of diagnosis.

[0049] 3. By using graph neural network models to learn and infer the global features of network topology, deep-seated and structural topological errors that are difficult to reveal through single connection analysis can be discovered, thereby improving the ability to detect complex topological errors and the level of intelligence in diagnosis.

[0050] 4. The final generated topology difference location report reflects the topology health status of the entire network. For confirmed anomalies, it provides information including location, fault type, and confidence level assessment, providing maintenance personnel with a basis for on-site inspection, reducing the manpower and time costs of fault diagnosis, and transforming maintenance work from passive response to proactive prediction and precise repair. Attached Figure Description

[0051] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. The drawings are used to provide a further understanding of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0052] Figure 1 A flowchart illustrating an embodiment of this application is disclosed.

[0053] Figure 2 Structural schematic diagrams of embodiments of this application are disclosed. Detailed Implementation

[0054] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0055] The following is in conjunction with the appendix Figure 1 - Figure 2 A preferred description of the present invention is provided below.

[0056] See attached document Figure 1 This invention proposes a method for diagnosing discrepancies between distribution network topology and actual data based on graph neural networks, comprising the following steps:

[0057] S1. Inject an initial detection signal sequence, including timing coding and specific frequency components, into the distribution network to stimulate multiple collaborative sensing terminals distributed in the distribution network to generate responses;

[0058] S2. Collect and integrate the responses generated by the collaborative sensing terminals to generate a structured set of original feature data for the map nodes;

[0059] S3. Perform cross-verification based on the original feature set of structured graph nodes to generate an enhanced initial topology connection feature map for the theoretical topology connection of the distribution network;

[0060] S4. Input the initial topological connectivity feature map into the preset graph neural network model to identify and generate a topological diagnostic probability map that marks abnormal regions;

[0061] S5. Generate a secondary detection signal sequence based on the abnormal regions marked in the topological diagnostic probability map;

[0062] S6. Inject the secondary detection signal sequence into the distribution network to generate a focused response dataset in the abnormal area;

[0063] S7. Based on the focused response dataset, perform iterative diagnosis of abnormal regions and generate the final topological difference localization report.

[0064] In one embodiment of the present invention, step S1 includes the following steps:

[0065] The initial detection signal sequence is injected into the power grid through the injection unit set at the beginning of the distribution network feeder or key connection point;

[0066] The initial detection signal sequence propagates along the physical path of the power grid, stimulating each cooperative sensing terminal along the line to synchronously acquire the transient electrical parameters and physical environment parameters of the response.

[0067] The injection process of the initial probe signal sequence consists of three stages:

[0068] The first step involves constructing a general-purpose weak electrical signal combination with specific characteristics. This combination includes a preset timing code and specific frequency components. The preset timing code refers to a pre-designed time pulse sequence, including pulse width, pulse interval, and repetition period. Specifically, the timing code should satisfy the following relationships: the pulse width should be in the millisecond range, generally between 1ms and 100ms; the pulse interval T should be in the second range, generally between 0.1s and 10s; and the repetition period is the sum of the two, forming a predictable yet identifiable time pattern. The selection of specific frequency components is based on the electromagnetic propagation characteristics of the distribution network and typically includes multiple discrete frequency points. These frequencies should avoid integer multiples of the power frequency of the distribution network (50Hz) to reduce interference with normal loads. For example, 105Hz, 505Hz, and 1005Hz are selected. The amplitude of the weak electrical signal is set to a low value that conforms to the transient tolerance range of the distribution network, such as 0.1% to 1% of the rated voltage, so that the signal can excite the collaborative sensing terminal without affecting the normal operation of the power grid.

[0069] The second step involves injecting signals into the power grid using programmable injection units deployed at key locations within the distribution network. These programmable injection units are power electronic devices with precise timing control and signal generation capabilities. Their core function is to generate and inject specific voltage or current signals at designated times according to programmed instructions. Injection units should be deployed at two types of key locations: the feeder head, which refers to the starting point of the main line from the low-voltage side of the distribution transformer to the next branch point, providing the most complete network topology observation view; and key interconnection points, which are the interconnection switches between different branches or distribution transformers in the distribution network, enabling synchronous signal propagation to multiple adjacent areas. Injection units typically require a synchronous clock with timestamp accuracy at the millisecond or even microsecond level to ensure the injected signal has a deterministic start time, facilitating subsequent time delay measurements.

[0070] In the third stage, the initial detection signal sequence propagates along the physical path of the power grid, stimulating distributed collaborative sensing terminals during propagation. The physical path refers to the signal's trajectory within the power grid topology, i.e., the network connectivity formed by voltage equipotentiality and current loops. When the signal originates from the injection point, it undergoes three propagation stages: in the near-field region, the signal propagates along directly connected cables and conductors, with the attenuation rate related to line impedance; in the far-field region, the signal propagates through the coupling of transformer windings, electromagnetic induction, and the interaction of distributed line parameters; at the end, the signal is absorbed or reflected at the load point. Collaborative sensing terminals are sensing and acquisition devices distributed at various key nodes in the distribution network, including transformer grounding, distribution switchgear, and branch line tap points. When these terminals receive the arriving signal, they synchronously capture it through internal detection circuits. Synchronous capture means that the time error at the start of signal reception by each terminal should be controlled within an acceptable range, typically not exceeding 1 ms, ensuring that the transient responses acquired by different terminals correspond to each other in the time dimension.

[0071] The initial probe signal sequence is a time-varying electrical signal composed of a preset timing code and specific frequency components superimposed linearly. The mathematical representation of this sequence is:

[0072]

[0073] in, This indicates the injected voltage signal. Let n be the amplitude of the nth frequency component. For the nth frequency component, For time variables, For phase shift, This is a timing coding function, whose value is 0 or 1, representing whether the signal is activated at that moment.

[0074] The setting of specific frequency components is based on research into the propagation characteristics of low-frequency bands in distribution networks. According to power system transient analysis theory, distribution networks exhibit good predictability in terms of attenuation and propagation delay for low-frequency signals, i.e., 1Hz to 10kHz. Signals in this frequency band can penetrate the entire distribution network with controllable attenuation. Amplitude The setting should be based on the rated operating voltage of the distribution network, generally between one-thousandth and one-hundredth of the rated voltage. For example, in a 380V distribution network, the amplitude should be controlled within the range of 0.4V to 3.8V. This range selection ensures that the signal is sufficient to excite the detection circuit of the collaborative sensing terminal without interfering with the normal operation of the power grid and the relay protection device.

