A method for identifying a power collection line fault section based on a physical information graph neural network and dynamic causal reasoning

Abstract

The application discloses a kind of based on physical information graph neural network and dynamic causal inference's current-carrying wire fault section identification method, it is related to current-carrying wire fault section identification technical field, by collecting multi-node millisecond level voltage transient data, and topological modeling is electrical diagram: node is monitoring point, edge contains the priori physical parameters of impedance, length, simultaneously constructs physical information graph neural network, in graph neural network training process embedding Kirchhoff's law and the differential equation constraint of fault transient propagation, make network model not only learn from data, more follow power system physical law;Real-time disturbance direction and source are distinguished in combination with dynamic causal inference;Finally fusion, realize intelligent distinction of fault and normal fluctuation, accurately determine that fault is in the specific section between adjacent monitoring nodes, output position, type and confidence for alarm linkage, the present method is strong, generalization is good, supports online fast positioning, reduces misjudgment.

Description

Technical Field

[0001] This invention relates to the field of collector wire fault section identification technology, and in particular to a collector wire fault section identification method based on physical information graph neural network and dynamic causal reasoning. Background Technology

[0002] With the high proportion of new energy sources such as wind power and photovoltaics being connected to the grid, multiple turbines and box-type transformers converge and connect to the grid via collection lines at the power station side, forming a radial-branch topology. This results in long line lengths, high node density, and a mix of cable and overhead installations. Grid connection switching, control strategy adjustments, switch operations, and initial fault states can all trigger cross-node transients within milliseconds and propagate along the lines. Although high-speed synchronous measurements have been deployed at the turbine outlet, the low-voltage side of the box-type transformer, and the busbar, slight asynchrony in field communication, clock drift, sampling jitter, multi-source harmonics, and background noise make it difficult to achieve high-precision alignment, stable triggering, and rapid utilization of multi-node data. Furthermore, the propagation phase velocity and attenuation of the mixed lines vary significantly due to the laying environment and operating mode, making it difficult to accurately characterize fixed parameters. This further restricts the ability to uniformly model, rapidly identify, and precisely locate disturbances.

[0003] Existing positioning technologies mainly include threshold or feature triggering, impedance or traveling wave ranging, synchronous phasor and state estimation, and pure data-driven learning. Threshold and feature methods are simple to implement, but sensitive to threshold settings and noise; impedance methods rely on equivalent parameters, grounding methods, and fault models, and errors increase in scenarios with uncertain parameters, mixed cables, and high resistance-to-reactance ratios; traveling wave two-end ranging requires high-bandwidth measurements and strict timing, is susceptible to reflection, bifurcation, and attenuation, and has high engineering costs. Synchronous phasor and state estimation is limited by measurement coverage, topology aging, and sampling rate, making it difficult to analyze millisecond-level transients. Data-driven methods are highly adaptive, but ignore Kirchhoff's current law and voltage law of the power grid, as well as electromagnetic propagation mechanisms, resulting in scarce training samples, insufficient cross-site migration capability, and weak interpretability, making it difficult to stably support site-level online applications; the consistency assessment of multi-source data also lacks a unified framework, affecting the mutual verification of conclusions and the credibility of commissioning.

[0004] In engineering sites, the voltage amplitude and phase of adjacent nodes under the same bus or feeder conditions are often highly similar, making it easy to cause ambiguity in source and direction determination. Narrowband harmonics introduced by inverter interfaces, control coupling, and power limiting strategies may mask or distort fault characteristics. Communication links commonly experience asynchrony, packet loss, and short-term disconnection on the order of microseconds to milliseconds, which can easily lead to misjudgment of arrival timing and "pseudo-causality." Sensor drift, anomalies, and missing values ​​further weaken the stability of segment-level criteria. Existing solutions are insufficiently selective in distinguishing between fault disturbances and normal fluctuations (such as grid connection / disconnection, reactive power switching, etc.), and also lack in rapid sub-segment-level location and low-overhead online operation. Therefore, it is necessary to integrate topological information and electromagnetic physics laws under voltage measurement conditions to form a technical path with lightweight dynamic direction determination, robustness to minor asynchrony and harmonics, and the ability to output confident segment results, in order to support large-scale deployment and improve location speed, accuracy, and operational efficiency. Summary of the Invention

[0005] In view of the above-mentioned prior art, the present invention provides a method for identifying fault zones in collector wires based on physical information graph neural networks and dynamic causal reasoning, which mainly solves the technical problems existing in the above-mentioned background art.

