Power distribution network fault accurate positioning method and system based on graph convolutional neural network

Abstract

The invention discloses a power distribution network fault accurate positioning method and system based on a graph convolutional neural network, and relates to the technical field of power systems, and the method comprises the steps: deploying monitoring equipment at a power distribution network node; in response to the distributed power supply switching event, generating a dynamic graph structure based on a pre-stored simulation model; taking the dynamic graph structure as a reference to initialize graph convolution kernel parameters, and generating two types of operation parameters based on a communication delay condition; fusing the new energy output prediction data, the electrical quantity monitoring data and the meteorological data to construct a dynamic causal graph; when a fault feature signal is detected, extracting electrical quantity monitoring data, a topological connection relationship and causal reasoning knowledge of the associated node; and constructing a graph convolutional network taking a dynamic graph structure as a network topology, selecting an operation parameter of a corresponding communication delay region as a convolution kernel weight, processing electrical quantity monitoring data, a topological connection relationship and causal reasoning knowledge of associated nodes, and outputting a fault coordinate.

Description

Technical Field

[0001] The present application relates to the field of power system technology, and more specifically to the field of intelligent distribution system technology. Specifically, it relates to a method and system for accurately locating distribution network faults based on graph convolutional neural networks. Background Art

[0002] With the advancement of smart grid construction, the operational stability of distribution networks, as the key hub connecting power systems and users, has attracted considerable attention. In recent years, the massive integration of distributed power sources, energy storage devices, and renewable energy sources has transformed distribution network topologies into complex and diverse meshes. This has significantly increased the dynamism and uncertainty of operational states, and external environmental factors have also posed challenges to the reliable operation of distribution networks.

[0003] Currently, fault location in distribution networks mostly relies on traditional fault indicators. However, these methods face significant limitations given the complex characteristics of new distribution networks. Traditional technologies struggle to adapt to the dynamic system changes brought about by distributed power sources and renewable energy, and fault location models lack applicability. Furthermore, their ability to integrate and analyze diverse data types is limited, resulting in poor location accuracy and real-time performance in complex fault scenarios, impacting fault handling efficiency and power supply reliability.

[0004] To address the above-mentioned problems, no effective solutions have been proposed so far. Summary of the Invention

[0005] The embodiments of the present application provide a method and system for accurately locating distribution network faults based on a graph convolutional neural network to solve the above-mentioned technical problems.

[0006] This application provides a method for accurately locating distribution network faults based on a graph convolutional neural network, including:

[0007] Deploy monitoring equipment at distribution network nodes to obtain topological connection relationships, electrical quantity monitoring data, environmental monitoring data, and new energy output forecast data;

[0008] In response to distributed power supply switching events, a dynamic graph structure is generated based on pre-stored simulation models;

[0009] Initializing graph convolution kernel parameters based on the dynamic graph structure and generating two types of operation parameters based on communication delay;

[0010] Integrating the new energy output forecast data, the electrical quantity monitoring data and the meteorological data to construct a dynamic cause-and-effect graph;

[0011] When a fault characteristic signal is detected, the electrical quantity monitoring data, topological connection relationship and causal reasoning knowledge of the associated nodes are extracted;

[0012] A graph convolution network is constructed with the dynamic graph structure as the network topology, and the operation parameters corresponding to the communication delay area are selected as the convolution kernel weight to process the electrical quantity monitoring data, the topological connection relationship and the causal reasoning knowledge of the associated nodes and output the fault coordinates.

[0013] Further, the dynamic graph structure is generated based on the pre-stored simulation model in response to the distributed power supply switching event, including:

[0014] The electrical coupling strength matrix between the nodes of the power distribution network is calculated, and the power distribution network topology network is divided into multiple dynamic autonomous areas according to the electrical coupling strength matrix;

[0015] A topological connection variation strategy is randomly generated in each dynamic autonomous area, and the regional electrical stability is verified through power flow calculation;

[0016] An reinforcement learning strategy is adopted to minimize the node voltage deviation as the target to train the agent to select the optimal topological combination in the topological connection variation strategy;

[0017] The topological connection characteristics of the high electrical stability area are migrated to the low electrical stability area; wherein the voltage deviation rate of the high electrical stability area is less than the deviation rate threshold, and the voltage deviation rate of the low electrical stability area is greater than the deviation rate threshold;

[0018] The line power constraint and voltage fluctuation range are verified, and the dynamic graph structure meeting the safe operation standard of the power distribution network is output.

[0019] Further, the graph convolution kernel parameters are initialized based on the dynamic graph structure, and two types of operation parameters are generated based on the communication delay, including:

[0020] The node impedance parameters in the dynamic graph structure are converted into a graph edge weight matrix as the initialization basis of the graph convolution kernel parameters;

[0021] The dynamic autonomous area with a communication delay lower than the delay threshold is updated in real time, including: analyzing the real-time gradient change of the electrical quantity monitoring data in the area, generating a convolution kernel parameter correction vector; aggregating the historical operation feature trend of the area to generate a parameter offset compensation amount;

[0022] For the dynamic autonomous area with a communication delay higher than the delay threshold, the prediction generation is performed, including: based on the line impedance parameters of the area, the load distribution characteristics are backstepped; combined with the historical switching frequency record of the distributed power supply in the area, the load change trend is predicted; the prediction result is mapped to the convolution kernel frequency domain response through orthogonal frequency domain transform;

[0023] Adaptive frequency domain filtering of key equipment of the power distribution network is performed to generate two types of operation parameters of real-time update type and delay tolerant type.

[0024] Further, the adaptive frequency domain filtering of the power distribution network key equipment priority includes:

[0025] Identifying key equipment nodes in the power distribution network, including distributed power supply access points and load center connection points with a capacity greater than a set capacity threshold;

[0026] Calculating the electrical influence factor of the key equipment node, which is determined by the equipment capacity weight, the topological centrality and the voltage sensitivity;

[0027] Constructing a frequency domain filtering mask matrix, and retaining the weight of the frequency domain component associated with the key equipment to 1.5 times;

[0028] Performing amplitude attenuation processing on the frequency domain component of the non-key area, and the attenuation coefficient is negatively related to the electrical influence factor;

[0029] Reconstructing the time domain convolution kernel parameters by inverse transformation to generate differentiated operation parameters that adapt to the operation characteristics of the power distribution network.

[0030] Further, the dynamic causal diagram is constructed by fusing the new energy output prediction data, the electrical quantity monitoring data and the meteorological data, including:

[0031] Aligning the collection time sequence of the new energy output prediction data and the electrical quantity monitoring data;

[0032] Determining the waveform correlation strength of lightning strike events and voltage sag by wavelet coherence analysis method;

[0033] Constructing a three-channel feature fusion network for processing data, including: an electrical quantity channel for extracting high-frequency mutation features of voltage and current waveforms; a meteorological channel for fusing spatial distribution data of temperature and humidity sensors; and a new energy channel for analyzing photovoltaic / wind power prediction deviation rate;

[0034] Training a Bayesian inference network to output a fault causal probability matrix, dynamically updating the causal weight of short-circuit fault and insulation aging; establishing a fault path reverse tracking mechanism, when a ground fault is detected, scanning the upstream equipment causal chain to generate a dynamic causal graph with weight allocation.

[0035] Further, the establishment of the fault path reverse tracking mechanism includes:

[0036] Defining a fault propagation depth threshold, and starting multi-level reverse tracking when a ground fault is detected;

[0037] First-level tracking: locating the upstream equipment directly connected to the fault point and calculating the fault current contribution degree;

[0038] Second-level tracking: scanning the associated equipment within two hops of the topological distance, and evaluating the fault correlation of the equipment insulation aging index;

[0039] Level 3 tracking: Analyze power supply points where the output fluctuation rate of renewable energy exceeds the fluctuation rate threshold, and calculate the probability of fault induction due to power mutation.

[0040] Furthermore, when a fault characteristic signal is detected, the electrical quantity monitoring data, topological connection relationship and causal reasoning knowledge of the associated nodes are extracted, including:

[0041] Activate dedicated analysis channels based on fault feature types, including: activating the electromagnetic transient analysis channel to extract current harmonic distortion features when the voltage drop exceeds the drop threshold; activating the thermal distribution calculation channel to generate a thermal stress distribution diagram for the equipment when the temperature rise exceeds the temperature change threshold;

[0042] The improved Dijkstra algorithm is used to search for associated nodes within three hops of the fault source under the power supply radius constraint;

[0043] extracting causal reasoning knowledge whose causal weight value is greater than a weight threshold from the dynamic causal graph;

[0044] Apply the causal probability propagation model to screen causal nodes whose correlation strength exceeds the correlation strength threshold;

[0045] Convert electrical quantity monitoring data into wavelet packet energy spectrum features;

[0046] Encode the topological connectivity relationship as the spectral feature vector of the node adjacency relationship;

[0047] The causal weight values ​​in the dynamic causal graph are embedded in the edge attribute matrix of the graph convolutional network, so that the causal knowledge participates in the message passing calculation.