[0075] A programmable injection unit is a physical actuator for signal injection. It comprises: hardware employing a power electronic topology, such as an H-bridge full-bridge topology or a boost chopper topology, to generate the required time-varying voltage; and software featuring a high-precision time synchronization mechanism, achieving millisecond- or microsecond-level time alignment with the collaborative sensing terminal via GPS or Precision Time Protocol (PTP). The operating frequency of the programmable injection unit should be at least ten times higher than the required injection signal frequency. For example, if the injection frequency is 1kHz, the switching frequency of the injection unit should be no less than 10kHz to ensure the fidelity of the generated signal waveform.

[0076] Collaborative sensing terminals are distributed IoT sensing devices with multi-channel synchronous sampling capabilities, enabling them to collect multiple physical quantities simultaneously. Typically, the sampling rate should be no less than 20 times the highest detection frequency; that is, if the highest detection frequency is 1kHz, the sampling rate should be no less than 20kHz. The sampling accuracy should reach a digital resolution of 12 bits or higher. Internally equipped with a GPS clock or timing device, it ensures that the time synchronization error between different terminals does not exceed 1ms.

[0077] Transient electrical parameters refer to rapidly changing electrical quantities measured at various points in the distribution network during signal propagation. They mainly include: instantaneous voltage waveforms, which are voltage value sequences recorded from high-frequency sampling point sequences; instantaneous current waveforms, which are current sampling sequences at corresponding moments; and phase changes, which are instantaneous phases extracted through Hilbert transform or phase-locked loops. The sampling time interval of these parameters should be compatible with the overall time synchronization accuracy of the system.

[0078] It should be understood that physical environment parameters are a preliminary concept in S1 and will be explained in detail in subsequent steps. Here, they can be understood as measurable non-electrical physical quantities in the geographical location of the distribution network and the operating environment of the equipment. Their acquisition is synchronized with transient electrical parameters.

[0079] In an exemplary implementation scenario of the present invention, assuming diagnostics are being performed in a nominal 380V distribution network, the programmable injection unit begins injecting an initial probe signal sequence at time t=0. This sequence consists of three frequency components: frequency... =105Hz, amplitude =0.5V; frequency =505Hz, amplitude =0.5V; frequency =1005Hz, amplitude =0.5V. Timing coding Defined as: within t∈[0,0.1s] =1, within t∈(0.1s,0.5s) =0, repeating with a period of 0.5s. Under this injection condition, the signal received by the cooperative sensing terminal on the distribution network at its location will exhibit attenuation characteristics. For example, the signal amplitude received by the terminal 100m away from the injection point may attenuate to 80% of its original value, at 500m away to 50%, and at 1000m away to 30%.

[0080] After receiving the injected signal, each collaborative sensing terminal records the received transient electrical parameters at its sampling time. For example, in the sampling at t=0.05s, terminal A records the instantaneous voltage value as... =1.2V, the corresponding value of terminal B is =0.8V. This amplitude difference reflects the signal attenuation and distance effect during propagation. These differences will become the basis for judging connectivity in subsequent cross-verification.

[0081] In one embodiment of the present invention, step S2 includes the following steps:

[0082] The collaborative sensing terminal synchronously collects transient electrical responses and environmental IoT data at its location;

[0083] The collected transient electrical responses are timestamped with the environmental IoT data to generate data units;

[0084] All data units generated by the collaborative sensing terminals are aggregated to form a structured set of original feature features for the graph nodes;

[0085] The transient electrical response includes instantaneous waveforms and phase changes of voltage and current; the environmental IoT data includes infrared temperature readings on the equipment surface and micro-vibration waveforms of the switch box.

[0086] After the initial detection signal sequence is injected and a response is triggered, the collaborative sensing terminals distributed throughout the distribution network will perform synchronous data acquisition and integration, ultimately generating a structured original feature set of the map nodes. This process includes data acquisition, data fusion, and dataset construction.

[0087] First, when the initial detection signal sequence arrives at each collaborative sensing terminal, the voltage detection circuit inside the terminal identifies preset signal characteristics. Using this as a trigger event, a synchronous data acquisition process is initiated. During this process, the high-speed data acquisition module within the terminal begins recording the transient electrical response at its location. Specifically, the module typically records the line voltage and the current flowing through the equipment or line continuously at a sampling frequency of no less than 20kHz, forming a digital sequence composed of discrete sampling points. This sequence depicts the instantaneous voltage and current waveforms over a period of time after the signal arrives. Simultaneously, the terminal's internal processor analyzes this waveform data in real time, extracting the instantaneous phase information of the signal through digital signal processing algorithms, thereby obtaining the phase changes of the voltage and current.

[0088] At the same time as acquiring transient electrical responses, the collaborative sensing terminal simultaneously collects environmental IoT data through its integrated non-electrical sensors: the infrared temperature sensor on the terminal is aimed at preset key equipment monitoring points, such as switch contacts or cable joints, to collect and record infrared temperature readings of the equipment surface at these points; simultaneously, the microelectromechanical system accelerometer installed on the switch box or transformer shell captures the minute mechanical vibrations generated by the equipment under signal excitation, recording their amplitude and frequency to form micro-vibration waveforms. These two types of non-electrical quantity data are strictly time-synchronized with the acquisition of transient electrical responses, sharing the same time reference.

[0089] Finally, after completing synchronous data acquisition, each collaborative sensing terminal will structurally bundle all the acquired data. The core of this process is timestamp alignment, which means that based on the moment the acquisition was triggered, the voltage waveform, current waveform, and phase change data from the transient electrical response, along with the infrared temperature readings and micro-vibration waveform data from the environmental IoT data, are packaged into a unified data packet. This data packet is defined as a data unit, which contains comprehensive response information of that node at a specific moment. After all the activated collaborative sensing terminals in the distribution network upload their respective generated data units to the central data server, the collection of these data units from different nodes constitutes a structured original feature set of the graph nodes for subsequent analysis.

[0090] It should be noted that transient electrical response is the change in electrical quantity triggered by the initial detection signal sequence, recorded by the collaborative sensing terminal. The instantaneous waveforms of voltage and current refer to a series of discrete data points of voltage and current amplitude acquired at a high sampling rate within a set sampling window; their data structure is a one-dimensional time series array. Phase change refers to the sequence of changes in the phase angle of voltage or current relative to a reference standard over time, obtained through analysis and calculation of the instantaneous waveforms.

[0091] Non-electrical quantities are parameters reflecting the physical state of equipment, synchronously collected by the collaborative sensing terminal. Among them, the infrared temperature reading on the equipment surface is a scalar value of the temperature at a specified point, in °C. Its setting is based on the equipment operating procedures; the temperature rise of the equipment joints during normal operation should not exceed a specific threshold, such as 40 °C. The micro-vibration waveform of the switch box is a vibration signal collected by an accelerometer. Its data structure is also a one-dimensional time series array, reflecting the mechanical response characteristics of the equipment under electrical disturbances.

