Power distribution network fault location method and system for distributed power supply access

Abstract

The invention discloses a distributed power supply access-oriented power distribution network fault distance measurement method and system, and relates to the technical field of power systems, and the method comprises the steps: collecting the electrical parameters and operation states of distributed power supply access nodes in a power distribution network in real time, building a dynamic manifold model based on an ecological niche theory, and carrying out the calculation of the dynamic manifold model; adaptively adjusting manifold learning neighborhood parameters according to power fluctuation data included in the electrical parameters, and updating node ecological niches to reconstruct a dynamic manifold model; based on the reconstructed dynamic manifold model, fault features are extracted from three scales of a current harmonic component, a feed line inter-harmonic propagation path and whole network voltage influence, and a three-dimensional feature vector is generated through fusion of a graph correlation algorithm; based on the expanded fault sample library and the power fluctuation data, constructing a fault transfer relation model to predict a ground fault and a short circuit risk area; a fault source is modeled by adopting a topological neural network, and a fault point distance measurement value is output through state prediction and strategy deduction.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power systems, more particularly to the technical field of intelligent power distribution system, power distribution switch control equipment manufacturing, and more particularly to a power distribution network fault location method and system for distributed power supply access. BACKGROUND

[0002] In the modern power distribution network operation system, fault location technology is a key link to ensure stable power supply. Accurate fault location can help operation and maintenance personnel quickly locate the fault position, greatly shorten the fault repair time, reduce the loss of power outage, and effectively improve the power supply reliability and user power experience. At the same time, it has important significance that cannot be ignored for optimizing power distribution network operation and maintenance strategy, reasonably allocating resources, and ensuring the safe and stable operation of the power system.

[0003] With the wide access of distributed power supply to the power distribution network, the traditional power distribution network fault location method is difficult to adapt to its complex and variable operating environment. Most of the traditional technologies are based on electrical parameters under steady-state operating conditions for fault analysis. After the access of distributed power supply, the power distribution network presents complex characteristics of multi-source power supply and bidirectional power flow. When facing the changes of system parameters caused by power fluctuations of distributed power supply, the traditional method is difficult to accurately capture fault characteristics, resulting in a decrease in the accuracy of fault location.

[0004] In view of the above problems, no effective solution has been proposed so far. SUMMARY

[0005] The embodiments of the present application provide a power distribution network fault location method and system for distributed power supply access to solve the above technical problems.

[0006] The present application provides a power distribution network fault location method for distributed power supply access, comprising:

[0007] Real-time collection of electrical parameters and operating conditions of distributed power supply access nodes in the power distribution network, construction of a dynamic manifold model based on the ecological niche theory, adaptive adjustment of manifold learning neighborhood parameters and updating of node ecological niche according to power fluctuation data included in the electrical parameters to reconstruct the dynamic manifold model;

[0008] Based on the reconstructed dynamic manifold model, fault features are extracted from current harmonic components, inter-feeder harmonic propagation paths, and overall network voltage influences in three scales, and a three-dimensional feature vector is generated by fusion through a graph association algorithm;

[0009] The three-dimensional feature vector is used to simulate distributed power supply fault scenarios, and the fault sample library is dynamically expanded;

[0010] Based on the expanded fault sample library and the power fluctuation data, a fault transmission relationship model is constructed to predict a grounding fault and a short-circuit risk area;

[0011] A space-time correlation matrix is established for fault features of three scales in the grounding fault and short-circuit risk area, and attention weights are dynamically allocated to locate a fault source;

[0012] The fault source is modeled by using a topological neural network, and a fault point distance measurement value is output by state prediction and strategy deduction.

[0013] Further, the dynamic manifold model is constructed based on the niche theory, the manifold learning neighborhood parameter is adaptively adjusted according to the power fluctuation data included in the electrical parameter, and the node niche is updated to reconstruct the dynamic manifold model, including:

[0014] The dynamic manifold model is constructed based on a niche overlap algorithm;

[0015] The manifold learning neighborhood parameter is adjusted according to the wind farm output power change rate in the power fluctuation data, specifically, the neighborhood range of the local linear embedding algorithm is increased, and the neighborhood range is represented by the number of neighborhood points;

[0016] The node niche is updated according to the charge and discharge states of the energy storage device in the operating state, the charging state node is defined as a voltage support niche, and the discharging state node is defined as a power regulation niche;

[0017] Based on the adjusted manifold learning neighborhood parameter and the updated node niche, the dynamic manifold model is reconstructed.

[0018] Further, the manifold learning neighborhood parameter is adjusted according to the wind farm output power change rate in the power fluctuation data, including: when the wind farm output power change rate exceeds a power change rate threshold, the number of neighborhood points of the local linear embedding algorithm is increased, and the neighborhood point increment is positively correlated with the wind farm output power change rate;

[0019] The node niche includes: assigning a voltage stability optimization coefficient to the voltage support niche; and assigning a power fluctuation suppression coefficient to the power regulation niche;

[0020] The dynamic manifold model is reconstructed, specifically: the increased number of neighborhood points is used as a spatial constraint condition, the voltage stability optimization coefficient and the power fluctuation suppression coefficient are used as node attribute weights; the local linear embedding algorithm is executed to generate a topological manifold structure adapted to distributed power access, and the topological manifold structure is the reconstructed dynamic manifold model.

[0021] Further, the fault features include distributed power harmonic anomaly and short-circuit fault features; based on the reconstructed dynamic manifold model, fault features are extracted from current harmonic component, inter-feeder harmonic propagation path and whole network voltage influence three scales, and a three-dimensional feature vector is generated by fusion through a graph correlation algorithm, including:

[0022] From the current harmonic component scale, current spectrum features are extracted based on the reconstructed dynamic manifold model, and the current spectrum features include specific frequency band amplitude corresponding to harmonic anomaly and fundamental distortion rate corresponding to short-circuit fault;

[0023] From the inter-feeder harmonic propagation path scale, path correlation degree features are extracted based on the reconstructed dynamic manifold model, and the path correlation degree features include path attenuation coefficient of harmonic propagation and diffusion radius of short-circuit fault current;

[0024] From the whole network voltage influence scale, voltage deviation rate features are extracted based on the reconstructed dynamic manifold model, and the voltage deviation rate features include steady-state voltage offset caused by harmonic and transient voltage drop depth caused by short-circuit fault;

[0025] Based on the path correlation degree features, a distribution network topology atlas is constructed, and nodes of the distribution network topology atlas correspond to distributed power access nodes, and edges correspond to feeder connection relationships;

[0026] The current spectrum features are mapped to node electrical properties;

[0027] The voltage deviation rate features are mapped to edge influence weights;

[0028] The shortest propagation path weights between nodes are calculated by Dijkstra algorithm;

[0029] The node fault influence is evaluated by PageRank algorithm combined with the node electrical properties and the edge influence weights;

[0030] The shortest propagation path weights and the node fault influence are fused to generate a graph correlation feature matrix;

[0031] The current spectrum features, the voltage deviation rate features and the graph correlation feature matrix are input into a fully connected neural network to generate a three-dimensional feature vector.

[0032] Further, based on the expanded fault sample library and the power fluctuation data, a fault transmission relationship model is constructed to predict grounding fault and short-circuit risk areas, including:

[0033] input the stereoscopic feature vector and the power fluctuation data into a dynamic Bayesian network to construct a fault transmission relationship model of meteorological parameters, equipment aging indicators, and electrical quantities, wherein the meteorological parameters include wind speed mutation gradient and illumination intensity decay rate, and the equipment aging indicators are characterized by impedance temperature rise coefficient and insulation medium loss;

[0034] The nonlinear dependence structure of the wind speed mutation gradient and the wind farm output power change rate is analyzed using the vine Copula function, and the time-varying influence of the equipment aging indicator on the feeder impedance is combined to calculate the fault transmission probability weight;

[0035] Based on the fault transmission probability weight, an association rule set of the equipment aging indicator and the harmonic abnormal feature is generated;

[0036] The mapping relationship between the short-circuit current probability density distribution and the ground fault transient feature is extracted by scanning the expanded fault sample library with a sliding time window;

[0037] The association rule set and the mapping relationship are fused, and a space-time Kriging interpolation algorithm is applied to generate a ground fault probability cloud chart and a short-circuit risk heat map, and output a spatial coordinate set of the ground fault and short-circuit risk area.