[0048] Furthermore, a graph convolutional network with the dynamic graph structure as the network topology is constructed, and the operation parameters corresponding to the communication delay area are selected as the convolution kernel weights. The electrical quantity monitoring data, topological connection relationship and causal reasoning knowledge of the associated nodes are processed and the fault coordinates are output, including:

[0049] A hierarchical convolutional architecture is configured to adapt to the radial topology of the distribution network, including: implementing a third-order neighborhood aggregation operation on the feeder trunk; performing a first-order neighborhood convolution on the branch lines;

[0050] Enhance fault feature propagation through causal weight values ​​in the edge attribute matrix;

[0051] Build a three-dimensional fault analysis engine, including: longitude dimension: using the reverse graph diffusion algorithm to generate fault probability distribution; latitude dimension: locating distance based on line impedance parameters and fault recording data; altitude dimension: identifying fault phase based on zero-sequence current phase angle difference;

[0052] Implementing parallel parameter loading and pipeline calculation on FPGA chip;

[0053] The output includes line number, tower geographical coordinates and fault coordinates of the fault phase.

[0054] Furthermore, the construction of the three-dimensional fault analysis engine includes:

[0055] When the fault probability distribution heat map shows high-probability points in multiple regions, the collaborative positioning mechanism is activated;

[0056] Identify the power supply partition that contributes most to the fault current;

[0057] fusing voltage phasor measurement unit data of adjacent zones in the longitude dimension to modify the fault probability distribution;

[0058] In the latitude dimension, a historical similar fault case library is called to optimize the distance positioning method;

[0059] In the height dimension, the zero-sequence impedance characteristics of the grounding transformer are combined to verify the phase identification results;

[0060] Generate cross-partition collaborative handling instructions, including load transfer path optimization plans and protection device linkage strategies.

[0061] This application provides a distribution network fault accurate location system based on graph convolutional neural network, including:

[0062] The distribution network monitoring module is used to deploy monitoring equipment at distribution network nodes to obtain topological connection relationships, electrical quantity monitoring data, environmental monitoring data, and new energy output forecast data;

[0063] A dynamic graph structure generation module is used to respond to distributed power supply switching events and generate a dynamic graph structure based on a pre-stored simulation model;

[0064] An operation parameter generation module is used to initialize graph convolution kernel parameters based on the dynamic graph structure and generate two types of operation parameters based on communication delay;

[0065] A dynamic cause-effect graph construction module, configured to integrate the new energy output forecast data, the electrical quantity monitoring data, and the meteorological data to construct a dynamic cause-effect graph;

[0066] The associated node data acquisition module is used to extract the electrical quantity monitoring data, topological connection relationship and causal reasoning knowledge of the associated nodes when a fault characteristic signal is detected;

[0067] The fault coordinate output module is used to construct a graph convolution network with the dynamic graph structure as the network topology, select the operation parameters corresponding to the communication delay area as the convolution kernel weight, process the electrical quantity monitoring data, topological connection relationship and causal reasoning knowledge of the associated nodes, and output the fault coordinates.

[0068] Based on the embodiments provided in the present application, by deploying monitoring equipment at the distribution network nodes and synchronously acquiring topological connection relationships, electrical quantity monitoring data, environmental monitoring data and new energy output forecast data, a multi-dimensional perception of the operating status of the distribution network is achieved. Compared with the traditional method that relies only on a single electrical quantity data, this method can comprehensively capture the impact of distributed power switching, new energy output fluctuations and environmental factors on fault characteristics, and provide a richer data source for subsequent precise positioning. When responding to a distributed power switching event, a dynamic graph structure is generated based on a pre-stored simulation model, so that the network model can match the topology changes of the distribution network caused by power switching in real time. This mechanism solves the defect in traditional technologies that static models are difficult to adapt to dynamic changes in the system, and ensures that the fault location model still has good applicability in scenarios where the topology is frequently adjusted.

[0069] The graph convolution kernel parameters are initialized based on a dynamic graph structure, and two types of computational parameters are generated based on communication delay, achieving adaptive optimization for different communication environments. By matching the computational parameters of the communication delay region as convolution kernel weights, the impact of data transmission delay on the real-time performance of fault location is reduced. Compared with traditional methods that do not consider communication differences, this significantly improves computational efficiency in complex network environments. A dynamic causal graph is constructed by integrating renewable energy output forecast data, electrical quantity monitoring data, and meteorological data, transforming potential correlations between multi-source data into computable graph-structured relationships. This mechanism overcomes the limitations of traditional methods in fusing multi-source data, especially in complex scenarios such as extreme weather or drastic fluctuations in renewable energy output. It can more accurately explore the causal relationship between fault characteristics and influencing factors, thereby improving location accuracy. A graph convolutional network with a dynamic graph structure as the network topology is constructed, combining the topological connectivity of associated nodes, electrical quantity data, and causal reasoning knowledge for joint processing. This architecture fully utilizes the advantages of graph convolutional neural networks in processing non-Euclidean spatial data, deeply integrating the distribution network topology structure with electrical quantity characteristics. Compared with traditional positioning methods based on rules or static models, it can more efficiently extract fault feature patterns and achieve accurate output of fault coordinates. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] The drawings described herein are used to provide a further understanding of the embodiments of the present invention and constitute a part of this application. The illustrative embodiments of this application and their descriptions are used to explain this application and do not constitute an improper limitation of this application. In the drawings:

[0071] Figure 1 This is a flowchart of an optional method for accurately locating distribution network faults based on a graph convolutional neural network according to an embodiment of the present application;

[0072] Figure 2 This is a flowchart of an optional method for accurately locating distribution network faults based on a graph convolutional neural network according to an embodiment of the present application;

[0073] Figure 3 This is a structural diagram of an optional distribution network fault precise location system based on graph convolutional neural network according to an embodiment of the present application.

[0074] The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings and in conjunction with the embodiments. DETAILED DESCRIPTION

[0075] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0076] Alternatively, as Figure 1 As shown, the present application provides a method for accurately locating distribution network faults based on graph convolutional neural networks, including:

[0077] S101, deploy monitoring equipment at distribution network nodes to obtain topological connection relationships, electrical quantity monitoring data, environmental monitoring data, and new energy output forecast data;

[0078] In this embodiment, monitoring equipment is deployed at distribution network nodes to collect physical quantities, electrical parameters, and environmental data. Monitoring equipment includes, but is not limited to, voltage / current sensors (collecting three-phase voltage and current waveforms), smart meters (monitoring active / reactive power), environmental monitoring units (temperature, humidity, and lightning strike sensors), PMU phasor measurement units (synchronously collecting electrical quantity phasors), and power prediction terminals at new energy stations (such as photovoltaic power prediction devices).

[0079] S102, responding to a distributed power supply switching event, generating a dynamic graph structure based on a pre-stored simulation model;

[0080] In this embodiment, a distributed power generation switching event is a system state change event caused by the grid connection or disconnection of a distributed power source (such as photovoltaic, wind power, or energy storage). For example, a rooftop photovoltaic power station in an industrial park may be automatically connected to the grid due to changes in sunlight intensity, or an energy storage device may be shut down due to grid frequency fluctuations.

[0081] Pre-existing simulation models are pre-built digital twins of distribution networks that simulate the system's dynamic response based on topology, device parameters, and operating rules. For example, an electromagnetic transient model of a distribution network with distributed power sources, built using PSCAD / EMTDC, can simulate power flow distribution and voltage fluctuations during power switching. Model inputs include line impedance parameters, transformer ratios, and distributed power source control strategies.

[0082] S103, initializing graph convolution kernel parameters based on the dynamic graph structure, and generating two types of operation parameters based on communication delay;

[0083] In this embodiment, two sets of convolution kernel parameters are generated based on the delay characteristics of the communication network, adapted for real-time data updates and delay-tolerant scenarios. For example, "real-time update parameters" (iteratively corrected based on the latest sampled data) are generated for areas where communication delay is less than or equal to the delay threshold, while "predictive generation parameters" (calculated in advance based on historical trends and load forecasts) are generated for areas where communication delay is greater than the delay threshold.