[0092] Environmental IoT data is a collection of the aforementioned non-electrical quantities. Its data structure consists of records containing multiple key-value pairs, such as fields for temperature and vibration.

[0093] Timestamp alignment is a crucial operation for ensuring the comparability of multi-source data. Specifically, all collaborative sensing terminals maintain synchronization with a unified standard time via GPS or a network time protocol, with the error controlled within 1ms. When data acquisition begins, all data points are marked with their corresponding timestamps.

[0094] A data unit is a structured combination of all the data generated by a single collaborative sensing terminal during acquisition. Its data structure can be an object containing multiple fields, such as {node ID, timestamp, voltage waveform array, current waveform array, phase change array, infrared temperature value, vibration waveform array}.

[0095] The structured feature set of the network nodes is a collection of data units generated by all collaborative sensing terminals in the entire distribution network. It is a two-dimensional data structure, which can be understood as a list, where each element is a data unit. This feature set serves as the direct input for subsequent topology diagnostic analysis.

[0096] For example, assume the initial detection signal sequence is injected at time t=0. Cooperative sensing terminal A, located in the middle of the feeder, and cooperative sensing terminal B, located at the end of the line, are simultaneously activated.

[0097] Collaborative sensing terminal A detects a signal and triggers data acquisition at t=0.015s. Assume a 500ms transient electrical response is acquired, where the instantaneous voltage waveform displays a superimposed waveform with an amplitude of 0.4V and frequencies including 105Hz and 505Hz. Simultaneously, it acquires an infrared temperature reading of 35.2℃ on the surface of the monitored switch contact device and records a weak micro-vibration waveform. Terminal A timestamps these data and bundles them to generate data unit A, whose content is {Node ID: A, Timestamp: t=0.015s, Voltage Waveform: [...], Current Waveform: [...], Phase Change: [...], Infrared Temperature: 35.2℃, Vibration Waveform: [...]}.

[0098] Because the collaborative sensing terminal B is farther from the injection point, it is assumed that the signal was detected and triggered for acquisition only at t=0.028s. It acquires a transient electrical response of 500ms, but due to signal attenuation, the instantaneous voltage waveform amplitude is 0.25V. Simultaneously, it acquires the infrared temperature reading of 34.8℃ on the surface of the monitored cable joint equipment and records the corresponding micro-vibration waveform. Terminal B also generates data unit B, whose content is {Node ID: B, Timestamp: t=0.028s, Voltage Waveform: [...], Current Waveform: [...], Phase Change: [...], Infrared Temperature: 34.8, Vibration Waveform: [...]}.

[0099] By combining the data units generated by terminal A, terminal B, and all other terminals in the network, a structured original feature set of the graph nodes is formed.

[0100] In one embodiment of the present invention, step S3 includes the following steps:

[0101] For two connected nodes in the theoretical topology, the corresponding transient electrical responses are extracted from the original feature set of the structured graph nodes. The attenuation and time delay during signal propagation are calculated to form a preliminary criterion for electrical connectivity. Simultaneously, environmental IoT data of the two connected nodes are extracted and compared to form an auxiliary criterion for physical connectivity. When the preliminary criterion and the auxiliary criterion corroborate each other, the connection is enhanced; otherwise, it is downweighted to generate an enhanced initial topology connectivity feature map. The auxiliary criterion includes at least the following:

[0102] Method 1: Extract the infrared temperature readings of the two associated nodes from the environmental IoT data and calculate the temperature difference between them; compare the temperature difference with the preset abnormal temperature difference threshold to determine whether there is an abnormal temperature difference, and use the determination result as an auxiliary criterion for the physical connection status.

[0103] Method 2: Extract the micro-vibration waveform data generated by the two associated nodes under the excitation of the detection signal from the environmental IoT data. Since the detection signal is a synchronous excitation source, the vibration responses of physically connected devices, like the two ends of a switch, should be highly synchronized in time. Therefore, calculate the normalized cross-correlation coefficient of the two micro-vibration waveforms within the signal response frequency band. Compare this cross-correlation coefficient with a preset vibration synchronization threshold. If the coefficient is lower than the threshold, it is determined that there is vibration loss of synchronization, indicating that the physical connection may be loose or there is mechanical isolation. This determination result is used as an auxiliary criterion for the physical connection status.

[0104] Specifically, based on the original feature set of structured graph nodes, a dual verification mechanism is used to assign confidence weights to each connection in the theoretical topology of the distribution network, ultimately generating an enhanced initial topology connection feature map. This is mainly accomplished through three parts: traversing theoretical connections, performing cross-verification, and adjusting weights.

[0105] The first part involves the system loading a pre-stored theoretical topology diagram of the power distribution network. This topology diagram, based on design drawings or historical operation and maintenance data, defines all device nodes in the network and their expected physical connections. Next, the system iterates through each connection in the theoretical topology diagram, processing a pair of theoretically directly connected nodes at a time, such as node A and node B.

[0106] In the second part, for selected nodes A and B, the system extracts their respective data units from the original feature set of the structured graph nodes. From these two data units, transient electrical response data is extracted to form a preliminary criterion for electrical connectivity. Specifically, the timestamps in the data units of nodes A and B are compared, and the time difference between them is calculated. This time difference is the signal delay of the initial probe signal sequence propagating between the two associated nodes. Simultaneously, the system analyzes the instantaneous voltage or current waveforms recorded by the two associated nodes, extracts the amplitude of the characteristic frequency components, and calculates the signal attenuation by comparing the ratio of the two amplitudes. If the calculated delay and attenuation conform to the propagation patterns on power lines—for example, the delay is proportional to the theoretical line length, and the attenuation is within a preset reasonable range—then the preliminary criterion for electrical connectivity of this connection is considered valid.

[0107] While performing electrical criterion analysis, the system simultaneously extracts and compares environmental IoT data from the corresponding data units of node A and node B to construct auxiliary criteria for physical connection status. Specifically, the system reads the infrared temperature readings of the two associated nodes and calculates their temperature difference. If the two associated nodes are physically closely connected components, such as the two ends of a switch, their temperatures should be very close under normal operating conditions, with the temperature difference within a small range. If the temperature difference exceeds this range, it may indicate a physical abnormality such as poor contact at the connection point.

[0108] The third part integrates the verification results from the two aspects mentioned above and dynamically assigns credibility values ​​to the connections. The assignment logic is to preset an initial credibility weight for each connection in the theoretical topology. When the preliminary criterion for electrical connectivity is valid, and the auxiliary criterion for physical connectivity also shows normality, i.e., the two verification results corroborate each other, the system enhances the credibility of the connection, for example, by increasing its weight by a fixed value. Conversely, if one or more of the electrical criteria and the physical auxiliary criteria are abnormal, such as excessive signal delay or abnormal temperature difference, the system determines that the two cannot corroborate each other and de-weights the connection, i.e., reduces its credibility weight. After traversing and processing all connections in the theoretical topology, a network graph containing all connections and their dynamically adjusted credibility-enhancing weights is obtained. This graph is defined as the initial topology connection feature graph with enhanced credibility.