[0038] Further, the time-space association matrix of the fault features of three scales in the ground fault and short-circuit risk area is established, and the attention weight is dynamically allocated to locate the fault source, including:

[0039] A three-dimensional time-space tensor is constructed by acquiring the fault feature time sequence fragments of the current harmonic component scale, the inter-feeder harmonic propagation path scale, and the full-network voltage influence scale in the ground fault and short-circuit risk area;

[0040] High-order Tucker decomposition is performed on the three-dimensional time-space tensor to separate the spatial mode matrix, the time mode matrix, and the feature dimension matrix;

[0041] A fault propagation speed correction factor is generated based on the dynamic convolution calculation results of the feeder wave impedance and the equivalent harmonic impedance of the distributed power source, and is embedded into the time mode matrix;

[0042] The spatial mode matrix and the time mode matrix embedded with the fault propagation speed correction factor are recombined to form a time-space association matrix;

[0043] The time-space association matrix is decoded by a bidirectional gated recurrent unit to predict the decay trajectory of the fault feature in the topological manifold structure;

[0044] According to the curvature change rate and the voltage deviation rate mutation point of the decay trajectory, a spatial attention weight vector and a time attention weight vector are generated;

[0045] The weighted Mahalanobis distance algorithm is used to fuse the spatial attention weight vector and the time attention weight vector to locate the fault source coordinates, wherein the short-circuit fault source positioning preferentially activates the spatial weight of the voltage sag depth mutation node, and the harmonic abnormal fault source positioning preferentially activates the time weight of the path attenuation abnormal moment.

[0046] Further, the generation of the fault propagation speed correction factor comprises:

[0047] Capturing the time difference sequence of each feeder traveling wave front within 1.5 milliseconds before and after the fault occurrence moment;

[0048] Based on the topological manifold structure of the dynamic manifold model, the equivalent length matrix of the traveling wave path is calculated;

[0049] An anti-bias estimation model of time difference and length mapping is constructed, and abnormal time difference points caused by grid-connected inverter switching noise are removed;

[0050] Extracting the frequency characteristic curve of the feeder wave impedance; fitting a nonlinear equation of the fault propagation speed varying with the harmonic frequency;

[0051] The nonlinear equation is converted into a time dimension scaling coefficient embedded space-time correlation matrix generation process. Further, the fault source is modeled by the topological neural network, and the fault point distance measurement value is output by state prediction and strategy deduction, comprising:

[0052] A k-order topological neighborhood subgraph is constructed centered on the located fault source coordinates; k is a positive integer;

[0053] The topological neural network is used to model the fault evolution process, and the topological neural network is composed of a graph convolution layer and a gated recurrent unit in cascade;

[0054] A double-path strategy deduction is performed by a rolling horizon optimization algorithm, wherein the first path deduces the power regulation strategy with the energy storage device charge-discharge state switching as the decision variable, and the second path deduces the fault isolation strategy with the feeder segment switch combination sequence as the decision variable;

[0055] The weight distribution of the voltage support niche and the power regulation niche is updated according to the power regulation strategy in each optimization horizon;

[0056] The edge connection relationship of the topological neighborhood subgraph is dynamically pruned according to the fault isolation strategy;

[0057] The updated niche weight and subgraph structure are input into the time series graph attention network to predict the multi-step fault current diffusion trajectory and traveling wave front arrival time difference;

[0058] A differential mapping relationship between the fault current diffusion trajectory and the traveling wave head arrival time difference is established;

[0059] A minimum energy solution of the differential mapping relationship is determined to output a fault point distance measurement value.

[0060] Further, the double-path strategy deduction includes:

[0061] A conflict detection mechanism of the power regulation strategy and the fault isolation strategy is established, and a conflict flag is triggered when a discharge state node is required to provide voltage support and the node is located in a feeder segment marked as an isolation section;

[0062] A multi-objective optimization is started on a topological subgraph region covered by the conflict flag, wherein a first objective function minimizes a voltage deviation rate, and a second objective function maximizes a fault isolation success rate;

[0063] An optimal strategy sequence is selected through a Pareto frontier solution set;

[0064] Node attributes and edge connection weights of a topological neighborhood subgraph are updated according to the optimal strategy sequence;

[0065] The updated subgraph structure is fed back to a fault current diffusion trajectory prediction module for iterative optimization.

[0066] The application provides a power distribution network fault location system for distributed power source access, comprising:

[0067] A dynamic manifold model construction and reconstruction module is configured to collect electrical parameters and operating states of distributed power source access nodes in a power distribution network in real time, construct a dynamic manifold model based on a niche theory, and update node niches to reconstruct the dynamic manifold model by adaptively adjusting manifold learning neighborhood parameters according to power fluctuation data included in the electrical parameters;

[0068] A three-dimensional feature vector generation module is configured to extract fault features from current harmonic components, inter-feeder harmonic propagation paths, and overall network voltage influences based on the reconstructed dynamic manifold model, and generate a three-dimensional feature vector by fusion through a graph correlation algorithm;

[0069] A fault sample library expansion module is configured to simulate distributed power source fault scenarios by using the three-dimensional feature vector, and dynamically expand a fault sample library;

[0070] A risk area prediction module is configured to construct a fault transmission relationship model based on the expanded fault sample library and the power fluctuation data, to predict grounding fault and short circuit risk areas;

[0071] A fault source positioning module is configured to establish a space-time correlation matrix for fault features of three scales in the grounding fault and short circuit risk areas, and dynamically allocate attention weights to locate a fault source;

[0072] The fault point distance measurement value output module is configured to model the fault source by using a topological neural network, and output fault point distance measurement values through state prediction and strategy deduction.

[0073] Based on the embodiments provided in the present application, by collecting the electrical parameters and operating states of the distributed power supply access nodes in the power distribution network in real time, and constructing a dynamic manifold model based on the ecological niche theory, the actual working conditions of the power fluctuation of the distributed power supply can be fully considered, the manifold learning neighborhood parameters are adaptively adjusted and the node ecological niche is updated by using the power fluctuation data in the electrical parameters, the dynamic changes of the operating state of the power distribution network are accurately described, compared with the traditional steady-state analysis method, the problem of variable system parameters after the distributed power supply is accessed can be effectively solved. Based on the reconstructed dynamic manifold model, the fault characteristics are extracted from three scales of current harmonic components, harmonic propagation paths between feeders and voltage influences of the whole network, and a three-dimensional feature vector is generated by fusion through a graph correlation algorithm. This multi-scale fault feature extraction and fusion method can comprehensively capture the characteristics of the fault in different dimensions under the distributed power supply access scenario, and overcomes the defects that the existing technology cannot comprehensively consider the influence of multiple factors, provides a more comprehensive and accurate information basis for subsequent fault analysis, and thus the accuracy and reliability of fault location are significantly improved. BRIEF DESCRIPTION OF DRAWINGS

[0074] The drawings described herein are used to provide further understanding of the embodiments of the present application, and form a part of the present application.

[0075] Figure 1 A flowchart of an optional power distribution network fault location method for distributed power supply access according to an embodiment of the present application;

[0076] Figure 2 A flowchart of another optional power distribution network fault location method for distributed power supply access according to an embodiment of the present application;

[0077] Figure 3 A structure diagram of an optional power distribution network fault location system for distributed power supply access according to an embodiment of the present application.