[0084] S104, integrating new energy output forecast data, electrical quantity monitoring data, and meteorological data to construct a dynamic cause-and-effect diagram;

[0085] S105, when a fault characteristic signal is detected, extracting electrical quantity monitoring data, topological connection relationships, and causal reasoning knowledge of associated nodes;

[0086] S106, construct a graph convolution network with a dynamic graph structure as the network topology, select the operation parameters corresponding to the communication delay area as the convolution kernel weight, process the electrical quantity monitoring data, topological connection relationship and causal reasoning knowledge of the associated nodes, and output the fault coordinates.

[0087] In this embodiment, corresponding communication delay zones are defined by different computational parameters, based on the latency metrics of the communication links in each area of ​​the distribution network. For example, the urban cable line area is classified as a low-latency area due to a latency of approximately 20ms due to 5G communication coverage, while the suburban overhead line area is classified as a high-latency area due to a latency of approximately 80ms due to wireless communication relays. Corresponding computational parameters are applied to each area.

[0088] Based on the embodiments provided in the present application, by deploying monitoring equipment at the distribution network nodes and synchronously acquiring topological connection relationships, electrical quantity monitoring data, environmental monitoring data and new energy output forecast data, a multi-dimensional perception of the operating status of the distribution network is achieved. Compared with the traditional method that relies only on a single electrical quantity data, this method can comprehensively capture the impact of distributed power switching, new energy output fluctuations and environmental factors on fault characteristics, and provide a richer data source for subsequent precise positioning. When responding to a distributed power switching event, a dynamic graph structure is generated based on a pre-stored simulation model, so that the network model can match the topology changes of the distribution network caused by power switching in real time. This mechanism solves the defect in traditional technologies that static models are difficult to adapt to dynamic changes in the system, and ensures that the fault location model still has good applicability in scenarios where the topology is frequently adjusted.

[0089] The dynamic graph structure is used as a benchmark to initialize the graph convolution kernel parameters, and two types of operation parameters are generated according to the communication delay, so as to realize adaptive optimization in different communication environments. By matching the operation parameters of the communication delay area as the convolution kernel weight, the influence of data transmission delay on the real-time performance of fault location can be reduced. Compared with the traditional method which does not consider the communication difference, the operation efficiency in the complex network environment is significantly improved. The dynamic causal graph is constructed by fusing new energy output prediction data, electrical quantity monitoring data and weather data, and the potential correlation between multiple source data is converted into a calculable graph structure relationship. This mechanism breaks through the limitation of the traditional method that the fusion ability of multiple source data is insufficient, especially in complex scenes such as extreme weather or sharp fluctuations of new energy output, and can more accurately mine the causal relationship between fault characteristics and influencing factors, thereby improving the positioning accuracy. The graph convolution network with dynamic graph structure as the network topology is constructed, and the topological connection relationship of the associated nodes, the electrical quantity data and the causal reasoning knowledge are jointly processed. The architecture makes full use of the processing advantage of the graph convolution neural network for non-Euclidean space data, deeply integrates the power distribution network topology structure and electrical quantity characteristics, and compared with the traditional positioning method based on rules or static model, the fault feature mode can be more efficiently extracted, and the accurate output of the fault coordinates can be realized.

[0090] Further, as shown in Figure 2 in response to the switching event of the distributed power supply, a dynamic graph structure is generated based on a pre-stored simulation model, including:

[0091] In S201, the electrical coupling strength matrix between the nodes of the power distribution network is calculated, and the power distribution network topology network is divided into multiple dynamic autonomous regions according to the electrical coupling strength matrix.

[0092] In this embodiment, the electrical coupling strength matrix represents a matrix representing the degree of electrical connection between nodes in the power distribution network. It is usually constructed based on the admittance matrix or the results of power flow calculation. For example, by calculating the short-circuit current distribution coefficient or voltage sensitivity between nodes, an n x n matrix is generated, and the larger the element value, the stronger the coupling between nodes. For example, the mutual admittance value between nodes i and j is used as the matrix element, which reflects the electrical connection strength between the two nodes through the line.

[0093] The division of the dynamic autonomous region includes: according to the electrical coupling strength matrix, the topology network is divided into several sub-regions with strong internal coupling and weak external coupling, and each region can be independently optimized for topology. For example, in a certain 10kV power distribution network, the low-voltage nodes (coupling strength > 0.7) under the same power distribution transformer are divided into an autonomous region, while the high-voltage nodes (coupling strength < 0.3) between different transformers are divided into different regions.

[0094] In one specific embodiment, the electrical coupling strength matrix is calculated based on the following formula:

[0095]

[0096] Among them, C ij Y is the electrical coupling strength between nodes i and j. A larger value indicates a tighter electrical connection. ij is the mutual admittance between nodes i and j (unit: Siemens S), which is obtained through the distribution network admittance matrix; n is the total number of distribution network nodes; k is the index of the distribution network node.

[0097] Based on the above equation, the normalized mutual admittance is used to quantify node coupling strength, providing a precise basis for dynamic autonomous region partitioning. For example, in a distribution network with distributed power sources, closely connected power and load nodes can be grouped into the same region. This allows subsequent topology optimization to better align with actual electrical characteristics, addresses the drawback of traditional topology partitioning that ignores electrical coupling, and improves the model's adaptability to dynamic changes such as power switching.

[0098] S202, randomly generating a topology connection variation strategy within each dynamic autonomous region, and verifying the regional electrical stability through power flow calculation;

[0099] In this embodiment, the topology connection variation strategy is a randomly generated topology adjustment scheme within the autonomous region, including combinations of switch state changes, tie line switching, and other operations. For example, within a certain autonomous region, a variation strategy of "closing section switch S1 and opening tie switch S2" or "activating backup line L3" is randomly generated, generating 5-10 different topology combinations each time.

[0100] S203, using a reinforcement learning strategy with the goal of minimizing node voltage deviation, to train the agent to select the optimal topology combination from the topology connection variation strategy;

[0101] In this embodiment, the optimal topology combination is defined as a topology that meets the voltage deviation minimization objective while also complying with power constraints. For example, through reinforcement learning, an agent is trained to select the solution that minimizes the mean node voltage deviation rate within a region from among multiple mutation strategies. For example, after power flow calculation, a solution with a node voltage deviation rate of less than 3% is superior to the 5% deviation rate of other solutions.

[0102] S204, migrating the topological connection features of the high electrical stability region to the low electrical stability region; wherein the voltage deviation rate of the high electrical stability region is less than the deviation rate threshold, and the voltage deviation rate of the low electrical stability region is greater than the deviation rate threshold;

[0103] In this embodiment, the deviation rate threshold is used to determine the critical voltage deviation for regional electrical stability, typically taking the upper limit of the allowable range of national standards. For example, a threshold of 5% is set. A voltage deviation rate of >5% in a region is considered low stability, while a voltage deviation rate of ≤5% indicates high stability. (Referring to GB / T12325-2022, which stipulates a ±7% tolerance for three-phase power supply voltages of 10 kV and below, a more stringent threshold is used here.)

[0104] S205 , verifying the line power constraint and voltage fluctuation range, and outputting a dynamic graph structure that meets the safe operation standards of the distribution network.

[0105] In this embodiment, the line power constraint and voltage fluctuation range refer to the maximum permitted transmission power limit of the line and the permissible range of voltage deviation from the rated value during system operation, respectively. For example, a 10kV line with a rated current of 300A corresponds to an active power constraint of 5.2MW; the voltage fluctuation range is set to ±7% of the rated value (e.g., the permissible voltage of a 10kV line is between 9.3kV and 10.7kV).

[0106] Safe operation standards are distribution network operation safety specifications stipulated by the state or industry, including requirements such as power quality, equipment capacity, and protection configuration.

[0107] Based on the embodiments provided in this application, the dynamic autonomous regions are divided by calculating the electrical coupling strength matrix, breaking the limitations of traditional overall modeling and realizing the refined management of the local topological characteristics of the distribution network. In each region, a topological connection variation strategy is randomly generated and combined with the flow calculation to verify the stability. In combination with the reinforcement learning strategy with the goal of minimizing the node voltage deviation, the optimal solution can be screened out from a large number of topological combinations, effectively improving the dynamic adaptability of the distribution network topology structure. At the same time, the topological features of the high electrical stability area are migrated to the low stability area to make up for the local weak links, and finally a dynamic graph structure that meets safety standards is output, providing a network model foundation that is more in line with actual operation for subsequent fault location.