[0109] The formula for updating the weights can be expressed as:

[0110]

[0111] in, Increase the confidence weight of the updated connection. The initial weight for this connection is usually set to 1, representing that it is theoretically reliable. This is an adjustment factor for the electrical connectivity criterion. Adjustment coefficients for auxiliary criteria of physical connection. The setting is based on the fact that when the time delay and attenuation conform to the physical model, A value greater than 1, such as 1.2, indicates enhancement; otherwise, A value less than 1, such as 0.5, represents a weakening effect. The setting is based on the fact that when physical parameters such as temperature difference are within the normal range, The value is greater than 1, for example, 1.1; when there is an anomaly, the value is less than 1, for example, 0.6. These coefficient values ​​can be optimized based on statistical analysis of n sets of field measurement data containing normal and abnormal connections.

[0112] It should be noted that cross-verification is a logical processing procedure that confirms the actual state of the theoretical connection through dual checks of electrical and physical dimensions. The preliminary criterion for electrical connectivity is the logical output of the cross-verification, which determines whether the connection is true or false, based on whether the signal propagation delay and attenuation meet expectations.

[0113] Signal attenuation and time delay during propagation are two key physical quantities that are calculated. Attenuation is quantified by comparing the amplitude ratio of the signals received by two associated nodes, while time delay is determined by comparing the timestamp difference of the data collected by the two associated nodes.

[0114] The auxiliary criterion for physical connection status is another logical output in cross-verification, used to confirm whether the electrical connection is reliable at the physical level. Its main basis is whether the environmental parameters of the equipment operation are normal. Abnormal temperature difference is a key physical auxiliary criterion indicator, referring to the absolute value of the difference between the infrared temperature readings of two connected nodes exceeding a preset normal threshold. This normal threshold is set according to the equipment type and operating standards; for example, for low-voltage switch contacts, the temperature difference threshold can be set to 5°C based on the equipment's own parameters.

[0115] It should be understood that the enhancement assignment and the weight reduction processing are two opposite operations that adjust the connection confidence weights based on the cross-verification results; the initial topology connection feature map with enhanced confidence is the final output of this step. It is in the form of a data structure and can be represented by an adjacency matrix or an adjacency list, which stores the connection relationships between all nodes in the distribution network, and each connection is accompanied by a confidence weight value after enhancement or weight reduction processing.

[0116] For example, continuing from the preceding steps, the system extracts data units A and B from the original feature set of the structured graph nodes. A's timestamp is t=0.015s, and its signal amplitude is 0.4V; B's timestamp is t=0.028s, and its signal amplitude is 0.25V. The calculated signal delay is 0.028s - 0.015s = 0.013s. Assuming the theoretical line length between A and B is 300 meters, the expected signal propagation delay is approximately 0.01s, and the attenuated amplitude is approximately 0.2V. Since the calculated delay of 0.013s and the attenuated amplitude of 0.25V are both within the reasonable fluctuation range of the expected values, the preliminary criterion for electrical connectivity in this example is valid. It is 1.2.

[0117] Next, the system extracts environmental IoT data. The infrared thermometer reading for device A is 35.2℃, and for device B it is 34.8℃. The temperature difference between the two is 0.4℃, far less than the abnormal temperature difference threshold of 5℃. Therefore, the auxiliary criterion for the physical connection status is normal, and the result is taken as... It is 1.1.

[0118] Because the electrical criteria and the physical auxiliary criteria corroborate each other, the system assigns enhanced values ​​to the connection between A and B. Assume initial weights. If the value is 1.0, then the updated confidence enhancement weight =1.0×1.2×1.1=1.32.

[0119] Suppose that when verifying the theoretical connection between CDs in the network, the signal delay is found to be much greater than expected, and the electrical criterion is invalid. The value is 0.5. At this point, even if the temperature difference is normal, that is... It remains at 1.1, but its final weight may be reduced to... =1.0×0.5×1.1=0.55.

[0120] After calculating the weights of all theoretical connections, an enhanced initial topology connection feature map is obtained, containing all connections such as (A,B,1.32), (C,D,0.55), and their weights.

[0121] In one embodiment of the present invention, step S4 includes the following steps:

[0122] The graph neural network model compares and analyzes the input enhanced initial topology connectivity feature map with a pre-set healthy topology feature library.

[0123] The graph neural network model infers and outputs the probability value of the graph-real difference at each connection in the network.

[0124] Visualize and render the probability values ​​of the graph-to-real-world differences of all connections to form a topology diagnostic probability map.

[0125] Specifically, based on the obtained enhanced initial topology connection feature map, the regions in the distribution network where the theoretical topology is most likely to differ from the actual situation are identified and presented in a visual manner.

[0126] First, the enhanced initial topology connectivity feature map is input into a pre-trained graph neural network model. In this embodiment, the graph neural network can be a graph attention network, i.e., GAT, with the following network structure design example: The model contains three graph attention layers, each with eight attention heads, implementing a multi-head attention mechanism to capture the importance of neighboring nodes from different perspectives. The output feature dimension of each attention layer is 64 dimensions. Exponential linear units (ELU) are used as the activation function between layers. During training, this model is treated as a binary classification task, classifying each edge in the graph, such as classifying normal connections from abnormal connections, using the cross-entropy loss function as the loss function. The Adam optimizer is used for training, with an initial learning rate of 0.001, a batch size of 32, and a total of 100 iterations. After the graph neural network model has been pre-trained on a large amount of labeled normal distribution network topology data, this pre-stored historical normal data constitutes a healthy topology feature library. Finally, each node and its features in the enhanced initial topology connection feature graph, along with each connection and its enhancement weights, are converted into a digital matrix format that the model can process, and then fed into the first layer of the model.

[0127] After receiving input data, the graph neural network model begins inference analysis. At each layer of the model, it aggregates information about its neighbors and connection weights for each node in the network. In this way, the model considers not only the weight of a single connection but also the overall performance of that connection within the local network environment. This process proceeds layer by layer, with the model performing a deep comparison between the input weight features and normal patterns learned from a healthy topology feature library. If the confidence weight of a connection, or the local pattern formed by it and its surrounding connections, deviates from the common range in healthy samples, the model determines that there is a suspected anomaly. After calculations across all layers, the final layer of the model outputs a value between 0 and 1 for each connection and each key switch node in the network. This value is defined as the graph-real difference probability value, which quantifies the likelihood of a topological inconsistency problem at that location.