[0006] To achieve the above objectives, the technical solution of this invention is implemented as follows: In a first aspect, the present invention provides a method for identifying fault zones in collector wires based on physical information graph neural networks and dynamic causal reasoning, the method comprising the following steps: S1. Real-time acquisition of multi-node voltage transient data of the collector line, and preprocessing. The preprocessed data is used to form multi-node voltage transient samples based on threshold or sudden change detection trigger event window. S2. Based on the principle of primary wiring, the monitoring points are modeled as graph nodes and the line branches are graph edges. The physical and geometric prior parameters are configured for the graph edges to obtain an electrical diagram with physical priors. S3. Construct a physical information graph neural network, taking multi-node voltage transient samples and electrical graph as input, and outputting the physical consistency score between the graph nodes and graph edges in the electrical graph, the candidate propagation path and the confidence of the source node; S4. Online calculation of dynamic causal reasoning for multi-node transient sequences based on sliding time windows, obtaining time-varying directional influence relationships and perturbation propagation directions, and providing candidate source nodes and arrival time sequences; S5. Integrate the scoring results of physical consistency with the results of dynamic causal inference. When the two are consistent, locate directly; when the two are inconsistent, first distinguish between fault and normal fluctuation, and make segment judgment based on the confidence level and the weight of the physical residual.

[0007] As a preferred embodiment of the present invention, step S1 specifically includes: High-speed synchronous voltage sensors are deployed at key nodes such as the wind turbine outlet, the low-voltage side of the box-type transformer, and the busbar junction of the collection line. Multi-node voltage transient sequences are obtained by sampling at the millisecond level, and data preprocessing is performed. The voltage change rate is calculated based on the obtained multi-node voltage transient sequence, and the maximum value is taken within the node set to characterize the voltage change intensity. An event is triggered when the maximum rate of change is not less than the rate of change threshold. At the same time, the square of the preprocessed voltage transient sequence of each node is integrated within the energy statistics window and summed over all nodes to characterize the cumulative energy of the transient intensity. An event is also triggered when the sum of this cumulative energy is not less than the energy threshold. When any triggering condition is met, an event time window is constructed by extending the event window forward and backward with the triggering time as the center. Within the event window, the preprocessed voltage transient sequence of each node is extracted to form a multi-node event dataset and output.

[0008] As a preferred embodiment of the present invention, the step S2 of configuring the physical and geometric prior parameters for the graph edges specifically includes: Each graph edge is configured with prior physical parameters and geometric information such as impedance, inductance, capacitance, conductance, and geometric length, and an edge feature vector is formed to construct multi-scale temporal features for graph nodes.

[0009] As a preferred embodiment of the present invention, obtaining the electrical diagram with physical prior knowledge in step S2 specifically includes: Based on the preprocessed voltage transient sequence of nodes within the event window, the feature components of five types of prior physical parameters are calculated on the same event window and then concatenated to form a node feature vector. The telegraph equations with distributed parameters are used as physical consistency constraints for the prior physical model of electromagnetic propagation. The propagation constant, phase velocity, and attenuation coefficient of the prior physical parameters in the frequency domain are calculated based on the electromagnetic propagation physical prior model. The propagation constant, phase velocity, and attenuation coefficient are then analytically mapped to the physical propagation index. This mapping result is then injected as a boundary-level physical prior into the graph data construction and model constraints to obtain an electrical graph with physical priors.

[0010] As a preferred embodiment of the present invention, step S3 specifically includes: Using node feature vectors and edge feature vectors as initial embeddings, an initial embedding vector is assigned to each graph node based on the multi-node voltage transient samples and the electrical graph. The hidden representations of graph nodes and graph edges are obtained through graph message passing and attention updates. Each layer update includes: generating a message from the source node to the target node based on the physical prior parameters of the source node embedding, target node embedding, and edge configuration; aggregating all neighbor messages received by the target node, and inputting the aggregation result and the target node's current layer embedding into the update function to obtain the target node's next layer embedding. Based on the final node embedding obtained after multiple iterations, the confidence of the fault source node, the node physical consistency score, the edge fault score, and the confidence of the candidate propagation path are generated through multiple parallel output heads.