[0078] The implementation, functional characteristics and advantages of the present application will be further described with reference to the embodiments and the accompanying drawings. DETAILED DESCRIPTION

[0079] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0080] Optionally, as shown in the present application, a power distribution network fault location method for distributed power supply access is provided, comprising: Figure 1

[0081] S101, real-time collection of electrical parameters and operating states of distributed power supply access nodes in a power distribution network, construction of a dynamic manifold model based on a niche theory, adaptive adjustment of manifold learning neighborhood parameters and updating of node niches according to power fluctuation data included in the electrical parameters to reconstruct the dynamic manifold model;

[0082] The electrical parameters are physical quantities representing power transmission and conversion in the power distribution network, and are used to quantify the system operating state, including: voltage (such as 10kV feeder voltage amplitude), current (photovoltaic grid-connected current effective value), active power (wind farm output power), reactive power (capacitor bank compensation amount), frequency (power grid frequency 50Hz), harmonic content (current 3rd harmonic distortion rate).

[0083] The operating state is the working mode and device state of the distributed power supply access node, reflecting the real-time operating condition, including: temperature, photovoltaic inverter grid-connected / off-grid state, energy storage battery charging / discharging mode, transformer tap position, circuit breaker on / off state, and power regulation mode (constant power / constant voltage) of the distributed power supply.

[0084] The manifold learning neighborhood parameter is a key parameter in the manifold learning algorithm for describing the local neighborhood relationship of data points, and determines the fitting accuracy of the low-dimensional manifold to the high-dimensional data. For example, in the local linear embedding (LLE) algorithm, the "number of neighborhood points k" (such as k=8, i.e. each voltage fluctuation data point refers to the correlation of the nearest 8 current data points) and the "neighborhood distance threshold ε" (such as ε=0.1V, defining the range of adjacent voltage fluctuation points). These parameters are dynamically adjusted with power fluctuation data to adapt to the nonlinear data distribution after the access of distributed power supply.

[0085] S102, based on the reconstructed dynamic manifold model, extracting fault features from current harmonic components, inter-feeder harmonic propagation paths and whole-network voltage influences in three scales, and generating a three-dimensional feature vector by graph association algorithm; S103, simulating a distributed power supply fault scenario by using the three-dimensional feature vector, and dynamically expanding a fault sample library;

[0086] ​S104: Based on the expanded fault sample library and power fluctuation data, a fault transmission relationship model is constructed to predict ground fault and short circuit risk areas;

[0087] S105, establishing a spatiotemporal correlation matrix for the fault characteristics of three scales in the ground fault and short circuit risk area and dynamically assigning attention weights to locate the fault source;

[0088] S106, using a topological neural network to model the fault source, and outputting a distance measurement value of the fault point through state prediction and strategy deduction.

[0089] Based on the embodiments provided by the present application, by collecting the electrical parameters and operating status of the distributed power access nodes in the distribution network in real time and constructing a dynamic manifold model based on the niche theory, it is possible to fully consider the actual operating conditions of the distributed power supply power fluctuations, and use the power fluctuation data in the electrical parameters to adaptively adjust the manifold learning neighborhood parameters and update the node niche, so as to achieve an accurate characterization of the dynamic changes in the distribution network operating state. Compared with the traditional steady-state analysis method, it can effectively deal with the problem of variable system parameters after the distributed power supply is connected. Based on the reconstructed dynamic manifold model, fault features are extracted from three scales: current harmonic components, inter-feeder harmonic propagation paths, and the impact of the entire network voltage. A three-dimensional feature vector is generated by fusing the graph association algorithm. This multi-scale fault feature extraction and fusion method can fully capture the characteristics of faults in different dimensions under the distributed power supply access scenario, overcome the defect that the existing technology cannot comprehensively consider the influence of multiple factors, and provide a more comprehensive and accurate information basis for subsequent fault analysis, thereby significantly improving the accuracy and reliability of fault distance measurement.

[0090] Further, if Figure 2 As shown in FIG, a dynamic manifold model is constructed based on the niche theory. The manifold learning neighborhood parameters are adaptively adjusted according to the power fluctuation data included in the electrical parameters and the node niche is updated to reconstruct the dynamic manifold model, including:

[0091] S201, constructing a dynamic manifold model based on the niche overlap algorithm;

[0092] The niche overlap algorithm, based on the concept of species niches in ecology, quantifies the degree of overlap in the power output characteristics of different distributed power sources, reflecting their complementarity or competitiveness. For example, the overlap area between the power curves of a wind farm and a photovoltaic power station over a 24-hour period is calculated. If the photovoltaic power generation is at full capacity during the summer daytime and the wind power output is low, the overlap is low (strong complementarity); if both power generation are low during a certain period, the overlap is high (poor power supply stability). Overlap = Overlap Area / (Total Wind Farm + Photovoltaic Power Coverage Area).

[0093] S202, adjusting a manifold learning neighborhood parameter according to a wind farm output power change rate in the power fluctuation data, specifically, increasing a neighborhood range of a locally linear embedding algorithm, the neighborhood range being represented by a number of neighborhood points;

[0094] In the embodiment, the wind farm output power change rate is a change amount of the wind farm output power in a unit time, used to measure the power fluctuation speed.

[0095] The locally linear embedding algorithm is a manifold learning algorithm, which realizes dimension reduction by maintaining local linear relationship of data, and is suitable for extracting manifold features of high-dimensional electrical parameters of a power grid.

[0096] The energy storage device includes but is not limited to a lithium battery energy storage system (response time ms level, suppressing short-time power fluctuation), a flywheel energy storage (high power density, coping with high-frequency charging and discharging), a super capacitor (instantaneous power compensation, suppressing voltage sag), and a compressed air energy storage (large capacity, daily peak regulation).

[0097] S204, reconstructing the dynamic manifold model based on the adjusted manifold learning neighborhood parameter and the updated node niche.

[0098] Based on the embodiments provided in the application, the dynamic manifold model is constructed based on the niche overlap algorithm, the manifold learning neighborhood parameter is adjusted according to the wind farm output power change rate, and the node niche is updated according to the charging and discharging state of the energy storage device, so that the dynamic manifold model can deeply integrate the actual operation characteristics of the distributed power source. Specifically, for wind farm power fluctuation, adjusting the neighborhood range can better capture the data change trend; according to the different states of the energy storage device, the node is given a specific niche, and its function positioning in the system is clear, thereby realizing more accurate and more practical modeling of the operation state of the distribution network, and effectively solving the problem that the traditional model is difficult to adapt to the dynamic changes of the distributed power source.

[0099] Further, adjusting the manifold learning neighborhood parameter according to the wind farm output power change rate in the power fluctuation data includes: when the wind farm output power change rate exceeds a power change rate threshold, increasing the number of neighborhood points of the locally linear embedding algorithm, the neighborhood point increment being positively correlated with the wind farm output power change rate.

[0100] The power change rate threshold includes but is not limited to 5% of the rated power, 10% of the rated power, etc.

[0101] The updating of the node niche includes: assigning a voltage stability optimization coefficient to the voltage support niche, improving the voltage dimension weight, and strengthening the suppression of voltage fluctuation; and assigning a power fluctuation suppression coefficient to the power regulation niche to suppress the power mutation of the distributed power supply.

[0102] The reconstructed dynamic manifold model is specifically: taking the increased number of neighborhood points as a spatial constraint condition, and taking the voltage stability optimization coefficient and the power fluctuation suppression coefficient as node attribute weights; and performing a local linear embedding algorithm to generate a topological manifold structure adapted to the access of the distributed power supply, and the topological manifold structure is the reconstructed dynamic manifold model.

[0103] In a specific embodiment, the formula of the dynamic manifold model reconstruction is:

[0104]

[0105] wherein M re is the reconstructed dynamic manifold model, used to abstractly represent the operating state topological structure of the distribution network after the access of the distributed power supply; LLE() is a local linear embedding algorithm, which is a core algorithm for constructing the manifold model; is the adjusted number of neighborhood points, in units of pieces, which is dynamically adjusted according to the wind farm output power change rate, and when the wind farm output power change rate exceeds a set threshold, the number of neighborhood points is increased by ΔN wherein ΔN is positively correlated with the power change rate; ω v is a voltage stability optimization coefficient, with a value range of [0.6, 0.8], used to enhance the attention of the model to the voltage stability dimension, and the value can be 0.8 when the voltage fluctuation is large; ω p is a power fluctuation suppression coefficient, with a value range of [0.2, 0.4], used to highlight the processing of the power fluctuation dimension, and the value can be 0.4 when the power fluctuation is severe.