[0128] The system visualizes all probability values ​​output by the graph neural network model, generating an intuitive topology diagnostic probability map. Specifically, the system draws the overall topology of the distribution network on the interface, and then colors the graph-to-physical discrepancy probability values ​​for each connection and node. A heatmap coloring scheme is typically used, mapping low-probability values ​​to cool colors such as blue and green, and high-probability values ​​to warm colors such as orange and red. This results in an image where the areas with the highest graph-to-physical discrepancy probability—i.e., abnormal areas—are highlighted in striking red or orange, providing maintenance personnel with clear guidance for troubleshooting. This heatmap, which intuitively displays the probability of topology anomalies across the entire network, is the topology diagnostic probability map.

[0129] Among them, the abnormal area refers to the set of line connections or switch nodes that are highlighted by color rendering on the topology diagnostic probability map and whose map-to-real-world difference probability value exceeds a preset threshold.

[0130] Graph neural network models are computational models specifically designed for processing and analyzing graph-structured data. They learn and identify complex connection patterns in a graph by simulating the propagation of information between nodes. These models are trained to distinguish between normal and abnormal topological connection patterns.

[0131] The Healthy Topology Feature Library is a benchmark dataset used to train graph neural network models. Its construction process involves repeatedly executing steps S1 to S3 on multiple distribution networks with known correct topologies, collecting enhanced initial topology connection feature maps generated under various normal loads and operating conditions. The collection of these maps constitutes the Healthy Topology Feature Library, representing various manifestations of normal topologies.

[0132] In one embodiment of the present invention, the health topology feature library construction process follows the principles outlined below:

[0133] Data Sources: Data is sourced from power grid companies in different regions of China, such as North China, East China, and South China, ensuring diversity in geographical and climatic conditions. Data includes GIS system maps of the power grid companies, the latest as-built drawings, and historical operational data of distribution automation systems for at least one year.

[0134] Sample Size and Representativeness: This feature library contains topology data for multiple distribution network feeders, including over 2000 typical 10kV distribution network feeder topologies. The correctness of these feeder topologies was verified through a three-pronged approach: comparing the latest as-built drawings with on-site maintenance personnel's inspection records; verifying lines that have consistently performed normally in line loss calculations, fault location, and other business operations; and conducting manual on-site verification for lines with questionable issues.

[0135] Operating Condition Coverage: To construct the feature library, steps S1 to S3 of this invention were repeatedly executed on these distribution networks with confirmed correct topologies, collecting enhanced topology connection feature maps under different operating conditions. The covered operating conditions include: summer peak load, winter off-peak load, spring and autumn stable load, and various typical operating modes simulating different tie switch and sectional switch switching operations. Multiple feature maps marked as healthy were generated and stored, exceeding 100,000 images; thus forming a healthy topology feature library; this feature library represents various manifestations of normal topologies.

[0136] The graph-actual discrepancy probability value is a value between 0 and 1 output by the graph neural network model for each connection or node in the network. It quantitatively represents how likely it is that the actual state of that element is inconsistent with the state depicted in the theoretical topology graph. A value close to 1 means that there may be a discrepancy, while a value close to 0 means that it is consistent with the theoretical topology.

[0137] Topology diagnostic probability maps are a form of data visualization that combines the topology of the distribution network with the probability values ​​of map-to-actual discrepancies, and visually displays potential anomaly areas across the entire network in the form of heat maps.

[0138] For example, the initial topological connectivity feature map that has been enhanced in the above example is input into a graph neural network model. It is known that the graph contains connection (A, B) with a weight of 1.32 and connection (C, D) with a weight of 0.55.

[0139] When analyzing data, the graph neural network model first consults its knowledge from a healthy topological feature database. If the database shows that for the (A,B) connection, the weights in a healthy state are typically distributed between 1.1 and 1.4, and since 1.32 falls within this range, the model determines that the (A,B) connection is normal. Therefore, the model outputs a very low graph-to-real difference probability value for the (A,B) connection, for example, 0.03.

[0140] For connection (C,D), if the healthy topology feature library shows its normal weight range should be 1.0 to 1.3, then an input weight of 0.55, below this range, is considered an abnormal signal. The model determines that the (C,D) connection may have a problem, such as a broken line or an unclosed switch. Therefore, the model outputs a high graph-to-real-world difference probability value for connection (C,D), for example, 0.95. Simultaneously, the model also assigns high difference probability values, for example, 0.88, to nodes C and D associated with this abnormal connection.

[0141] Finally, on the generated topology diagnostic probability map, the connection (A, B) will be displayed in blue or green, while the connection (C, D) and nodes C and D will be displayed in bright red. This red area is the identified anomalous area, representing the location that needs to be focused on for further detection and diagnosis.

[0142] In one embodiment of the present invention, step S5 includes the following steps:

[0143] Analyze the topological diagnostic probability map, identify connections whose map-to-real discrepancy probability values ​​exceed a preset probability threshold, and define them as target diagnostic regions.

[0144] Based on the electrical characteristics of the target diagnostic area, select signal frequency and timing combinations from a preset waveform library that match the potential fault types in that area;

[0145] The selected signal frequency and timing combination are used to construct a secondary detection signal sequence.

[0146] First, the topology diagnostic probability map is analyzed. This analysis process is essentially image processing or data filtering. By scanning the entire map, network elements with high probability values ​​for map-to-real-world discrepancies are identified. The system sets a probability threshold, such as 0.8, and then identifies all connections or switching devices with probability values ​​exceeding this threshold, forming a set. The network area represented by this set is defined as the target diagnostic area, which delineates the specific lines or devices in the entire distribution network most likely to have problems.

[0147] Then, after identifying the target diagnostic area, a second detection is performed using a tailored secondary detection signal. The steps include: querying a pre-established waveform library for electrical characteristic parameters related to the target diagnostic area; for example, if the target is a cable line, the system queries whether the cable type is overhead or underground, and its approximate length. Based on these electrical characteristics, the system dynamically selects the signal combination most sensitive to the potential fault type in that area from the waveform library. For example, for a switch suspected of having poor contact, a high-frequency pulse signal is selected due to its high sensitivity to changes in contact resistance; while for a cable suspected of having a grounding fault, a signal containing specific low-frequency components might be selected because such signals have more unique attenuation characteristics when propagating in the grounding path. Therefore, the system selects one or more signal frequency and timing combinations with the best detection effect.

[0148] Finally, the system programs the selected signal frequency and timing combination to generate a completely new electrical signal sequence. This sequence is designed to probe only the identified target diagnostic region, with its energy and characteristics optimized for this specific task. This specially customized signal sequence is defined as the secondary probe signal sequence. It differs from the initial probe signal sequence in the first step; the former is broad-spectrum and universally applicable, while the latter is narrow-spectrum and highly targeted. After generation, the secondary probe signal sequence will be stored and prepared for injection into the power grid by the programmable injection unit.