[0108] Furthermore, the graph convolution kernel parameters are initialized based on the dynamic graph structure, and two types of operation parameters are generated based on the communication delay, including:

[0109] The node impedance parameters in the dynamic graph structure are converted into graph edge weight matrices as the basis for initializing the graph convolution kernel parameters;

[0110] In this embodiment, the node impedance parameter is the equivalent impedance value of each node in the distribution network, reflecting the node's resistance to current flow. It is typically obtained through short-circuit testing or power flow calculations. For example, the equivalent positive-sequence impedance of a 10kV busbar node is 0.5 + j2.0Ω, and the zero-sequence impedance is 1.2 + j4.5Ω. These parameters are used to characterize the electrical characteristics of the node under different fault types.

[0111] Among them, the fault types include but are not limited to short circuit faults (such as three-phase short circuit, two-phase short circuit, single-phase ground short circuit), line break faults (single-phase line break, two-phase line break), ground faults (metallic grounding, arc grounding); internal equipment faults (transformer winding short circuit, poor contact of switchgear); distributed power supply related faults (photovoltaic inverter failure, energy storage battery overheating); and compound faults of short circuit and grounding coexisting.

[0112] The conversion into a graph edge weight matrix includes: mapping the impedance values ​​between nodes to the weights of the edges in the graph structure and constructing a weighted adjacency matrix that reflects the strength of the electrical connection.

[0113] Real-time updates are performed on dynamic autonomous regions where communication delays are below a threshold. This includes analyzing the real-time gradient changes in electrical quantity monitoring data within the region to generate convolution kernel parameter correction vectors; and aggregating historical operating characteristic trends within the region to generate parameter offset compensation.

[0114] In a specific embodiment, the convolution kernel parameter correction vector is generated based on the following formula:

[0115]

[0116] Where Δθ is the convolution kernel parameter correction vector; α is the real-time gradient weight coefficient (range 0.1-0.5), which adjusts the impact of real-time data; τ is the time scale coefficient (unit: s); is the real-time gradient of the electrical quantity (unit: V / s or A / s), reflecting the rate of change of the electrical quantity; β is the historical trend weight coefficient (dimensionless, ranging from 0.5 to 0.9), which adjusts the influence of historical data; T is the historical operating characteristic trend (unit: V or A), extracted by sliding average.

[0117] Based on the above equation, the real-time electrical quantity gradient and historical operating trends are combined to generate a correction vector, addressing the parameter update flaws of traditional single-data-source systems. When the switching of distributed generation sources causes a sudden change in current, the real-time gradient term quickly responds to the change. During periods of data stability, the historical trend term ensures parameter stability, enabling the graph convolutional network to adaptively optimize parameters in areas with varying communication delays, improving the real-time and accuracy of fault location.

[0118] In this embodiment, the delay threshold is a critical value that distinguishes communication delay levels and is used to determine whether a region belongs to real-time update mode or predictive generation mode. For example, if the threshold is set to 50ms and the measured communication delay in one region is 30ms (< the threshold), the real-time update mode is used; while in another region, the delay is 80ms (> the threshold), the predictive generation mode is used. (The threshold can be adjusted based on the actual performance of the distribution network communication system, for example, 4G communication typically has a delay of 30-100ms, while 5G delay is <20ms.)

[0119] Real-time gradient refers to the rate of change of electrical quantity monitoring data over time, reflecting the speed of the system's dynamic processes. For example, if a line current suddenly increases from 200A to 500A within 10ms, the gradient is (500-200) / 0.01 = 30kA / s; if a node voltage drops from 10kV to 8kV within half a cycle, the gradient is -4kV / cycle (a 50Hz system cycle is 20ms).

[0120] Historical operating characteristic trends are device operating patterns or system change trends extracted from historical data and used to predict future conditions. For example, the daily power curve for a photovoltaic power station over the past seven days shows peak output from 10:00 AM to 2:00 PM, with a normal fluctuation trend. The load curve for a distribution transformer over the past three months shows a period of increasing load from 8:00 AM to 10:00 AM every Monday. This trend can be extracted for parameter compensation.

[0121] For dynamic autonomous regions where communication delay exceeds the delay threshold, predictions are generated, including: inferring load distribution characteristics based on the line impedance parameters in the region; predicting load change trends based on the historical switching frequency records of distributed power sources in the region; and mapping the prediction results into convolution kernel frequency domain responses through orthogonal frequency domain transformation.

[0122] In this embodiment, the line impedance parameters are the resistance and reactance values ​​per unit length of the line, which are determined by the material, cross-section, and laying method of the conductor and are used for power flow calculation and fault analysis.

[0123] Adaptive frequency domain filtering is performed with priority given to key equipment in the distribution network, generating two types of operation parameters: real-time update type and delay tolerance type.

[0124] Based on the embodiment provided by the present application, the node impedance parameters are converted into graph edge weight matrices to initialize the graph convolution kernel parameters, so that the network parameters are closely related to the electrical characteristics of the distribution network. In view of the difference in communication delay, the gradient changes of electrical quantity data and historical operating characteristics are analyzed in real time in the low-latency area, and correction vectors and compensation amounts are generated to ensure that the parameters are updated in time to adapt to the real-time working conditions; load forecasting is performed for the high-latency area based on line impedance and historical switching records, and mapped to the convolution kernel frequency domain response through orthogonal frequency domain transformation, and a parameter model that adapts to the delay is constructed in advance. Finally, adaptive frequency domain filtering with priority on key equipment is performed to generate two types of operation parameters, which not only guarantees the priority of key equipment data processing, but also takes into account the effectiveness of the model under different communication environments, significantly improving the network's adaptability to complex communication conditions.

[0125] Furthermore, adaptive frequency domain filtering with priority given to key equipment in the distribution network is performed, including:

[0126] Identify key equipment nodes in the distribution network, including distributed power access points and load center connection points with capacities greater than a set capacity threshold;

[0127] In this embodiment, the capacity threshold is used to define the critical capacity value of key equipment. Devices exceeding this threshold are considered nodes with significant system impact. For example, the capacity threshold for distributed power access points is set at 100kVA (e.g., a photovoltaic power plant with a capacity of 150kVA is considered a critical node if it exceeds the threshold), and the capacity threshold for load center connection points is set at 500kVA (e.g., a commercial complex with a load of 600kVA is considered a critical node).

[0128] Calculate the electrical impact factor of key equipment nodes, which is determined by equipment capacity weight, topological centrality, and voltage sensitivity;

[0129] In this embodiment, the electrical impact factor is an indicator that comprehensively considers device capacity, topological location, and voltage sensitivity, and is used to evaluate the importance of a device to system operation.

[0130] Equipment capacity weight: calculated based on the ratio of the equipment's rated capacity to the total regional capacity. For example, if a 300kVA distributed power supply accounts for 15% of the regional total capacity, its weight is 0.15.

[0131] Topological centrality: Using the degree centrality index, a node is connected to 3 lines, and the centrality is 3 / (n-1) (n is the number of regional nodes);

[0132] Voltage sensitivity: The ratio of the node voltage change to the reactive power change. The higher the sensitivity, the greater the electrical impact factor.

[0133] Construct a frequency domain filter mask matrix, and increase the retention weight of frequency domain components associated with key equipment to 1.5 times;

[0134] In this embodiment, the frequency domain filter mask matrix is ​​a matrix that differentially weights the signals of key devices and non-key areas in the frequency domain space to enhance key features. For example, a diagonal matrix is ​​constructed, with the elements of the frequency band corresponding to the key devices set to 1.5 (for example, retaining the weight of the 100Hz harmonic component) and the elements of the non-key areas set to 0.5 (for example, attenuating high-frequency components above 500Hz). The matrix dimension is consistent with the number of components after frequency domain decomposition.

[0135] Perform amplitude attenuation processing on the frequency domain components in non-critical areas, and the attenuation coefficient is negatively correlated with the electrical impact factor;

[0136] In this embodiment, non-critical frequency domain components are frequency components that are not directly associated with critical equipment, or signal components that have little impact on fault location. For example, in a distribution network, high-order harmonics (such as 11th harmonics and above) at non-critical load nodes, or fundamental component fluctuations during normal operation, are amplitude-attenuated during filtering.

[0137] The time domain convolution kernel parameters are reconstructed through inverse transformation to generate differentiated operation parameters that adapt to the operating characteristics of the distribution network.