[0011] As a preferred embodiment of the present invention, step S3 further includes constructing a joint loss function to optimize the physical information graph neural network during the model training stage. The joint loss function includes a data fitting loss term and residual constraint terms constructed based on Kirchhoff's current law, Kirchhoff's voltage law and the fault transient propagation partial differential equation, respectively.

[0012] As a preferred embodiment of the present invention, step S4 specifically includes: A sliding sub-window is set within the event window. The time delay cross-correlation function is calculated based on the voltage transient time series data of each pair of monitoring nodes within the sub-window. The lag time between nodes is obtained by finding the cross-correlation peak. After standardization, a directed time delay estimation matrix is ​​formed for each sub-window, which serves as a preliminary criterion for arrival time series. On the preset target frequency band, the coherence function and phase spectrum are calculated based on the time series data of each pair of monitoring nodes, and then the phase slope index is calculated as a directional index. The net propagation direction between nodes is determined according to the positive and negative signs and significance of the directional index, and the directional index matrix of each sub-window is obtained. Based on the directed time delay estimation matrix and the directional index matrix, weights are constructed by combining the signal-to-noise ratio and coherence of each node pair. The global time offset vector is estimated by weighted optimization and the original time series data is compensated to obtain the corrected time delay matrix. Based on the corrected time delay matrix and the directional index matrix, the directed influence matrix of each sub-window is constructed by fusing them together, and the directed influence matrix on the sliding sub-window sequence is subjected to exponential moving average to obtain the smoothed time-varying directed influence matrix. Based on the smoothed time-varying directed influence matrix, the net outflow intensity of each node is calculated. The net outflow intensity is the sum of the influence intensity from the current node to all other nodes minus the sum of the influence intensity from all other nodes to the current node. The candidate source node set is sorted and selected according to the magnitude of the net outflow intensity, and the evolution trajectory of the time-varying directed influence matrix, the candidate source node set, and the arrival time sequence of each candidate source are output.

[0013] As a preferred embodiment of the present invention, step S5 specifically includes: Based on the topological connection relationship of the electrical diagram, the line branches between adjacent monitoring nodes are constructed as a set of candidate segments, and the physical consistency score of each graph node and the edge fault score of each graph edge are obtained, as well as the set of candidate source nodes and the arrival time sequence of each node are obtained. For each candidate segment, adaptive fusion weights are calculated based on the physical consistency scores of the nodes at both ends of the segment and the physical residuals that characterize the deviation between the observation time difference and the physical model. Based on the edge fault score, the adaptive fusion weight, the candidate source node indication, and the observation arrival time difference of the nodes at both ends of the segment, the basic fusion score of each candidate segment is calculated. Based on the configured line physical prior parameters, the theoretical time difference of arrival for the fault disturbance to propagate along the candidate segment is calculated. The theoretical time difference is compared with the observed time difference of arrival to obtain the time difference error. A reward decay function is constructed based on the time difference error, and the basic fusion score is corrected using the reward decay function to obtain the final fusion score for each candidate segment. When determining the faulty segment based on the final fusion score, a hysteresis comparison mechanism is introduced. When the segment corresponding to the maximum final fusion score at the current moment is different from the faulty segment maintained at the previous moment, and the maximum final fusion score exceeds the sum of the final fusion score of the faulty segment maintained at the previous moment and the preset hysteresis boundary value, the faulty segment is switched to the segment corresponding to the current maximum score; otherwise, the faulty segment at the previous moment is maintained.

[0014] In a second aspect, the present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements any of the above-described methods for identifying fault zones in a collector wire based on a physical information graph neural network and dynamic causal reasoning.

[0015] Thirdly, the present invention also provides an electronic device, including a processor and a memory, the memory storing a plurality of instructions, the processor loading instructions from the memory to execute the steps in any of the above-described methods for identifying fault zones in a collector wire based on a physical information graph neural network and dynamic causal reasoning.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention explicitly embeds power grid topology and electromagnetic physics laws into a graph neural network, combining propagation priors and causal direction evidence to clearly distinguish between normal fluctuations and fault disturbances; it accurately locates fault sections between adjacent monitoring nodes and outputs location, type, and confidence level for linked alarms. It achieves low-latency and highly selective online positioning without requiring full and accurate parameter identification, and supports parallel operation of multiple lines and incremental updates, reducing false positives and false negatives, while meeting the needs of large-scale operation and maintenance and rapid response of new energy power plants. Attached Figure Description

[0017] Figure 1 This is a schematic diagram illustrating the steps of a method for identifying fault zones in a collector wire based on a physical information graph neural network and dynamic causal reasoning. Detailed Implementation

[0018] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. In the following description, the expression "some embodiments" refers to a subset of all possible embodiments; however, it should be understood that "some embodiments" can be the same subset or different subsets of all possible embodiments and can be combined with each other without conflict.