[0149] The target diagnostic area is a set of specific lines or switching equipment identified in the topology diagnostic probability map whose map-to-actual discrepancy probability value exceeds a preset high threshold. It delineates the geographical and electrical scope of the secondary detection. Electrical characteristics are data describing the physical and electrical properties of power equipment or lines, such as cable type, material, cross-sectional area, nominal length, and the rated current and breaking capacity of switches. This information is typically stored in the power grid's asset management database.

[0150] The waveform library is a pre-built database that stores waveform templates of various detection signals with different characteristics. Each template is associated with its most applicable diagnostic scenario and target fault type. For example, a template may contain a pulse train with a frequency of 5 kHz and a pulse width of 100 μs, labeled as suitable for detecting minute contact defects in switch contacts. The waveform library is built based on extensive laboratory simulations and analysis of field measurement data from n sets of different fault types.

[0151] Signal frequency and timing combination refers to the specific parameter configuration of one or more signal templates selected from the waveform library, including the frequency components of the signal, the amplitude of each frequency, the pulse width of the signal, the repetition period and other time series characteristics.

[0152] The secondary detection signal sequence is a time-varying electrical signal sequence dynamically selected and constructed from a waveform library based on the characteristics of the target diagnostic region and the potential fault type. Compared with the initial signal, it may have a higher frequency, narrower pulse width, or more complex timing coding, aiming to stimulate the response characteristics of the target region.

[0153] For example, suppose the topology diagnostic probability map is analyzed and the discrepancy between the map and reality for connection (C,D) is found to be 0.95, far exceeding the threshold of 0.8. Therefore, the system identifies connection (C,D) and the switching equipment at both ends as the target diagnostic area. Next, the system queries the database and learns that connection (C,D) is a 50m long buried copper core cable, and its most likely potential fault type is high impedance caused by joint oxidation or localized line damage.

[0154] Based on this information, the system queries the waveform library. The library suggests that for this type of high-impedance fault, a combination of high-frequency square waves and low-frequency sine waves is most sensitive in terms of frequency and timing. For example, the system selects a 10kHz square wave signal with an amplitude of 1% of the injected voltage to detect resistance changes; simultaneously, a 200Hz sine wave signal with an amplitude of 0.5% of the injected voltage is superimposed to assess the integrity of the overall circuit. The timing of this combination is designed with a detection window of 100ms.

[0155] Finally, a secondary detection signal sequence is generated based on the selected signal frequency and timing combination. In the time domain, this sequence exhibits a rapidly oscillating high-frequency square wave superimposed on a slowly varying sinusoidal fundamental wave. This sequence is packaged and labeled for high-impedance fault detection at connection (C,D), ready to be injected into the grid in the next step to obtain a stronger and clearer response signal in the targeted diagnostic area compared to the initial grid-wide detection.

[0156] In one embodiment of the present invention, step S6 includes the following steps:

[0157] Generate a focused response dataset, including the following steps:

[0158] The secondary detection signal sequence is then injected back into the distribution network.

[0159] The energy of the secondary detection signal sequence will be concentrated on the target diagnostic area, triggering a response on the collaborative sensing terminal in that area;

[0160] The response data of each terminal within the target diagnostic area is collected again to form a focused response dataset.

[0161] The secondary detection signal sequence is used to excite and collect response data in the target area, providing a basis for final fault confirmation. First, the system sends a command to the programmable injection unit deployed in the distribution network to inject the aforementioned secondary detection signal sequence. After receiving the command and signal sequence data, the programmable injection unit will inject the electrical signal again into the same injection point in the distribution network according to the voltage, frequency, and timing characteristics defined in the sequence. The process is similar to the first injection operation, but the injected signal content has changed from a general detection signal to a targeted, focused detection signal.

[0162] When the secondary detection signal sequence is injected into the power grid, its propagation behavior differs from the initial signal. Because the frequency and timing combination of this signal are specifically optimized for the electrical characteristics of the target diagnostic area, its signal energy dissipates less into non-target areas during propagation. Instead, it propagates along the physical path leading to the target diagnostic area, concentrating most of the energy on that area. This energy focusing effect will elicit stronger and more distinctive electrical and physical responses at the collaborative sensing terminals within that area compared to the first detection. For example, the signal voltage amplitude may be several orders of magnitude higher than the first time, or it may be able to induce sufficiently identifiable micro-vibrations in the equipment.

[0163] Only the collaborative sensing terminals located within the target diagnostic area will be activated to perform data acquisition tasks. After receiving the focused detection signal, these terminals will synchronously acquire transient electrical responses and environmental IoT data at their location again. Due to the enhanced strength and characteristics of the excitation signal, the signal-to-noise ratio of the acquired data will be significantly improved compared to the first acquisition, meaning that useful signal features are far more prominent than background noise. After acquisition, the response data from each terminal within the target diagnostic area are integrated to form a dataset with superior data quality and feature recognition compared to the second full-network data. This dataset is defined as the focused response dataset.

[0164] The focused response dataset is the core output of this step. Its data structure is similar to the original feature set of structured graph nodes, both being collections of multiple data units. However, its key features are high signal-to-noise ratio (SNR) and focus. Focus means that this dataset only contains data from collaborative sensing terminals within the target diagnostic area. High SNR refers to the very high power ratio of the effective signal component to the noise component in the data. This is because the energy of the excitation signal is concentrated, making the amplitude of the response signal much larger than the inherent electrical noise level of the power grid.

[0165] It should be noted that the concentration of energy on the target diagnostic area is a description of a physical phenomenon. This is based on the selection of the signal frequency. By choosing a frequency with minimal transmission loss along the path to the target area, or one that most easily interacts with potential fault characteristics, the signal energy naturally converges towards that area during propagation. Stronger amplitude and more pronounced electrical and physical responses are a result of the energy focusing effect. Stronger amplitude refers to a larger change in physical quantities such as voltage, current, or vibration measured by the co-sensing terminal; more pronounced characteristics mean a clearer waveform in the response signal, making it easier to identify specific fault-related patterns, such as clear reflection peaks or resonances at specific frequencies.

[0166] For example, the programmable injection unit receives a command to inject a secondary detection signal sequence designed for high-impedance faults at connection (C,D) into the distribution network. The energy of this signal sequence is highly concentrated and propagated to the area where connection (C,D) is located. Collaborative sensing terminals C and D located in this area receive this focused signal. Assume that in the first detection, the signal amplitude received by terminal C is only 0.3V, and the waveform is mixed with a lot of background noise; while in this focused detection, due to the targeted design of the signal, the signal amplitude received by terminal C reaches 1.8V, which is 6 times that of the first time; the 10kHz square wave component in the signal waveform is clearly distinguishable, and its signal-to-noise ratio improves from about 5dB in the first time to 20dB. This demonstrates a stronger amplitude and more distinct electrical and physical response.