[0138] Based on the embodiments provided in this application, by identifying the key equipment nodes of the distribution network, the electrical impact factor is calculated in combination with the equipment capacity weight, topological centrality and voltage sensitivity to clarify the importance of the equipment in the network. Based on this, a frequency domain filter mask matrix is ​​constructed, the weight of the frequency domain components associated with key equipment is increased, and the amplitude attenuation processing is performed on non-key areas, so that the graph convolution network can focus more on the electrical characteristics of key equipment when processing data, highlighting the core information required for fault location. The time domain convolution kernel parameters are reconstructed through inverse transformation, and the generated differentiated operation parameters can accurately adapt to the operating characteristics of the distribution network. Compared with the traditional unified processing method, the accuracy and pertinence of fault feature extraction are greatly improved.

[0139] Furthermore, a dynamic cause-and-effect diagram is constructed by integrating renewable energy output forecast data, electrical quantity monitoring data, and meteorological data, including:

[0140] Align the collection time series of renewable energy output forecast data and electrical quantity monitoring data;

[0141] The waveform correlation strength between lightning events and voltage sags is determined by wavelet coherence analysis.

[0142] Wavelet coherence analysis, a time-frequency analysis method based on the wavelet transform, is used to quantify the strength of correlation between different signals in the time-frequency domain. For example, the waveforms of lightning strikes (recorded by a lightning location system) and voltage sags (collected by a PMU) are analyzed to calculate the correlation coefficient between the two in the 10-100 Hz frequency band. For example, if a voltage sag occurs 5 milliseconds after a lightning strike, the correlation coefficient reaches 0.85, indicating a strong correlation between the two.

[0143] A three-channel feature fusion network was constructed for data processing, including an electrical quantity channel for extracting high-frequency mutation features of voltage and current waveforms; a meteorological channel for fusing spatially distributed data from temperature and humidity sensors; and a new energy channel for analyzing the deviation rate of photovoltaic / wind power forecasts.

[0144] High-frequency mutations are the sudden appearance of high-frequency components in electrical waveforms, often associated with faults or transient processes. For example, during a short-circuit fault, harmonic components above 1000 Hz may appear in the current waveform (this frequency band has low energy during normal operation), or high-frequency oscillations (such as steep-front traveling waves) may appear in the voltage waveform within 0.1 ms.

[0145] The PV / wind power forecast deviation rate is the percentage of the difference between the actual output of renewable energy and the predicted value, reflecting the forecast accuracy. For example, a PV power station predicted an output of 500 kW at 12:00 PM, but the actual output was 450 kW. The deviation rate is (450-500) / 500 = -10%. A wind farm predicted a wind speed of 8 m / s, corresponding to a power of 800 kW. The actual wind speed was 6 m / s, resulting in a power of 500 kW. The deviation rate is (500-800) / 800 = -37.5%.

[0146] Train the Bayesian inference network to output the fault causal probability matrix and dynamically update the causal weights of short circuit faults and insulation aging;

[0147] In a specific embodiment, the failure causal probability is calculated based on the following formula:

[0148]

[0149] Among them, P(F k |E i ) is the observed event E i Time Fault F k The conditional probability of P(E i |F k ) is fault F k When event E occurs i The conditional probability of P(E i |F j ) is fault F j When event E occurs i The conditional probability of P(F k ) is fault F k The prior probability of is calculated based on the historical fault frequency; m is the total number of fault types, and j is the index of the fault type.

[0150] Based on the above formula, fault causal weights are dynamically updated by integrating multi-source data (such as abnormal renewable energy output and meteorological events). For example, when a lightning strike and a sudden drop in PV output are observed simultaneously, traditional methods may misidentify it as a PV equipment failure (with a probability of 0.6). However, this formula calculates a line short-circuit fault probability of 0.82, more accurately reflecting the actual fault cause and improving fault location accuracy in complex scenarios.

[0151] In this embodiment, the Bayesian inference network outputs a causal probability matrix, a network model constructed based on Bayesian statistical theory, that outputs a probabilistic association matrix between fault types and influencing factors. For example, the network inputs are electrical quantity characteristics (such as zero-sequence current), environmental data (such as humidity), and the new energy deviation rate. In the output matrix P, P(short circuit | insulation aging) = 0.75, indicating a 75% probability of a short circuit fault under insulation aging conditions. This matrix is ​​continuously updated with new data.

[0152] A fault path reverse tracing mechanism is established to scan the upstream equipment causal chain when a ground fault is detected to generate a dynamic causal graph with weight distribution.

[0153] In this embodiment, the upstream device causal chain is a sequence of devices traced backward from the fault point along the power transmission direction, sorted by the strength of the causal relationship to form a chain. For example, if a ground fault occurs on a 10kV line, the causal chain scan order is: fault point → upstream sectionalizer → distribution transformer → substation outgoing line switch. The causal weight of each device is calculated (e.g., transformer insulation aging weight 0.6, switch contact failure weight 0.3), and high-weighted devices are prioritized for investigation.

[0154] Based on the embodiment provided in this application, by aligning the time series of new energy output and electrical quantity data, and combining the wavelet coherence analysis method to quantify the correlation strength between meteorological factors such as lightning strikes and voltage drops, a reliable basis is provided for multi-source data fusion. A three-channel feature fusion network is constructed to extract features from the three dimensions of electrical quantity, meteorology, and new energy respectively. The Bayesian inference network is then used to output the fault causal probability matrix, and the fault causal weights are dynamically updated so that the model can learn the correlation between faults and influencing factors as the operating conditions change. The established fault path reverse tracing mechanism, when a ground fault occurs, scans the upstream equipment causal chain hierarchically to generate a dynamic causal graph, comprehensively sorts out the potential causes of the fault, and provides a multi-dimensional analysis perspective for accurately locating the root cause of the fault.

[0155] Furthermore, a fault path reverse tracing mechanism is established, including:

[0156] Define the fault propagation depth threshold and initiate multi-level reverse tracing when a ground fault is detected;

[0157] In this embodiment, the fault propagation depth threshold defines the hierarchical parameter of the fault reverse tracing range, measured in topological connection hops, to control the hierarchical depth of the device being traced. For example, if the threshold is set to 1 hop, only upstream devices directly connected to the fault point (such as section switches and transformers) are traced; if the threshold is set to 3 hops, the tracing range is extended to devices connected to the fault point via three topological levels (such as substation outgoing line switches and adjacent feeder tie switches).

[0158] First-level tracing: locates the upstream equipment directly connected to the fault point and calculates the contribution of the fault current;

[0159] The directly connected upstream device to the fault point is a device located on the power supply side, directly connected to the fault point via a line in the distribution network topology. This device is the direct source path of the fault current. For example, if a short circuit occurs on a 10kV line L1, the directly connected upstream devices are the section switch S1 and the distribution transformer T1 (on the power supply side).

[0160] Second-level tracing: Scans associated devices within two hops of topological distance and evaluates the correlation between device insulation aging indicators and faults.

[0161] Two-hop associated devices are devices connected to the fault point via two intermediate nodes, meaning they have a topological distance of 2. This expands the scope of fault cause investigation. For example, if the fault point is at node A on line L1, the one-hop device is switch S1, and the two-hop devices are switch S2 and transformer T2 on line L2 connected via S1.

[0162] Level 3 tracking: Analyze power supply points where the output fluctuation rate of renewable energy exceeds the fluctuation rate threshold, and calculate the probability of fault induction due to power mutation.

[0163] The volatility threshold is the critical value for determining whether renewable energy output fluctuations are abnormal, expressed as a percentage of the power change rate per unit time. For example, if the photovoltaic output volatility threshold is set at 20% / 10 minutes, a photovoltaic power station's power will drop from 500kW to 350kW within 10 minutes (a 30% change rate), exceeding the volatility threshold. Similarly, if the wind power output volatility threshold is set at 15% / 5 minutes, a wind farm's power will increase from 800kW to 1000kW within 5 minutes (a 25% change rate), exceeding the volatility threshold.

[0164] Based on the embodiments provided by this application, by defining the fault propagation depth threshold, the fault source is tracked hierarchically. The first level locates the directly connected upstream device and calculates the fault current contribution, quickly locking the directly affected device; the second level scans the related devices within two hops to evaluate the insulation aging index and check for potential hidden dangers; the third level analyzes the power points where the new energy output fluctuation rate exceeds the threshold to determine the impact of the power mutation on the fault. This hierarchical tracking method gradually checks from near to far, from direct to potential factors. Compared with a single method, it can more comprehensively and systematically determine the cause of the fault, avoid missing key fault clues, and provide a more complete chain of evidence for fault location.