[0106] Based on the above formula, the model parameters are adjusted based on the niche theory, and by integrating the dynamically changing number of neighborhood points and the voltage and power related weight coefficients into the local linear embedding algorithm, the reconstructed dynamic manifold model can accurately adapt to the complex operating scenarios of the power fluctuation and equipment state change of the distribution network after the access of the distributed power supply. Compared with the traditional fixed-parameter manifold model, it can dynamically adjust the data fitting range and feature weight, more accurately capture the correlation between electrical parameters, lay a solid foundation for subsequent extraction of fault features and construction of fault transmission relationship model, and effectively improve the description ability of the model to the dynamic characteristics of the distribution network and the capture accuracy of the fault features.

[0107] The local linear embedding algorithm is executed to generate a topological manifold structure, including: dimension reduction of power distribution network topological data (including distributed power supply access points) by the local linear embedding algorithm to generate a low-dimensional manifold structure reflecting the electrical correlation of nodes. For example, after the access of wind power, the local linear embedding algorithm maps 100-dimensional parameters such as voltage and current to a 2-dimensional manifold, and the distance between nodes reflects the electrical coupling strength (for example, the wind power access point is close to the adjacent load node, indicating that the two have strong voltage fluctuation correlation), which is used to identify topological weak points (for example, a node is isolated in the manifold, which is easy to cause a chain failure).

[0108] Based on the embodiments provided in the present application, when the wind farm output power change rate exceeds the threshold value, the number of neighborhood points is increased in a positive correlation, the specific coefficients for different ecological niches are allocated, and they are integrated into the dynamic manifold model construction process. This precise quantitative adjustment method can expand the data capture range of the model when the power fluctuates violently, and by assigning different functional weights to the nodes through the coefficients, the reconstructed topological manifold structure more accurately reflects the operation rules of the power distribution network after the access of distributed power supply, greatly enhancing the adaptability and representation ability of the model to complex operation conditions.

[0109] Further, the fault features include distributed power supply harmonic abnormality and short-circuit fault features; based on the reconstructed dynamic manifold model, fault features are extracted from current harmonic components, harmonic propagation paths between feeders, and network voltage influences on three scales, and a three-dimensional feature vector is generated by fusion through a graph association algorithm, including:

[0110] From the current harmonic component scale, the current frequency spectrum features are extracted based on the reconstructed dynamic manifold model, and the current frequency spectrum features include the specific frequency band amplitude corresponding to the harmonic abnormality and the fundamental wave distortion rate corresponding to the short-circuit fault;

[0111] From the harmonic propagation path scale between feeders, the path correlation degree features are extracted based on the reconstructed dynamic manifold model, and the path correlation degree features include the path attenuation coefficient of harmonic propagation and the diffusion radius of short-circuit fault current;

[0112] From the network voltage influence scale, the voltage deviation rate features are extracted based on the reconstructed dynamic manifold model, and the voltage deviation rate features include the steady-state voltage offset caused by harmonics and the transient voltage drop depth caused by short-circuit faults;

[0113] A power distribution network topology atlas is constructed based on the path correlation degree features, and the nodes of the power distribution network topology atlas correspond to the distributed power supply access nodes, and the edges correspond to the feeder connection relationship;

[0114] The feeder connection relationship is the physical connection topology of the feeder and the power supply, load, transformer and other devices in the power distribution network, which determines the power transmission path.

[0115] The current frequency spectrum features are mapped to node electrical properties;

[0116] mapping voltage deviation rate features to edge influence weights;

[0117] calculating shortest propagation path weights between nodes by Dijkstra algorithm;

[0118] evaluating node fault influence by PageRank algorithm combined with node electrical properties and edge influence weights;

[0119] Dijkstra algorithm: calculating the shortest path between two points in a graph (e.g., calculating the minimum impedance path for fault current from wind farm to busbar for fault location);

[0120] PageRank algorithm: evaluating the importance of nodes in a graph (e.g., considering distribution network topology as a web link graph, the higher the PageRank value of a photovoltaic access point, the greater the influence weight of its fault on the voltage of the entire network)

[0121] fusing shortest propagation path weights and node fault influence to generate graph correlation feature matrix;

[0122] inputting current spectrum features, voltage deviation rate features, and graph correlation feature matrix into fully connected neural network to generate a three-dimensional feature vector.

[0123] Fully connected neural network generating a three-dimensional feature vector: through a multi-layer neural network, multi-dimensional input data is nonlinearly mapped to generate a vector containing comprehensive fault features. For example, input layer (e.g., 10 features: voltage deviation, current harmonic, path attenuation, etc.) → hidden layer (2 layers, 50 nodes per layer, ReLU activation) → output layer (5-dimensional three-dimensional feature vector, e.g., [harmonic anomaly index, short-circuit risk value, voltage stability, topology propagation efficiency, fault influence]).

[0124] In a specific embodiment, the three-dimensional feature vector generation formula is:

[0125] V s = FNN(CSF, VDF, GCF)

[0126] where V s is the generated three-dimensional feature vector, integrating multi-dimensional feature information of distribution network faults; FNN() is a fully connected neural network used to realize the fusion of fault features of different scales; CSF is the current spectrum feature; VDF is the voltage deviation rate feature; and GCF is the graph correlation feature matrix.

[0127] Based on the above formula, the technical content of extracting and fusing fault features from multiple scales, using a fully connected neural network to deeply fuse fault features of current harmonic components, inter-feeder harmonic propagation paths, and full-network voltage influences at three scales to generate a three-dimensional feature vector containing rich fault information. In the scenario of distributed power supply access leading to complex and diversified fault features, compared with single-scale or simply combined fault feature analysis methods, this formula can comprehensively and systematically integrate electrical quantity and topological structure information, effectively capturing the performance of faults in different dimensions. The generated three-dimensional feature vector can be used to subsequently build a fault transmission relationship model and predict fault risk areas, significantly improving the accuracy and comprehensiveness of fault diagnosis, and providing a more representative and comprehensive feature expression for fault analysis.

[0128] Based on the embodiments provided in the present application, fault features are extracted in detail from three scales, and a feature vector is generated by using a graph correlation algorithm and a neural network. In the current harmonic component scale, current frequency spectrum features are extracted, in the inter-feeder harmonic propagation path scale, path correlation degree features are extracted, and in the full-network voltage influence scale, voltage deviation rate features are extracted, comprehensively covering the performance of faults in different dimensions. By constructing a topological graph, the features are mapped as node attributes and edge weights, and then a variety of algorithms are used to evaluate the fault influence of the nodes. The finally generated three-dimensional feature vector integrates multi-dimensional information, can comprehensively and meticulously depict fault characteristics, and provides a richer and more accurate information basis for subsequent fault analysis, significantly improving the extraction and analysis ability of fault features.

[0129] Further, based on the expanded fault sample library and power fluctuation data, a fault transmission relationship model is constructed to predict grounded fault and short circuit risk areas, including:

[0130] The three-dimensional feature vector and power fluctuation data are input into a dynamic Bayesian network to construct a fault transmission relationship model of meteorological parameters, equipment aging indicators, and electrical quantities, wherein the meteorological parameters include wind speed mutation gradient and illumination intensity decay rate, and the equipment aging indicators are represented by impedance temperature rise coefficient and insulation dielectric loss;

[0131] The nonlinear dependence structure of wind speed mutation gradient and wind farm output power change rate is analyzed using the vine Copula function, and the time-varying influence of the equipment aging indicator on the feeder impedance is calculated to calculate the fault transmission probability weight;

[0132] The vine Copula function is an algorithm for handling multivariate correlation, which decomposes high-dimensional joint distribution through a "vine" tree structure, and is suitable for analyzing the dependence relationship between distributed power supply and meteorological parameters. For example, when analyzing the correlation of wind speed, illumination intensity, and wind farm power, the vine Copula is decomposed as follows: ① wind speed-illumination Copula; ② power-(wind speed+illumination) conditional Copula, and then the probability of wind farm power sudden drop under extreme weather conditions (such as strong wind + dark clouds) is predicted.