[0167] Meanwhile, due to the enhanced signal energy, if there is indeed high-resistance oxidation at the junction of (C,D), the Joule heating effect generated when this high-frequency, high-current signal flows through will be more pronounced; for example, the infrared temperature reading of the collaborative sensing terminal C may rise from 35°C to 40°C in a short period of time, producing a temperature rise that is difficult to observe in the first detection, which provides key physical evidence for judging the nature of the fault.

[0168] Terminals C and D collect this response data again and package it into new data units. These new data units, when combined, constitute the focused response dataset. The data quality of this dataset is higher than that obtained from the first full-network scan, providing a solid foundation for the next step of final, high-confidence diagnostics.

[0169] In one embodiment of the present invention, step S7 includes the following steps:

[0170] The focused response dataset is repeatedly input into the cross-verification process of step S3 and the diagnostic process of step S4 to re-infer the topological state of the target diagnostic region and obtain the updated map-real difference probability value.

[0171] By comparing the changes in the probability values ​​of the difference between the initial diagnosis and the subsequent reasoning, we can obtain confidence conclusions about the fault points and abnormal states within the target diagnosis area.

[0172] Based on the confidence level conclusion, the root cause of the map-to-surface discrepancy is identified and output, forming the final topology discrepancy location report. The final topology discrepancy location report includes: the fault location causing the map-to-surface discrepancy, the fault type, and the confidence level assessment calculated based on the change in the probability value of the map-to-surface discrepancy.

[0173] Specifically, using the obtained focused response dataset, a decisive, high-confidence confirmation diagnosis is made on the abnormal areas identified in the initial diagnosis, and the final conclusions that can guide operation and maintenance are output.

[0174] First, the system uses the focused response dataset as new input and re-executes the cross-verification process in step S3 and the graph neural network diagnostic process in step S4. The system focuses only on connections and nodes within the target diagnostic region, using this higher-quality dataset to recalculate signal attenuation and latency, and compares it with environmental IoT data. Due to the significantly improved signal-to-noise ratio, the calculated electrical parameters and observed physical parameters are more accurate and stable, making the cross-verification results more reliable. Subsequently, the updated connection weights are input again into the same graph neural network model, which infers the topological state of the target diagnostic region again and outputs updated graph-real-world difference probability values.

[0175] The system eliminates potential uncertainties from the initial diagnosis by comparing the results of two diagnostic tests. The initial network-wide diagnosis may contain a certain probability of false positives due to weak signals or interference, while the focused diagnosis uses signals with more distinct characteristics, resulting in higher reliability. If a connection showed a high probability of difference in the initial diagnosis, and this probability further increases and stabilizes at a high level in the focused diagnosis, the system can arrive at a high-confidence conclusion that the connection has a genuine fault. Conversely, if a high-probability point in the initial diagnosis decreases in the focused diagnosis, it indicates that the initial assessment may have been a false alarm. Through this iterative comparison, fault points and abnormal states can be further confirmed.

[0176] After obtaining a high-confidence diagnostic conclusion, key information from the entire diagnostic process is integrated to form a final topology discrepancy location report. This report identifies the cause of the discrepancy between the map and reality, such as an open-circuit fault at the CD connection point of a certain line. The report will include three core elements: fault location, including device number and geographical coordinates; fault type, such as open circuit, short circuit, high-impedance connection, or incorrect switching state; and confidence assessment, which is a quantitative score of the reliability of the diagnostic conclusion. This score is derived based on the changing trend of the diagnostic probability during the iteration process and the final probability value. This report can be directly sent to maintenance personnel as a direct basis for their on-site inspection and recovery work.

[0177] Iterative diagnosis refers to repeatedly applying the process to its own output in order to obtain more accurate or convergent results. In this scheme, it refers to repeatedly applying the cross-verification process of S3 and the diagnostic process of S4 to the focused response dataset. A high-confidence conclusion refers to a diagnostic result with a high and stable probability value of map-reality difference after confirmation by secondary focused diagnosis. The reliability of this conclusion is higher than that of a conclusion based solely on the initial full-network diagnosis.

[0178] It should be noted that the final topology difference location report is the final output of the entire diagnostic method. It is a structured document or data display interface. Its data structure should contain clearly defined fields, such as fault ID, device location, fault type description, confidence score, diagnosis time, etc., aiming to provide users with clear, accurate and actionable decision support.

[0179] Confidence assessment is a quantitative score of the reliability of a diagnostic conclusion, and its value is usually between 0 and 100%. It is based on the consistency of the probability values ​​of the two diagnoses and the high degree of confidence of the final value. For example, if the initial probability is 0.95 and the probability of diagnosis rises to 0.99 in a second focused diagnosis, a 99% confidence assessment can be given.

[0180] For example, the focused response dataset containing the connection (C,D) is input into S3 and S4 for iteration. During the iterative diagnosis, the signal propagation characteristics between the connection (C,D) are recalculated. For instance, the results show that the propagation delay is infinite and no effective signal is received at terminal D, which constitutes the typical electrical characteristics of an open circuit. Simultaneously, regarding the physical auxiliary criteria, when a slight increase in the temperature of terminal C is observed, but the temperature of terminal D remains unchanged, it further confirms that energy has failed to be conducted to point D.

[0181] The results are then input into the graph neural network model for further inference. Based on the open-path features described above, the model outputs updated graph-real difference probability values ​​for the connection (C,D), such as increasing from 0.95 initially to 0.998.

[0182] By comparing the two diagnoses, the system found that the difference probability of the connection (C,D) further increased from a high level of 0.95 to an extremely high level of 0.998, which ruled out the possibility that the initial diagnosis was misjudged due to noise interference, thus obtaining a high confidence conclusion that there is an open circuit fault in the connection (C,D).

[0183] Finally, a final topological difference localization report is generated based on the conclusions.

[0184] The example report content is as follows:

[0185] Fault location: Feeder No. 3, CD cable connection point under switch cabinet K3;

[0186] Fault type: Open circuit fault;

[0187] Root cause: Diagnostic data shows that the signal has difficulty propagating from point C to point D, indicating a physical connection interruption;

[0188] Confidence level assessment: 99.8%;

[0189] Recommended action: Please send someone immediately to check the integrity of the CD cable connector or the cable itself.

[0190] This report will provide maintenance personnel with clear operational guidelines, well-defined causes, and reliable conclusions.

[0191] See appendix Figure 2 The present invention also proposes a distribution network topology map discrepancy diagnosis system based on graph neural networks, comprising the following modules:

[0192] The initial signal injection module is used to inject an initial detection signal sequence, including timing coding and specific frequency components, into the distribution network to excite multiple collaborative sensing terminals distributed in the distribution network to generate responses.