[0165] Furthermore, when a fault characteristic signal is detected, the electrical quantity monitoring data, topological connection relationship and causal reasoning knowledge of the associated nodes are extracted, including:

[0166] Activate dedicated analysis channels based on fault feature types, including: activating the electromagnetic transient analysis channel to extract current harmonic distortion features when the voltage drop exceeds the drop threshold; activating the thermal distribution calculation channel to generate a thermal stress distribution diagram for the equipment when the temperature rise exceeds the temperature change threshold;

[0167] In this embodiment, the voltage drop threshold is the critical value for determining a voltage sag fault, typically expressed as a percentage of the rated voltage. For example, the voltage drop threshold is set to 90% of the rated voltage (e.g., a 10kV line voltage drops below 9kV); for sensitive load scenarios, the voltage drop threshold is set to 95% (initiating a special analysis when a 10kV line voltage drops below 9.5kV).

[0168] The temperature change threshold is the critical value for abnormal equipment temperature increases, expressed as the amount of temperature rise per unit time. For example, if the transformer oil temperature change threshold is set to 5°C / min, thermal analysis will be activated if the oil temperature rises by 6°C within 1 minute. If the cable connector temperature change threshold is set to 3°C / 5min, thermal stress calculation will be initiated if the temperature rises by 4°C within 5 minutes.

[0169] The improved Dijkstra algorithm is used to search for associated nodes within three hops of the fault source under the power supply radius constraint;

[0170] The power supply radius constraint is the maximum allowable physical distance between the power source and the load point in the distribution network. It limits the search range for fault-associated nodes. For example, a city distribution network sets a power supply radius constraint of 5km. When searching for nodes within three hops of the fault source, devices within 5km of the power source are also screened. Rural distribution networks can set the radius constraint to 10km to accommodate a wider range.

[0171] Three-hop: In a distribution network topology, the distance from a fault point through three intermediate devices (such as switches and transformers) is used to quantify the topological hierarchy between nodes. For example, fault point → switch S1 (1 hop) → switch S2 (2 hops) → transformer T1 (3 hops), where T1 is the associated device within three hops.

[0172] Extracting causal reasoning knowledge with causal weight values ​​greater than a weight threshold from the dynamic causal graph;

[0173] The weight threshold is the critical value for screening high-confidence causal relationships in the dynamic causal graph. Knowledge with a weight exceeding this value is extracted for fault analysis. For example, setting the weight threshold to 0.6 extracts relationships in the causal matrix with P(lightning strike → insulator flashover) = 0.72. For complex fault scenarios, raising the threshold to 0.8 retains only strong relationships such as P(insulation aging → short-circuit fault) = 0.85.

[0174] Apply the causal probability propagation model to screen causal nodes whose correlation strength exceeds the correlation strength threshold;

[0175] In this embodiment, the causal probability propagation model is an inference model based on a probabilistic graphical model (such as a Bayesian network), which calculates the fault correlation strength by transferring conditional probabilities between nodes. For example, a directed acyclic graph containing fault types, device states, and environmental factors is constructed. For example, the "lightning strike" node influences the "insulator flashover" node through the conditional probability P (flashover | lightning strike) = 0.8, and then influences the "line short circuit" node through P (short circuit | flashover) = 0.9. Finally, the posterior probability of each node is calculated as the correlation strength.

[0176] Correlation strength threshold: a critical value for screening strong correlation fault causes. Nodes exceeding this value are considered to have a direct correlation with the fault. For example, the correlation strength threshold is 0.5, and the causal correlation strength of a certain device is calculated as 0.63, which is included in the analysis. For multiple fault scenarios, the correlation strength threshold is 0.7, and only nodes with a correlation strength greater than or equal to 0.7 are retained (such as a transformer aging correlation strength of 0.75 and a line overload of 0.71).

[0177] Convert electrical quantity monitoring data into wavelet packet energy spectrum features;

[0178] Encode the topological connection relationship into a spectral feature vector of node adjacency relationship;

[0179] Embed the causal weight value in the dynamic causal graph into the edge attribute matrix of the graph convolution network, so that the causal knowledge participates in the message passing calculation.

[0180] Based on the embodiments provided in the present application, dedicated analysis channels are activated according to different fault feature types, such as starting electromagnetic transient analysis to extract current harmonic features when voltage sag occurs, and activating thermal distribution calculation when temperature anomaly occurs, to realize targeted and deep mining of fault features. Improved Dijkstra algorithm is used to search for associated nodes under the constraint of power supply radius, ensuring that the nodes included in the analysis meet both topological relationships and actual power supply logic. High-weight causal reasoning knowledge is extracted from the dynamic causal graph, combined with the causal probability propagation model to screen strong correlation nodes, and the fault analysis is focused on key factors. Convert electrical quantity data into wavelet packet energy spectrum features, encode topological relationships into spectral feature vectors, and embed causal weights into the edge attribute matrix of the graph convolution network, to realize effective fusion and collaborative calculation of multiple source information, and provide more rich and accurate input features for fault location.

[0181] Further, a graph convolution network with a dynamic graph structure as the network topology is constructed, and the operation parameters corresponding to the communication delay area are selected as the convolution kernel weights to process the electrical quantity monitoring data of the associated nodes, the topological connection relationship, and the causal reasoning knowledge and output the fault coordinates, including:

[0182] Configure a hierarchical convolution architecture that adapts to the radial topology of the distribution network, including: implementing a three-order neighborhood aggregation operation on the main road of the feeder line; and performing a first-order neighborhood convolution on the branch line;

[0183] In this embodiment, a third-order neighborhood aggregation operation is performed: in the graph convolutional network of the distribution network feeder trunk, the electrical quantity characteristics of the current node and its two-layer topological neighbors (a total of three-layer node range) are aggregated to capture the characteristics of long-distance fault propagation. For example, the third-order neighborhood of a feeder trunk node A includes: the directly connected node B (1st order), the node C connected to node B (2nd order), and the node D connected to node C (3rd order). The voltage and current characteristics of A, B, C, and D are aggregated to analyze the global electrical quantity changes during long-distance trunk faults (such as short circuits at substation outlets).

[0184] First-order neighborhood convolution: In a branch line graph convolutional network, only the features of the current node's directly connected neighbors (1-layer topology scope) are aggregated, adapting to the local feature extraction of short-distance branch line faults. For example, the first-order neighborhood of a branch line node E is the directly connected node F (the branch end load). Only the electrical quantity data of E and F are aggregated to analyze the local current mutation characteristics when the branch line end is short-circuited, reducing computational redundancy for non-relevant nodes.

[0185] Enhance fault feature propagation through causal weight values ​​in edge attribute matrix;

[0186] Build a three-dimensional fault analysis engine, including: longitude dimension: using the reverse graph diffusion algorithm to generate fault probability distribution; latitude dimension: locating distance based on line impedance parameters and fault recording data; altitude dimension: identifying fault phase based on zero-sequence current phase angle difference;

[0187] In this embodiment, the reverse graph diffusion algorithm reversely derives the probability distribution of fault sources from fault-signature nodes within a graph structure, iteratively updating node fault probabilities to ultimately generate a heat map. Specific implementation: Initialization: Nodes with detected fault signatures are set as high-probability sources (probability = 1), and all other nodes have probabilities = 0; Iterative Diffusion: Probabilities are transferred to adjacent nodes based on topological connections, with the transfer weight inversely proportional to the line impedance (e.g., if node A has a fault probability of 0.8, when it diffuses to node B via a line with impedance Z = 1Ω, the transfer probability = 0.8 × (1 / Z) = 0.8); Convergence Detection: When the rate of change of all node probabilities is < 1%, the algorithm stops and generates a fault probability heat map (e.g., if a line segment has a probability > 0.7, it is identified as a high-probability fault area).

[0188] The reverse graph diffusion algorithm generates a probability distribution, including: starting from a fault feature detection node, propagating the fault probability in reverse according to the topological connection, iteratively calculating the fault probability of each node, and finally generating a power distribution network fault probability heat map. For example, if a current mutation (fault feature) is detected in node X, the initial fault probability of X is set to P(X) = 1, and the adjacent node P(Y) = 0.5; the second iteration: node Y transmits the probability P(Z) = P(Y) x (1 / line impedance YZ) to its neighbor node Z, for example, if the YZ impedance is 2Ω, then P(Z) = 0.5 x 0.5 = 0.25; after iteration to convergence, the probability of a certain line segment node is all > 0.6, and it is determined as a high probability area of fault (for example, the probability of 5-8km of main road L1 is 0.72).