[0019] In the following description, numerous specific details are set forth in order to provide a more thorough understanding of the invention. However, it will be apparent to those skilled in the art that the invention can be practiced without one or more of these details. In other instances, certain technical features well-known in the art have not been described in order to avoid obscuring the invention.

[0020] It should be understood that the present invention can be embodied in various forms and should not be construed as being limited to the embodiments set forth herein. Rather, providing these embodiments will make the disclosure thorough and complete, and will fully convey the scope of the invention to those skilled in the art. Furthermore, the terminology used herein is intended only to describe particular embodiments and is not intended to limit the invention. When used herein, the singular forms “a,” “an,” and “the” are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the terms “compose” and / or “comprising,” when used in this specification, identify the presence of the stated features, integers, steps, operations, elements, and / or components, but do not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups. When used herein, the term “and / or” includes any and all combinations of the associated listed items.

[0021] It should also be noted that when an element is referred to as being "fixed to" another element, it can be directly attached to the other element or there may be an intervening element. When an element is referred to as being "connected to" another element, it can be directly connected to the other element or there may be an intervening element. The terms "vertical," "horizontal," "inner," "outer," "left," "right," and similar expressions used herein are for illustrative purposes only and do not represent the only possible implementation.

[0022] To fully understand this invention, a detailed structure will be presented in the following description to illustrate the technical solution proposed by this invention. Optional embodiments of the invention are described in detail below; however, in addition to these detailed descriptions, the invention may have other embodiments.

[0023] Firstly, this invention provides a method for identifying fault zones in collector wires based on physical information graph neural networks and dynamic causal reasoning. Please refer to the appendix for details. Figure 1 The method includes the following steps: S1. Real-time acquisition of multi-node voltage transient data of the collector wire, and preprocessing. The preprocessed data is used to form multi-node voltage transient samples based on threshold or sudden change detection trigger event windows.

[0024] As a preferred embodiment of the present invention, step S1 specifically includes: High-speed synchronous voltage sensors are deployed at key nodes such as the wind turbine outlet, the low-voltage side of the box-type transformer, and the busbar junction of the collection line. Multi-node voltage transient sequences are obtained by sampling at the millisecond level, and data preprocessing is performed. In this embodiment, voltage sensors with nanosecond-level time alignment capabilities are deployed at key nodes such as the wind turbine outlet, the low-voltage side of the box-type transformer, and the busbar to ensure high-fidelity acquisition of the three-phase voltage channels. Each monitoring node acquires at least three-phase voltage, and if necessary, simultaneously acquires the neutral point voltage. The sampling period is preferably 0.5–2 ms. The analog front-end is configured with anti-aliasing filtering, and the range and accuracy meet the requirements of not overloading transient amplitudes and not distorting details. Simultaneously, the acquired raw data is packaged by node. The preprocessed sequence includes node numbers, line / branch identifiers, and equipment health status such as smudging rate, saturation rate, signal-to-noise ratio, and harmonic distortion. It undergoes basic cleaning processes such as DC and trend term removal, bandpass or wavelet threshold noise suppression, outlier identification, and necessary short-missing imputation to obtain the final preprocessed sequence. To avoid redundant calculations over long periods, a joint criterion of derivative threshold and energy threshold is used to trigger event times on multiple nodes. When any node satisfies the condition that the peak value of the derivative exceeds the threshold Or the energy exceeds the threshold for a short period of time This means that a disturbance has occurred.

[0025] The voltage change rate is calculated based on the obtained multi-node voltage transient sequence, and the maximum value is taken within the node set to characterize the voltage change intensity. An event is triggered when the maximum rate of change is not less than a rate of change threshold. Simultaneously, the squares of the preprocessed voltage transient sequences at each node are integrated within the energy statistics window and summed over all nodes to represent the cumulative energy characterizing the transient intensity. When the accumulated energy The same event is triggered when the energy threshold is not less than the specified value. When any triggering condition is met, an event time window is constructed by extending the event window forward and backward with the triggering time as the center. Within the event window, the preprocessed voltage transient sequence of each node is extracted to form a multi-node event dataset and output.