[0193] The response data acquisition module is used to collect and integrate the responses generated by the collaborative sensing terminal to generate a structured set of original feature sets of the map nodes.

[0194] The topology cross-verification module performs cross-verification based on the original feature set of structured graph nodes, generating an enhanced initial topology connection feature map for the theoretical topology connection of the distribution network.

[0195] The anomaly probability diagnosis module is used to input the initial topological connectivity feature map into a preset graph neural network model, identify and generate a topological diagnosis probability map that marks the abnormal regions;

[0196] The secondary detection sequence generation module generates a secondary detection signal sequence based on the abnormal regions marked in the topological diagnostic probability map.

[0197] The focused response acquisition module is used to inject secondary detection signal sequences into the distribution network and generate focused response datasets in abnormal areas;

[0198] The iterative diagnostic report module performs iterative diagnostics on abnormal regions based on the focused response dataset, generating a final topological difference localization report.

[0199] Each of the modules can be implemented in whole or in part through software, hardware, or a combination thereof. It supports hardware embedded in or independent of the processor in the computer device, and also supports software stored in the memory of the computer device, so that the processor can call and execute the operations corresponding to each of the above modules.

[0200] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for diagnosing discrepancies in distribution network topology based on graph neural networks, characterized in that, Includes the following steps: S1. Inject an initial detection signal sequence including timing coding and a specific frequency component into the distribution network to stimulate multiple collaborative sensing terminals distributed in the distribution network to generate a response; wherein, the specific frequency component refers to a frequency whose frequency is an integer multiple of the power frequency of the distribution network. S2. Collect and integrate the responses generated by the collaborative sensing terminals to generate a structured set of original feature data for the map nodes; S3. Perform cross-verification based on the original feature set of structured graph nodes to generate an enhanced initial topology connection feature map for the theoretical topology connection of the distribution network; S4. Input the initial topological connectivity feature map into the preset graph neural network model to identify and generate a topological diagnostic probability map that marks abnormal regions; S5. Generate a secondary detection signal sequence based on the abnormal regions marked in the topological diagnostic probability map; S6. Inject the secondary detection signal sequence into the distribution network to generate a focused response dataset in the abnormal area; S7. Based on the focused response dataset, perform iterative diagnosis of abnormal regions and generate the final topological difference localization report.

2. The method for diagnosing distribution network topology discrepancies based on graph neural networks according to claim 1, characterized in that, Step S1 includes the following steps: The initial detection signal sequence is injected into the power grid through the injection unit set at the beginning of the distribution network feeder or key connection point; The initial detection signal sequence propagates along the physical path of the power grid, stimulating each cooperative sensing terminal along the line to synchronously acquire the transient electrical parameters and physical environment parameters of the response.

3. The method for diagnosing distribution network topology discrepancies based on graph neural networks according to claim 1, characterized in that, Generating the original feature set of the graph nodes includes the following steps: The collaborative sensing terminal synchronously collects transient electrical responses and environmental IoT data at its location; The collected transient electrical responses are timestamped with the environmental IoT data to generate data units; All data units generated by the collaborative sensing terminals are aggregated to form a structured set of original feature features for the graph nodes; The transient electrical response includes instantaneous waveforms and phase changes of voltage and current; the environmental IoT data includes infrared temperature readings on the equipment surface and micro-vibration waveforms of the switch box.

4. The method for diagnosing distribution network topology discrepancies based on graph neural networks according to claim 1, characterized in that, Generating an enhanced initial topology connectivity feature map includes the following steps: For two connected nodes in the theoretical topology, the corresponding transient electrical response is extracted from the original feature set of the structured graph nodes, the attenuation and time delay during signal propagation are calculated, and a preliminary criterion for electrical connectivity is formed. Simultaneously extract and compare the environmental IoT data of two related nodes to form auxiliary criteria for judging the physical connection status; When the preliminary criteria and the auxiliary criteria corroborate each other, the connection is enhanced and assigned a higher value; otherwise, it is downweighted to generate an enhanced initial topological connection feature map.

5. The method for diagnosing distribution network topology discrepancies based on graph neural networks according to claim 4, characterized in that, The process of forming auxiliary criteria includes the following steps: Extract the infrared temperature readings of two related nodes from the environmental IoT data and calculate the temperature difference between them; The temperature difference is compared with a preset abnormal temperature difference threshold to determine whether an abnormal temperature difference exists, and the determination result is used as an auxiliary criterion for the physical connection status.

6. The method for diagnosing discrepancies in distribution network topology based on graph neural networks according to claim 1, characterized in that, Generating a topological diagnostic probability map includes the following steps: The graph neural network model compares and analyzes the input enhanced initial topology connectivity feature map with a pre-set healthy topology feature library. The graph neural network model infers and outputs the probability value of the graph-real difference at each connection in the network. Visualize and render the probability values ​​of the graph-to-real-world differences of all connections to form a topology diagnostic probability map.

7. The method for diagnosing discrepancies in distribution network topology based on graph neural networks according to claim 1, characterized in that, Generating a secondary detection signal sequence includes the following steps: Analyze the topological diagnostic probability map, identify connections whose map-to-real discrepancy probability values ​​exceed a preset probability threshold, and define them as target diagnostic regions. Based on the electrical characteristics of the target diagnostic area, select signal frequency and timing combinations from a preset waveform library that match the potential fault types in that area; The selected signal frequency and timing are combined to construct a secondary detection signal sequence.

8. The method for diagnosing distribution network topology discrepancies based on graph neural networks according to claim 1, characterized in that, Generate a focused response dataset, including the following steps: The secondary detection signal sequence is then injected back into the distribution network. The energy of the secondary detection signal sequence will be concentrated on the target diagnostic area, stimulating a response on the collaborative sensing terminal in that area; The response data of each terminal within the target diagnostic area is collected again to form a focused response dataset.

9. The method for diagnosing discrepancies in distribution network topology based on graph neural networks according to claim 1, characterized in that, Generate the final topology difference localization report, including the following steps: The focused response dataset is repeatedly input into the cross-verification process of step S3 and the diagnostic process of step S4 to infer the topological state of the target diagnostic region again and obtain the updated map-real difference probability value. By comparing the changes in the probability values ​​of the difference between the initial diagnosis and the subsequent reasoning, we can obtain confidence conclusions about the fault points and abnormal states within the target diagnosis area. Based on the confidence level conclusion, the root cause of the map-to-reals discrepancy is identified and output, forming the final topological discrepancy localization report.

10. The method for diagnosing distribution network topology discrepancies based on graph neural networks according to claim 9, characterized in that, The final topology discrepancy localization report includes: the location of the fault that caused the map-to-surface discrepancy, the fault type, and a confidence assessment calculated based on the change in the probability value of the map-to-surface discrepancy.

Images