[0189] The parameters are loaded in parallel and pipeline calculation is implemented in the FPGA chip.

[0190] The output includes line number, tower geographic coordinates and fault phase.

[0191] Based on the embodiments provided in the present application, a hierarchical convolution architecture suitable for the radial topology of the power distribution network is configured, different order neighborhood operations are used in the feeder main road and branch line, which fits the characteristics of the power distribution network topology structure and effectively extracts network features at different levels. The causal weight in the edge attribute matrix is used to enhance the propagation of fault features, so that the network highlights the key fault clues in the information transmission process. The three-dimensional fault analysis engine is constructed, which uses the reverse graph diffusion algorithm, line impedance and fault recording data, and zero sequence current phase angle difference in longitude, latitude and height dimensions respectively for fault location and phase identification, and cross verifies the fault information in multiple dimensions. The parameters are loaded in parallel and pipeline calculation is implemented in the FPGA chip, which fully utilizes the hardware acceleration advantage, and finally outputs the accurate fault coordinates including line, coordinates and phase, realizing the all-round and efficient processing of fault location.

[0192] Further, a three-dimensional fault analysis engine is constructed, including:

[0193] When the fault probability distribution heat map shows multiple high probability points, the cooperative positioning mechanism is started;

[0194] In the present embodiment, the multiple high probability points: in the fault probability heat map, multiple geographical or topological regions simultaneously appear with fault probability exceeding the set threshold, which usually indicates a complex fault or a cross-region fault. For example: in a certain power distribution network, the middle section of feeder L1 (probability 0.65) and the end of feeder L2 (probability 0.68) simultaneously appear high probability points, which may be a cross-feeder fault near the tie switch; the urban cable area A (probability 0.7) and the suburban overhead area B (probability 0.72) simultaneously appear high probability, which indicates that there may be a multi-point grounding fault or a misjudgment caused by system resonance.

[0195] Identify the power supply partition that contributes most to the fault current;

[0196] Fusing voltage phasor measurement unit data from adjacent zones in the longitude dimension to correct the fault probability distribution;

[0197] In the latitude dimension, a historical similar fault case library is called to optimize the distance positioning method;

[0198] In this embodiment, the line impedance parameters and fault recording data are combined to locate the distance. This involves calculating the physical distance between the fault point and the measurement end by combining the line impedance per unit length and the traveling wave arrival time from the fault recording. For example, a 10kV line with a total length of 10km and a unit impedance of 0.2Ω / km shows a traveling wave time difference of 100μs from the measurement end to the fault point (traveling wave speed is approximately 3×10^5km / s). Therefore, distance = time difference × speed = 30km (but since the line is 10km long, correction is required). Furthermore, the corrected distance is calculated using the line impedance: total impedance = distance to fault point × unit impedance. If the measured fault phase impedance is 1.2Ω, the fault distance = 1.2 / 0.2 = 6km. Cross-validating this with the traveling wave location result confirms that the fault is located 6km away.

[0199] In the height dimension, the zero-sequence impedance characteristics of the grounding transformer are combined to verify the phase identification results;

[0200] In this embodiment, the zero-sequence impedance characteristic of the grounding transformer is an impedance parameter of the grounding transformer in the zero-sequence network, and its value affects the distribution and phase angle of the zero-sequence current, and is used to verify the accuracy of the fault phase identification result.

[0201] For example, it is known that the zero-sequence impedance of a grounding transformer is Z0=3+j15Ω. When a phase A grounding fault occurs, the theoretical zero-sequence current phase angle φ=arctan(15 / 3)=78.7°; if the measured zero-sequence current phase angle is 75° (within the error range of ±5°), the phase identification is verified to be correct; if the measured phase angle is 120°, it indicates that there may be a cross-phase fault or abnormal grounding transformer parameters, and re-analysis is required.

[0202] Identify the fault phase based on the zero-sequence current phase angle difference: The phase difference of the zero-sequence current in each phase under different fault phases is used to determine the specific phase (A / B / C phase or between phases) where the fault occurred. For example, in the case of a single-phase grounding fault (such as phase A), the phase angle of the zero-sequence current of phase A differs from the phase angles of the zero-sequence current of phases B and C by approximately 180°, while the phase angle difference between phases B and C is approximately 0°; if the measured zero-sequence current phase angle of phase A is 30°, that of phase B is 210°, and that of phase C is 215°, then the phase angle difference Δφ(AB) = 180° and Δφ(AC) = 185°, and it is determined to be a phase A grounding fault (the phase angle difference threshold is usually set to ±15°).

[0203] Generate cross-partition cooperative treatment instructions, including load transfer path optimization scheme and protection device linkage strategy.

[0204] Based on the embodiments provided in the present application, in the complex scenario of multiple regional high probability points of fault probability distribution, the cooperative positioning mechanism is started. By determining the power supply partition with the largest fault current contribution, the key fault investigation area is determined. In the longitude dimension, the adjacent partition voltage phasor measurement unit data is fused to correct the fault probability distribution, and the information reference range is expanded in the spatial dimension; in the latitude dimension, the historical similar fault case library is called to optimize the distance positioning, and the historical experience is used to improve the positioning accuracy; in the height dimension, the zero sequence impedance characteristics of the grounding transformer are combined to verify the phase identification, and the inherent electrical parameters of the equipment are used to enhance the judgment basis. Finally, cross-partition cooperative treatment instructions are generated, which not only realize accurate fault positioning, but also output load transfer and protection device linkage strategy, providing a complete solution for rapid recovery and stable operation of the system after the fault.

[0205] Optionally, as shown in Figure 3 , the present application provides a power distribution network fault accurate positioning system based on graph convolutional neural network, comprising:

[0206] The power distribution network monitoring module 301 is used for deploying monitoring equipment at the nodes of the power distribution network, acquiring topological connection relationship, electrical quantity monitoring data, environmental monitoring data and new energy output prediction data;

[0207] The dynamic graph structure generation module 302 is used for generating a dynamic graph structure based on a pre-stored simulation model in response to a distributed power supply switching event;

[0208] The operation parameter generation module 303 is used for initializing the graph convolution kernel parameters based on the dynamic graph structure, and generating two types of operation parameters based on the communication delay;

[0209] The dynamic causal graph construction module 304 is used for constructing a dynamic causal graph by fusing new energy output prediction data, electrical quantity monitoring data and meteorological data;

[0210] The associated node data acquisition module 305 is used for extracting the electrical quantity monitoring data, topological connection relationship and causal reasoning knowledge of the associated nodes when detecting a fault characteristic signal;

[0211] The fault coordinate output module 306 is used for constructing a graph convolution network with the dynamic graph structure as the network topology, selecting the operation parameters corresponding to the communication delay area as the convolution kernel weight, processing the electrical quantity monitoring data, topological connection relationship and causal reasoning knowledge of the associated nodes and outputting the fault coordinate.

[0212] Reference Figure 3 , the power distribution network fault accurate positioning system based on graph convolutional neural network of the present application belongs to the technical field of intelligent power distribution system.

[0213] It should be noted that in this application, the embodiments implemented on the distribution network fault precise positioning system side based on graph convolutional neural network can be referenced with the embodiments implemented on the distribution network fault precise positioning method side based on graph convolutional neural network, and this application will not go into details one by one.

[0214] The above are only preferred embodiments of the present invention and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention description and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A distribution network fault accurate location method based on graph convolutional neural network, characterized by: include: Deploy monitoring equipment at distribution network nodes to obtain topological connection relationships, electrical quantity monitoring data, environmental monitoring data, and new energy output forecast data; In response to distributed power supply switching events, a dynamic graph structure is generated based on pre-stored simulation models; Initializing graph convolution kernel parameters based on the dynamic graph structure, and generating two types of operation parameters based on communication delay; Integrating the new energy output forecast data, the electrical quantity monitoring data and the meteorological data to construct a dynamic cause-and-effect graph; When a fault characteristic signal is detected, the electrical quantity monitoring data, topological connection relationship and causal reasoning knowledge of the associated nodes are extracted; A graph convolutional network with the dynamic graph structure as the network topology is constructed, and the operation parameters corresponding to the communication delay area are selected as the convolution kernel weights. The electrical quantity monitoring data, topological connection relationship and causal reasoning knowledge of the associated nodes are processed and the fault coordinates are output.