[0133] generate a set of association rules between the device aging indicators and the harmonic abnormal features based on the fault transmission probability weights;

[0134] The harmonic abnormal feature refers to an abnormal parameter or phenomenon in the power signal due to the harmonic component (current or voltage component with an integer multiple of the fundamental frequency) exceeding the normal range. These features are usually characterized by harmonic content, distortion rate, amplitude fluctuation, phase shift, etc., and can be used to identify power equipment failure, abnormal distributed power supply access, or unbalanced power grid operation state.

[0135] The mapping relationship between the short-circuit current probability density distribution and the ground fault transient characteristics is extracted by scanning the expanded fault sample library with a sliding time window;

[0136] The association rule set and the mapping relationship are fused and the spatiotemporal Kriging interpolation algorithm is applied to generate a ground fault probability cloud map and a short-circuit risk heat map, and output a spatial coordinate set of the ground fault and short-circuit risk area.

[0137] The spatiotemporal Kriging interpolation algorithm is an interpolation method that combines spatial position and time series correlation, used to predict the spatiotemporal distribution of power grid parameters. For example, given the voltage sag values at points A, B, and C at a certain time, the spatiotemporal Kriging is used to predict the voltage at point D (considering the spatial distance between D and A / B / C, and the time correlation of historical sag data), to generate a spatiotemporal cloud map of the entire network voltage sag, to assist in locating the fault source position and impact range.

[0138] Based on the embodiments provided in the present application, the dynamic Bayesian network is used to construct the fault transmission relationship model, the nonlinear dependence structure is analyzed by the vine Copula function, the association rule set and the mapping relationship are generated, and the risk area is output by the spatiotemporal Kriging interpolation algorithm. This scheme fully considers the complex mutual relationship between meteorological parameters, device aging indicators and electrical quantities, can mine the influence law of each factor on fault transmission, accurately predicts the ground fault and short-circuit risk area, provides a strong decision basis for operation and maintenance personnel to take preventive measures in advance and reasonably arrange maintenance plans, and effectively improves the fault prevention and risk control ability of the distribution network.

[0139] Further, a spatiotemporal association matrix is established for the three scales of fault features in the ground fault and short-circuit risk area, and attention weights are dynamically allocated to locate the fault source, including:

[0140] A three-dimensional spatiotemporal tensor is constructed by acquiring the fault feature time sequence segments of the current harmonic component scale, the inter-feeder harmonic propagation path scale, and the full-network voltage influence scale in the ground fault and short-circuit risk area;

[0141] Performing high-order Tucker decomposition on the three-dimensional space-time tensor separates the spatial mode matrix, the time mode matrix, and the feature dimension matrix;

[0142] In this embodiment, high-order Tucker decomposition: splits the multi-dimensional electrical data in the power grid (such as voltage and current data of different time, location, and parameter type) into core features and basic factors, like disassembling a complex power grid operation panoramic picture into several key "puzzle modules", to facilitate the extraction of key information related to faults. For example, in a distribution network, a three-dimensional data block containing "voltage data of each feeder within 10 ms after the fault occurs", "node location", and "harmonic frequency" is decomposed by high-order Tucker decomposition to extract three types of core features: "the most significant time point of voltage sag", "the most seriously affected feeder section", and "the key harmonic frequency component", which are used for subsequent fault analysis.

[0143] Generate a fault propagation speed correction factor based on the dynamic convolution calculation results of the feeder wave impedance and the equivalent harmonic impedance of the distributed power source, and embed it in the time mode matrix;

[0144] In this embodiment, the feeder wave impedance is the resistance of the feeder (transmission line) to the traveling wave (electromagnetic fluctuation generated at the time of fault), similar to the resistance of a water pipe to water flow. The greater the wave impedance, the more obvious the energy loss and waveform distortion of the traveling wave during propagation. For example, when a short-circuit fault occurs on a feeder in a distribution network, the traveling wave propagates from the fault point to both ends. Different wave impedances of different feeders will result in different times and waveforms of the traveling wave reaching the measurement point. For example, a high wave impedance feeder will slow down the propagation speed of the traveling wave, and the time of reaching the monitoring point will be later than the theoretical value.

[0145] Equivalent harmonic impedance of distributed power source: treat the distributed power source (such as photovoltaic and wind power) as a "virtual resistor" at harmonic frequency. The size of this resistor reflects the power source's ability to affect the grid's harmonics - a small impedance is easy to inject harmonics into the grid, and a large impedance has a restraining effect on harmonics. For example, a photovoltaic inverter generates 5th and 7th harmonics when it is working. If its equivalent harmonic impedance is small (e.g., 0.5 Ω), these harmonics are easy to enter the grid, causing voltage distortion of other equipment. If the equivalent harmonic impedance is increased through control strategies (e.g., adjusted to 2 Ω), the harmonic injection will decrease.

[0146] The dynamic convolution generates a fault propagation speed correction factor. According to the real-time operation state of the power grid (such as distributed power fluctuation, topology change), the "filter" parameter of data processing is dynamically adjusted. The deviation coefficient of the actual fault propagation speed and the theoretical value is calculated to correct the time calculation of fault location. For example, when the power of the wind farm suddenly increases, the harmonic component in the power grid changes, causing the traveling wave propagation speed to be different from the ideal situation. By dynamically convolving the historical traveling wave data, it is found that the actual propagation speed is 10% slower than the theoretical value, and a correction factor of 0.9 is generated. In subsequent fault location, the calculated distance is multiplied by 0.9 to improve the positioning accuracy.

[0147] The space-time correlation matrix is formed by the recombination of the space mode matrix and the time mode matrix embedded with the fault propagation speed correction factor;

[0148] The space-time correlation matrix is decoded by a bidirectional gated recurrent unit to predict the decay trajectory of the fault feature in the topological manifold structure;

[0149] The bidirectional gated recurrent unit (Bi-GRU) is a time series data processing model that can simultaneously "look forward" and "look backward", which is used to analyze the law of fault signal change over time, just like finding the key frame by playing a video forward and backward. For example, input the voltage time series data of each feeder (such as voltage change from 1ms to 5ms after fault), Bi-GRU simultaneously analyzes the signal change "from fault occurrence to current time" and "from current time to fault occurrence", and decodes the "voltage drop start time" and "the accurate order of wave head arrival at each feeder", providing key information in time dimension for fault location.

[0150] The space attention weight vector and the time attention weight vector are generated according to the curvature change rate of the decay trajectory and the voltage deviation rate mutation point;

[0151] The space attention weight vector and the time attention weight vector are fused by using the weighted Mahalanobis distance algorithm to locate the fault source coordinates, wherein the short-circuit fault source positioning preferentially activates the space weight of the voltage drop depth mutation node, and the harmonic abnormal fault source positioning preferentially activates the time weight of the path attenuation abnormal time.

[0152] The weighted Mahalanobis distance algorithm is a "data similarity measurement tool" that considers the importance of different electrical parameters, similar to setting different scores for different test subjects, and calculating the matching degree of new data and historical fault data. For example, in fault location, the harmonic distortion rate is more important for judging distributed power faults, while the fundamental voltage is more important for judging short-circuit faults. The weighted Mahalanobis distance will assign a higher weight (such as 0.6) to the harmonic distortion rate and 0.4 to the fundamental voltage, and calculate the similarity of the new fault data and the "photovoltaic inverter fault" sample in the sample library to assist in locating the fault type.