2. The method for accurately locating distribution network faults based on graph convolutional neural networks according to claim 1 is characterized in that: The step of responding to a distributed power supply switching event and generating a dynamic graph structure based on a pre-stored simulation model includes: Calculating an electrical coupling strength matrix between distribution network nodes, and dividing the distribution network topology network into a plurality of dynamic autonomous regions according to the electrical coupling strength matrix; Randomly generate topology connection variation strategies in each dynamic autonomous region and verify regional electrical stability through power flow calculations; A reinforcement learning strategy with the goal of minimizing node voltage deviation is used to train the agent to select the optimal topology combination among the topology connection mutation strategies. Migrating the topological connection features of the high electrical stability region to the low electrical stability region; wherein the voltage deviation rate of the high electrical stability region is less than a deviation rate threshold, and the voltage deviation rate of the low electrical stability region is greater than the deviation rate threshold; Verify the line power constraints and voltage fluctuation range, and output a dynamic graph structure that meets the safe operation standards of the distribution network.

3. The method for accurately locating distribution network faults based on graph convolutional neural networks according to claim 2 is characterized in that: The graph convolution kernel parameters are initialized based on the dynamic graph structure, and two types of operation parameters are generated based on the communication delay, including: Converting the node impedance parameters in the dynamic graph structure into a graph edge weight matrix as a basis for initializing graph convolution kernel parameters; Real-time updates are performed on dynamic autonomous regions where communication delays are below a threshold. This includes analyzing the real-time gradient changes in electrical quantity monitoring data within the region to generate convolution kernel parameter correction vectors; and aggregating historical operating characteristic trends within the region to generate parameter offset compensation. For dynamic autonomous regions where the communication delay is higher than the delay threshold, prediction generation is performed, including: inferring the load distribution characteristics based on the line impedance parameters of the region; predicting the load change trend based on the historical switching frequency records of distributed power sources in the region; and mapping the prediction results into a convolution kernel frequency domain response through orthogonal frequency domain transformation; Adaptive frequency domain filtering is performed with priority given to key equipment in the distribution network, generating two types of operation parameters: real-time update type and delay tolerance type.

4. The method for accurately locating distribution network faults based on graph convolutional neural networks according to claim 3 is characterized in that: The method of performing adaptive frequency domain filtering with priority given to key equipment in the distribution network includes: Identify key equipment nodes in the distribution network, including distributed power access points and load center connection points with capacities greater than a set capacity threshold; Calculating the electrical impact factor of key device nodes, where the electrical impact factor is determined by device capacity weight, topological centrality, and voltage sensitivity; Construct a frequency domain filter mask matrix, and increase the retention weight of frequency domain components associated with key equipment to 1.5 times; Performing amplitude attenuation processing on the frequency domain components of the non-critical area, where the attenuation coefficient is negatively correlated with the electrical impact factor; The time domain convolution kernel parameters are reconstructed through inverse transformation to generate differentiated operation parameters that adapt to the operating characteristics of the distribution network.

5. The method for accurately locating distribution network faults based on graph convolutional neural networks according to claim 3 is characterized in that: The new energy output forecast data, the electrical quantity monitoring data and the meteorological data are integrated to construct a dynamic cause-effect graph, including: Aligning the acquisition time series of the new energy output forecast data and the electrical quantity monitoring data; The waveform correlation strength between lightning events and voltage sags is determined by wavelet coherence analysis. A three-channel feature fusion network was constructed for data processing, including an electrical quantity channel for extracting high-frequency mutation features of voltage and current waveforms; a meteorological channel for fusing spatially distributed data from temperature and humidity sensors; and a new energy channel for analyzing the deviation rate of photovoltaic / wind power forecasts. Train the Bayesian inference network to output the fault causal probability matrix and dynamically update the causal weights of short circuit faults and insulation aging; A fault path reverse tracing mechanism is established to scan the upstream equipment causal chain when a ground fault is detected to generate a dynamic causal graph with weight distribution.

6. The method for accurately locating distribution network faults based on graph convolutional neural networks according to claim 5 is characterized in that: The establishment of a fault path reverse tracing mechanism includes: Define the fault propagation depth threshold and initiate multi-level reverse tracing when a ground fault is detected; First-level tracing: locates the upstream equipment directly connected to the fault point and calculates the contribution of the fault current; Second-level tracing: Scans associated devices within two hops of topological distance and evaluates the correlation between device insulation aging indicators and faults. Level 3 tracking: Analyze power supply points where the output fluctuation rate of renewable energy exceeds the fluctuation rate threshold, and calculate the probability of fault induction due to power mutation.

7. The method for accurately locating distribution network faults based on graph convolutional neural networks according to claim 5 is characterized in that: When a fault characteristic signal is detected, the electrical quantity monitoring data, topological connection relationship and causal reasoning knowledge of the associated nodes are extracted, including: Activate dedicated analysis channels based on fault feature types, including: activating the electromagnetic transient analysis channel to extract current harmonic distortion features when the voltage drop exceeds the drop threshold; activating the thermal distribution calculation channel to generate a thermal stress distribution diagram for the equipment when the temperature rise exceeds the temperature change threshold; The improved Dijkstra algorithm is used to search for associated nodes within three hops of the fault source under the power supply radius constraint; extracting causal reasoning knowledge whose causal weight value is greater than a weight threshold from the dynamic causal graph; Apply the causal probability propagation model to screen causal nodes whose correlation strength exceeds the correlation strength threshold; Convert electrical quantity monitoring data into wavelet packet energy spectrum features; Encode the topological connectivity relationship as the spectral feature vector of the node adjacency relationship; The causal weight values ​​in the dynamic causal graph are embedded in the edge attribute matrix of the graph convolutional network, so that the causal knowledge participates in the message passing calculation.

8. The method for accurately locating distribution network faults based on graph convolutional neural networks according to claim 7 is characterized in that: A graph convolutional network with the dynamic graph structure as the network topology is constructed, and the operation parameters corresponding to the communication delay area are selected as the convolution kernel weights. The electrical quantity monitoring data, topological connection relationships, and causal reasoning knowledge of the associated nodes are processed and the fault coordinates are output, including: A hierarchical convolutional architecture is configured to adapt to the radial topology of the distribution network, including: implementing a third-order neighborhood aggregation operation on the feeder trunk; performing a first-order neighborhood convolution on the branch lines; Enhance fault feature propagation through causal weight values ​​in the edge attribute matrix; Build a three-dimensional fault analysis engine, including: longitude dimension: using the reverse graph diffusion algorithm to generate fault probability distribution; latitude dimension: locating distance based on line impedance parameters and fault recording data; altitude dimension: identifying fault phase based on zero-sequence current phase angle difference; Implementing parallel parameter loading and pipeline calculation on FPGA chip; The output includes line number, tower geographical coordinates and fault coordinates of the fault phase.

9. The method for accurately locating distribution network faults based on graph convolutional neural networks according to claim 8 is characterized in that: The construction of the three-dimensional fault analysis engine includes: When the fault probability distribution heat map shows high-probability points in multiple regions, the collaborative positioning mechanism is activated; Identify the power supply partition that contributes most to the fault current; fusing voltage phasor measurement unit data of adjacent zones in the longitude dimension to modify the fault probability distribution; In the latitude dimension, a library of historical similar fault cases is called to optimize the distance positioning method; In the height dimension, the zero-sequence impedance characteristics of the grounding transformer are combined to verify the phase identification results; Generate cross-partition collaborative handling instructions, including load transfer path optimization plans and protection device linkage strategies.

10. A distribution network fault accurate positioning system based on graph convolutional neural network, characterized in that: include: The distribution network monitoring module is used to deploy monitoring equipment at distribution network nodes to obtain topological connection relationships, electrical quantity monitoring data, environmental monitoring data, and new energy output forecast data; A dynamic graph structure generation module is used to respond to distributed power supply switching events and generate a dynamic graph structure based on a pre-stored simulation model; An operation parameter generation module is used to initialize graph convolution kernel parameters based on the dynamic graph structure and generate two types of operation parameters based on communication delay; A dynamic cause-effect graph construction module, configured to integrate the new energy output forecast data, the electrical quantity monitoring data, and the meteorological data to construct a dynamic cause-effect graph; The associated node data acquisition module is used to extract the electrical quantity monitoring data, topological connection relationship and causal reasoning knowledge of the associated nodes when a fault characteristic signal is detected; The fault coordinate output module is used to construct a graph convolution network with the dynamic graph structure as the network topology, select the operation parameters corresponding to the communication delay area as the convolution kernel weight, process the electrical quantity monitoring data, topological connection relationship and causal reasoning knowledge of the associated nodes, and output the fault coordinates.

Images