[0153] In an alternative embodiment, the fault source positioning formula is:

[0154] P f = WMD(SAW ￠ TSF, TAW ￠ TSF)

[0155] wherein P f is the fault source coordinates, used to determine the specific spatial location of the fault occurrence in the power distribution network, is the final output result of fault location; WMD() is a weighted Mahalanobis distance algorithm, used to measure the similarity between the fault feature vector and the known fault mode, considering the correlation between the features and giving different weights to different dimensions; SAW is a spatial attention weight vector, with a value range of [0, 1], generated according to the distribution of the fault feature in space, used to highlight the areas significantly affected by the fault (such as the nodes or feeder sections with the most severe voltage sag); TAW is a time attention weight vector, with a value range of [0, 1], generated according to the change of the fault feature in the time series, used to emphasize the features at the key time points of the fault occurrence (such as the arrival time of the traveling wave, the moment of voltage mutation); TSF is the fault feature vector decoded by the bidirectional gated recurrent unit (Bi-GRU), including: electrical quantity features (such as current amplitude in amperes A, voltage in volts V); time features (such as fault occurrence time in seconds s); ￠ is a vector element corresponding multiplication operation, used to multiply the attention weight and the fault feature vector element by element, to realize the weighted enhancement of the key information.

[0156] Based on the above formula, the technical scheme for locating the fault source based on the establishment of the spatio-temporal correlation matrix and the allocation of attention weights, by combining the spatial and temporal attention weight vectors with the fault feature vector, realizes the accurate positioning of the fault source using the weighted Mahalanobis distance algorithm. In the power distribution network with distributed power sources, the fault propagation presents complex spatio-temporal characteristics, and different types of faults have different performances in the time-space dimension. This formula can dynamically allocate weights according to the spatio-temporal importance of the fault features, giving priority to the spatial weight of the voltage sag mutation node for short-circuit faults, and giving priority to the time weight of the path attenuation anomaly time for harmonic abnormal faults, effectively improving the accuracy and pertinence of the positioning compared with traditional fault location methods. By dynamically calculating the fault propagation speed correction factor through convolution, combining the feeder wave impedance and the equivalent harmonic impedance of the distributed power source, the defect of assuming constant wave speed in traditional traveling wave positioning is corrected, which is especially suitable for scenarios where the wave speed changes due to the access of distributed power sources; at the same time, the robustness of the positioning is improved by using the robust estimation model to eliminate abnormal data points, and the fault features are dynamically updated in each time window based on the rolling time domain optimization, which adapts to the dynamic changes of the power distribution network topology, realizing the fine positioning of the spatio-temporal location of the fault source, providing key technical support for rapid fault handling and ensuring the safe and stable operation of the power distribution network.

[0157] Based on the embodiments provided in the present application, the fault source is located by establishing a space-time correlation matrix of fault features in the fault risk area and dynamically allocating attention weight. A three-dimensional space-time tensor is constructed and decomposed, the fault propagation speed correction factor is calculated by combining the feeder wave impedance and the equivalent harmonic impedance of the distributed power supply, the matrix is reorganized to predict the fault feature decay trajectory through the bidirectional gate recurrent unit, and finally the fault source is located based on the trajectory and voltage deviation rate. This process fully considers the change law of fault features in the time-space dimension, differentiates the weight for different fault types (short circuit fault, harmonic abnormal fault), realizes the accurate positioning of the fault source, and effectively solves the problem of fault source positioning in complex distribution network environment.

[0158] Further, the generation of the fault propagation speed correction factor includes:

[0159] Capture the time difference sequence of each feeder wave head within 1.5 milliseconds before and after the fault occurs;

[0160] In the present embodiment, the time difference sequence of each feeder wave head is a group of data arranged in order according to the time difference of the fault generated wave reaching different feeder measurement points, which is the core basis for fault location, similar to the time difference of seismic waves reaching different monitoring stations for locating the earthquake source. For example, a certain distribution network has three monitoring points A, B and C, the wave reaches point A at 10 ms, point B at 12 ms, and point C at 15 ms, then the time difference sequence is [B-A=2 ms, C-B=3 ms, C-A=5 ms], and the fault point position can be deduced from these time differences.

[0161] Based on the topological manifold structure of the dynamic manifold model, the equivalent length matrix of the wave propagation path is calculated;

[0162] An anti-difference estimation model is constructed to map the time difference and length, and the RANSAC algorithm is used to remove abnormal time difference points caused by grid-connected inverter switching noise;

[0163] In the present embodiment, the anti-difference estimation model is used as a "filter" to exclude abnormal value interference in the data, avoiding the misdirection of noise (such as interference signals generated by inverter switching) to fault analysis. For example, when an abnormal time difference (such as actually 2 ms, but measured as 10 ms) caused by noise is mixed in the collected wave arrival time, the anti-difference estimation model will identify that the value is too different from other data, mark it as an abnormal value and remove it, ensuring that the time difference data used for positioning is accurate.

[0164] The frequency characteristic curve of the feeder wave impedance is extracted by wavelet packet decomposition; a nonlinear equation of the fault propagation speed changing with the harmonic frequency is fitted;

[0165] The nonlinear equation is converted into a time dimension scaling coefficient embedded in the space-time correlation matrix generation process.

[0166] Based on the embodiments provided in the present application, in the process of generating the fault propagation speed correction factor, the time difference sequence of the arrival of the traveling wave is captured, the equivalent path length matrix is calculated, the outlier is removed by constructing a robust estimation model, and the nonlinear relationship between the fault propagation speed and the harmonic frequency is fitted to convert it into a scaling coefficient embedded in the space-time correlation matrix. This series of operations can effectively exclude noise interference, accurately consider the influence of harmonic frequency on fault propagation speed, make the constructed space-time correlation matrix more accurately reflect the actual situation of fault propagation, provide a more reliable basis for fault source positioning, and improve the accuracy and stability of fault positioning.

[0167] Further, a topological neural network is used to model the fault source, and a fault point distance measurement value is output by state prediction and strategy deduction, including:

[0168] A k-order topological neighborhood subgraph is constructed with the positioned fault source coordinates as the center;

[0169] A topological neural network is used to model the fault evolution process, and the topological neural network is composed of a graph convolution layer and a gated recurrent unit in cascade; the graph convolution layer dynamically adjusts the adjacency matrix according to the weight distribution of the voltage support niche and the power regulation niche;

[0170] A double-path strategy deduction is performed by a rolling horizon optimization algorithm, wherein the first path deduces a power regulation strategy with the state switching of the energy storage device as the decision variable, and the second path deduces a fault isolation strategy with the combination sequence of the feeder section switches as the decision variable;

[0171] The rolling horizon optimization algorithm divides the fault handling process into multiple short time stages, optimizes the control strategy according to the current state in each stage, and updates the strategy according to the opponent's reaction, similar to considering only the optimal move in the next few steps when playing chess. For example, in fault ranging, the first stage (0-10ms) optimizes the discharge strategy of the energy storage device to stabilize the voltage, and the second stage (10-20ms) optimizes the on-off strategy of the feeder switch to isolate the fault. The optimization of each stage is based on the results of the previous stage to dynamically adjust the strategy to ensure real-time effectiveness.

[0172] The weight distribution of the voltage support niche and the power regulation niche is updated according to the power regulation strategy in each optimization time domain;

[0173] In this embodiment, the optimization time domain is the time range of each independent optimization stage in the rolling horizon optimization, similar to the "time period granularity" when making plans. For example, the optimization time domain is set to 5ms, that is, the strategy optimization is performed every 5ms: 0-5ms optimizes the power regulation of the power supply, 5-10ms optimizes the switch action, and 10-15ms optimizes the fault positioning algorithm parameters. The decision of each time domain only focuses on the optimal solution in the current short time.

[0174] dynamically pruning the edge connection relationship of the topological neighborhood subgraph according to the fault isolation strategy;

[0175] inputting the updated ecological niche weight and the subgraph structure into a time series graph attention network to predict the multi-step fault current diffusion trajectory and the arrival time of the traveling wave front;

[0176] establishing a differential mapping relationship between the fault current diffusion trajectory and the arrival time difference of the traveling wave front;

[0177] solving the minimum energy solution of the differential mapping relationship by a particle swarm optimization algorithm to output the fault point distance measurement.

[0178] Based on the embodiments provided in the present application, the fault source is modeled by a topological neural network, the fault source coordinates are located to construct a topological neighborhood subgraph, the fault current diffusion trajectory and the arrival time of the traveling wave are predicted by a double-path strategy deduction and a time series graph attention network, and then the fault point distance is determined. This scheme combines topological structure and neural network, can effectively simulate the evolution process of the fault in the distribution network, considers the influence of different control and isolation strategies through the double-path strategy deduction, updates the model parameters according to the strategy, realizes the accurate prediction of the fault point distance, provides key distance information for fault handling, and improves the accuracy and practicality of fault distance measurement.

[0179] Further, the double-path strategy deduction includes:

[0180] a conflict detection mechanism of the power control strategy and the fault isolation strategy is established, and a conflict flag is triggered when a discharge state node is required to provide voltage support and the feeder segment where the node is located is marked as an isolation segment;

[0181] NSGAII multi-objective optimization is started on the topological subgraph area covered by the conflict flag, wherein a first objective function minimizes the voltage deviation rate, and a second objective function maximizes the fault isolation success rate;

[0182] The NSGAII multi-objective optimization algorithm is an optimization tool for handling multiple conflicting objectives, such as simultaneously pursuing “the highest voltage stability” and “the fastest fault isolation speed”, finding the best balance point between the two, similar to finding the optimal solution between “performance” and “battery life” when choosing a mobile phone. For example, when the power control strategy requires the energy storage device to discharge to stabilize the voltage, and the fault isolation strategy requires the feeder containing the energy storage to be disconnected, NSGAII will simultaneously optimize the two objectives of “minimizing the voltage deviation rate” and “maximizing the fault isolation success rate” to generate a set of Pareto optimal solutions (e.g., voltage deviation rate 3%, isolation success rate 95% or voltage deviation rate 5%, isolation success rate 100%) for the system to choose the most suitable strategy.

[0183] The optimal strategy sequence is selected through the Pareto front solution set;

[0184] updating the node attribute and edge connection weight of the topology neighborhood subgraph according to the optimal strategy sequence;

[0185] feeding back the updated subgraph structure to the fault current diffusion trajectory prediction module for iterative optimization.

[0186] Based on the embodiments provided in the present application, a conflict detection mechanism is established in the double-path strategy deduction. When there is a conflict between the power regulation and fault isolation strategies, the multi-objective optimization is started to select the optimal strategy sequence, and the topology subgraph structure is updated. This mechanism can effectively solve the contradictions between strategies, maximize the fault isolation success rate, ensure that the distribution network can operate efficiently and stably during fault handling, and improve the strategy optimization and coordinated control ability of the distribution network when dealing with faults.

[0187] Optionally, as shown in Figure 3 The present application provides a distribution network fault location system for distributed power supply access, comprising:

[0188] The dynamic manifold model construction and reconstruction module 301 is used to collect the electrical parameters and operating state of the distributed power supply access node in the distribution network in real time, construct a dynamic manifold model based on the ecological niche theory, and update the node ecological niche to reconstruct the dynamic manifold model by adaptively adjusting the manifold learning neighborhood parameters according to the power fluctuation data included in the electrical parameters.

[0189] The three-dimensional feature vector generation module 302 is used to extract fault features from current harmonic components, inter-feeder harmonic propagation paths, and overall network voltage influences based on the reconstructed dynamic manifold model, and generate a three-dimensional feature vector by fusion through a graph correlation algorithm.

[0190] The fault sample library expansion module 303 is used to simulate distributed power supply fault scenarios using the three-dimensional feature vector, and dynamically expand the fault sample library.

[0191] The risk area prediction module 304 is used to construct a fault transmission relationship model based on the expanded fault sample library and power fluctuation data to predict the grounding fault and short-circuit risk area.

[0192] The fault source positioning module 305 is used to establish a space-time correlation matrix for the fault features of the three scales in the grounding fault and short-circuit risk area and dynamically allocate attention weights to locate the fault source.

[0193] The fault point distance measurement output module 306 is used to model the fault source using a topology neural network, and output the fault point distance measurement value through state prediction and strategy deduction.

[0194] Reference Figure 3The power distribution network fault location system for distributed power supply access belongs to the technical field of intelligent power distribution system, facility and power distribution switch control equipment manufacturing.

[0195] It should be noted that, in the present application, the embodiments implemented on the side of the power distribution network fault location system for distributed power supply access can be mutually referenced with the embodiments implemented on the side of the power distribution network fault location method for distributed power supply access, and the present application will not be described one by one.

[0196] The above is only the preferred embodiment of the present application, and does not limit the patent scope of the present application, and any equivalent structure or equivalent flow transformation made by using the content of the specification and drawings of the present application, or direct or indirect application in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. A distribution network fault location method for distributed power access, characterized in that: include: The electrical parameters and operating status of distributed power generation access nodes in the distribution network are collected in real time, a dynamic manifold model is constructed based on the ecological niche theory, and the manifold learning neighborhood parameters are adaptively adjusted according to the power fluctuation data included in the electrical parameters and the node ecological niche is updated to reconstruct the dynamic manifold model; Based on the reconstructed dynamic manifold model, fault features are extracted from three scales: current harmonic components, inter-feeder harmonic propagation paths, and overall network voltage impact. Three-dimensional feature vectors are generated through fusion using a graph association algorithm. Using the three-dimensional feature vector to simulate distributed power supply fault scenarios and dynamically expand the fault sample library; Based on the expanded fault sample library and the power fluctuation data, a fault transmission relationship model is constructed to predict ground fault and short circuit risk areas; Establishing a spatiotemporal correlation matrix for the fault characteristics at three scales within the ground fault and short circuit risk area and dynamically assigning attention weights to locate the fault source; A topological neural network is used to model the fault source, and a distance measurement value of the fault point is output through state prediction and strategy deduction.

2. The method for fault location in a distribution network for distributed power access according to claim 1, characterized in that: The method of constructing a dynamic manifold model based on the ecological niche theory, adaptively adjusting manifold learning neighborhood parameters according to power fluctuation data included in electrical parameters, and updating node ecological niches to reconstruct the dynamic manifold model includes: Constructing a dynamic manifold model based on the niche overlap algorithm; Adjusting a manifold learning neighborhood parameter according to a wind farm output power change rate in the power fluctuation data, specifically increasing a neighborhood range of a local linear embedding algorithm, wherein the neighborhood range is represented by the number of neighborhood points; updating the node niche according to the charge and discharge status of the energy storage device in the operating state, wherein the charging state node is defined as the voltage support niche and the discharging state node is defined as the power regulation niche; The dynamic manifold model is reconstructed based on the adjusted manifold learning neighborhood parameters and the updated node niches.

3. The method for fault location in a distribution network for distributed power access according to claim 2, characterized in that: The adjusting of the manifold learning neighborhood parameters according to the wind farm output power change rate in the power fluctuation data includes: when the wind farm output power change rate exceeds a power change rate threshold, increasing the number of neighborhood points of the local linear embedding algorithm, wherein the increase in neighborhood points is positively correlated with the wind farm output power change rate; The updating of the node niche includes: allocating a voltage stability optimization coefficient to the voltage support niche; allocating a power fluctuation suppression coefficient to the power regulation niche; The reconstructing of the dynamic manifold model is specifically as follows: taking the increased number of neighborhood points as a spatial constraint condition, and taking the voltage stability optimization coefficient and the power fluctuation suppression coefficient as node attribute weights; executing a local linear embedding algorithm to generate a topological manifold structure adapted to the access of distributed power sources, and the topological manifold structure is the reconstructed dynamic manifold model.

4. The method for fault location in a distribution network for distributed power access according to claim 1, characterized in that: The fault characteristics include distributed power supply harmonic anomaly and short circuit fault characteristics; Based on the reconstructed dynamic manifold model, fault features are extracted from three scales: current harmonic components, inter-feeder harmonic propagation paths, and overall grid voltage impact. A three-dimensional feature vector is generated through fusion using a graph association algorithm, including: Extracting current spectrum features from the current harmonic component scale based on the reconstructed dynamic manifold model. The current spectrum features include the amplitude of a specific frequency band corresponding to the harmonic anomaly and the fundamental wave distortion rate corresponding to the short-circuit fault. From the inter-feeder harmonic propagation path scale, the path correlation characteristics are calculated and extracted based on the reconstructed dynamic manifold model. The path correlation characteristics include the path attenuation coefficient of harmonic propagation and the diffusion radius of the short-circuit fault current. Extract voltage deviation rate features from the grid-wide voltage impact scale based on the reconstructed dynamic manifold model. The voltage deviation rate features include the steady-state voltage offset caused by harmonics and the transient voltage sag depth caused by short-circuit faults. Constructing a distribution network topology map based on the path correlation characteristics, wherein the nodes of the distribution network topology map correspond to the distributed power access nodes, and the edges correspond to the feeder connection relationships; Mapping the current spectrum characteristics into node electrical properties; Mapping the voltage deviation rate feature into an edge influence weight; Calculate the shortest propagation path weight between nodes through Dijkstra algorithm; Evaluate the node fault influence by combining the node electrical properties and the edge influence weight using the PageRank algorithm; The shortest propagation path weight and node failure influence are integrated to generate a graph correlation feature matrix. The current spectrum feature, the voltage deviation rate feature and the graph correlation feature matrix are input into a fully connected neural network to generate a three-dimensional feature vector.

5. The method for fault location in a distribution network for distributed power access according to claim 4, characterized in that: The method of constructing a fault transmission relationship model based on the expanded fault sample library and the power fluctuation data to predict ground fault and short circuit risk areas includes: Inputting the three-dimensional feature vector and the power fluctuation data into a dynamic Bayesian network to construct a fault transmission relationship model between meteorological parameters, equipment aging indicators, and electrical quantities, wherein the meteorological parameters include wind speed mutation gradient and light intensity attenuation rate, and the equipment aging indicators are characterized by impedance temperature rise coefficient and insulation dielectric loss; The vine copula function is used to analyze the nonlinear dependence structure of the wind speed sudden change gradient and the wind farm output power change rate. The fault transmission probability weight is calculated by combining the time-varying effect of equipment aging indicators on feeder impedance. generating a set of association rules between equipment aging indicators and harmonic abnormality characteristics based on the fault transfer probability weight; The expanded fault sample library is scanned through a sliding time window to extract the mapping relationship between the short-circuit current probability density distribution and the transient characteristics of the ground fault. The association rule set and the mapping relationship are integrated and a spatiotemporal Kriging interpolation algorithm is applied to generate a ground fault probability cloud map and a short circuit risk heat map, and a spatial coordinate set of the ground fault and short circuit risk areas is output.

6. The method for fault location in a distribution network for distributed power access according to claim 5, characterized in that: The method of establishing a spatiotemporal correlation matrix for the fault characteristics at three scales within the ground fault and short circuit risk area and dynamically allocating attention weights to locate the fault source includes: Obtaining fault feature time series fragments of the current harmonic component scale, inter-feeder harmonic propagation path scale, and overall network voltage impact scale in the ground fault and short circuit risk area to construct a three-dimensional space-time tensor; Performing a high-order Tucker decomposition on the three-dimensional space-time tensor to separate a spatial pattern matrix, a temporal pattern matrix, and a characteristic dimension matrix; Based on the dynamic convolution calculation results of the feeder wave impedance and the equivalent harmonic impedance of the distributed generation, the fault propagation speed correction factor is generated and embedded into the time mode matrix; Recombining the spatial pattern matrix and the time pattern matrix embedded with the fault propagation speed correction factor to form a time-space correlation matrix; Decoding the spatiotemporal correlation matrix through a bidirectional gated recurrent unit to predict the attenuation trajectory of the fault signature in the topological manifold structure; Generate a spatial attention weight vector and a temporal attention weight vector according to the curvature change rate of the attenuation trajectory and the voltage deviation rate mutation point; The weighted Mahalanobis distance algorithm is used to fuse the spatial attention weight vector and the temporal attention weight vector to locate the coordinates of the fault source. The short-circuit fault source location prioritizes the activation of the spatial weight of the voltage sag depth mutation node, and the harmonic anomaly fault source location prioritizes the activation of the temporal weight of the path attenuation anomaly moment.

7. The method for fault location in a distribution network for distributed power access according to claim 6, characterized in that: The generation of the fault propagation speed correction factor includes: Capture the arrival time difference sequence of the traveling wave fronts of each feeder within 1.5 milliseconds before and after the fault occurs; Based on the topological manifold structure of the dynamic manifold model, the equivalent length matrix of the traveling wave path is calculated; Construct a robust estimation model for time difference and length mapping, and eliminate abnormal time difference points caused by switching noise of grid-connected inverters; Extract the frequency characteristic curve of the feeder wave impedance; fit the nonlinear equation of the fault propagation speed varying with the harmonic frequency; The nonlinear equation is converted into a time dimension scaling coefficient and embedded into the space-time correlation matrix generation process.

8. The method for fault location in a distribution network for distributed power access according to claim 2, characterized in that: The method of using a topological neural network to model the fault source and outputting a distance measurement value of the fault point through state prediction and strategy deduction includes: Construct a k-order topological neighborhood subgraph with the located fault source coordinates as the center; The topological neural network is used to model the fault evolution process. The topological neural network is composed of a cascade of graph convolutional layers and gated recurrent units. A dual-path strategy deduction is performed using a rolling horizon optimization algorithm. The first path uses the charge and discharge state switching of the energy storage device as the decision variable to deduce the power control strategy, while the second path uses the feeder section switch combination sequence as the decision variable to deduce the fault isolation strategy. In each optimization time domain, the weight distribution of the voltage support niche and the power regulation niche is updated according to the power control strategy; Dynamically prune the edge connections of the topological neighborhood subgraph according to the fault isolation strategy; The updated niche weights and subgraph structures are fed into the temporal graph attention network to predict the multi-step fault current diffusion trajectory and the arrival timing of the traveling wavefront. Establish the differential mapping relationship between the fault current diffusion trajectory and the arrival time difference of the traveling wave front; The minimum energy solution of the differential mapping relationship is determined to output the fault point distance measurement value.

9. The method for fault location in a distribution network for distributed power access according to claim 8, characterized in that: The dual-path strategy deduction includes: Establishing a conflict detection mechanism between the power regulation strategy and the fault isolation strategy, triggering a conflict flag when a node in a discharge state is required to provide voltage support and the feeder segment where the node is located is marked as an isolation segment; Initiate multi-objective optimization for the topology subgraph area covered by the conflict marker, where the first objective function minimizes the voltage deviation rate and the second objective function maximizes the fault isolation success rate; Screen the optimal strategy sequence through the Pareto frontier solution set; Update the node attributes and edge connection weights of the topological neighborhood subgraph according to the optimal strategy sequence; The updated subgraph structure is fed back to the fault current diffusion trajectory prediction module for iterative optimization.

10. A distribution network fault location system for distributed power access, characterized in that: include: A dynamic manifold model construction and reconstruction module is used to collect the electrical parameters and operating status of distributed power access nodes in the distribution network in real time, construct a dynamic manifold model based on the ecological niche theory, adaptively adjust the manifold learning neighborhood parameters according to the power fluctuation data included in the electrical parameters, and update the node ecological niche to reconstruct the dynamic manifold model; The 3D feature vector generation module is used to extract fault features from three scales: current harmonic components, inter-feeder harmonic propagation paths, and overall network voltage impact, based on the reconstructed dynamic manifold model. The 3D feature vector is then generated through fusion using a graph association algorithm. A fault sample library expansion module, configured to simulate a distributed power supply fault scenario using the three-dimensional feature vector and dynamically expand the fault sample library; A risk area prediction module is used to construct a fault transmission relationship model based on the expanded fault sample library and the power fluctuation data to predict ground fault and short circuit risk areas; A fault source location module is used to establish a spatiotemporal correlation matrix for the fault characteristics of three scales in the ground fault and short circuit risk area and dynamically allocate attention weights to locate the fault source; The fault point distance measurement value output module is used to model the fault source using a topological neural network and output the fault point distance measurement value through state prediction and strategy deduction.

Images