[0026] In this embodiment, to avoid redundant calculations over long periods, a joint criterion of derivative threshold and energy threshold is used to trigger event times on multiple nodes. When any node satisfies the condition that the peak value of the derivative exceeds the threshold or accumulated energy Exceeding the threshold This means that a disturbance has occurred.

[0027] Specifically, accumulated energy can be defined as: ; The mathematical expression for its triggering condition is as follows:

[0028]

[0029] in, , It can be adaptively set based on the historical 95% to 99% percentile, and the window length L is set to tens to hundreds of sampling points depending on the sampling period.

[0030] Sure The event window constructed afterward is: ; in, To extend the duration in front of the event window forward, Extend the duration of the event window backward.

[0031] Specifically, only the multi-node transient samples within the event window, aligned to a unified time axis, and their quality control indicators are used as input for subsequent topology modeling and identification. Event samples that do not meet the standards are removed or downweighted.

[0032] S2. Based on the principle of primary wiring, the monitoring points are modeled as graph nodes and the line branches are graph edges. The physical and geometric prior parameters are configured for the graph edges to obtain an electrical diagram with physical priors.

[0033] As a preferred embodiment of the present invention, the step S2 of configuring the physical and geometric prior parameters for the graph edges specifically includes: Each graph edge is configured with prior physical parameters and geometric information such as impedance, inductance, capacitance, conductance, and geometric length, and an edge feature vector is formed to construct multi-scale temporal features for graph nodes.

[0034] As a preferred embodiment of the present invention, obtaining the electrical diagram with physical prior knowledge in step S2 specifically includes: Based on the preprocessed voltage transient sequence of nodes within the event window, the feature components of five types of prior physical parameters are calculated on the same event window and then concatenated to form a node feature vector. The telegraph equations with distributed parameters are used as physical consistency constraints for the prior physical model of electromagnetic propagation. The propagation constant, phase velocity, and attenuation coefficient of the prior physical parameters in the frequency domain are calculated based on the electromagnetic propagation physical prior model. The propagation constant, phase velocity, and attenuation coefficient are then analytically mapped to the physical propagation index. This mapping result is then injected as a boundary-level physical prior into the graph data construction and model constraints to obtain an electrical graph with physical priors.

[0035] In this embodiment, an electrical diagram is established based on electrical wiring and electrical connectivity. Monitoring points are mapped as graph nodes. , Given the total number of nodes, the branch lines are mapped to graph edges. , This represents the total number of edges. Physical and geometric prior parameters are configured for each graph edge. Including resistance per unit length ,inductance ,capacitance Electrical conductivity Branch length These prior parameters can be represented by deterministic values ​​or by intervals or statistics to characterize uncertainty. Multi-scale features are extracted for each node. These include instantaneous values, derivatives, short-time energy, spectral entropy, kurtosis, and bandpass energy ratio. The telegraph equations for distributed parameter lines are used as the propagation prior model to calculate the propagation prior parameters. The specific calculation formulas are as follows: Propagation constant: ; Phase velocity: ; Attenuation coefficient: ; in, This indicates taking the imaginary part of the propagation constant. This indicates taking the real part of the propagation constant. It represents angular frequency.

[0036] In representative frequency bands Based on this, the theoretical time difference of arrival and amplitude attenuation between adjacent monitoring nodes are given, which serve as the physical benchmark for subsequent verification and fusion. The specific formula is as follows:

[0037] in, Indicates the theoretical time difference of arrival. Indicates the theoretical amplitude. This represents the initial amplitude.

[0038] Simultaneously, multi-scale temporal features of graph nodes are extracted from the event window. Parameters such as instantaneous values, derivatives, short-time energy, spectral entropy, and bandpass energy ratio, together with electrical diagrams and propagation prior parameters, form the input to the subsequent neural network.

[0039] S3. Construct a physical information graph neural network, taking multi-node voltage transient samples and electrical graph as input, and outputting the physical consistency score of the nodes and edges in the electrical graph, the confidence of the candidate propagation path and the source node.

[0040] As a preferred embodiment of the present invention, step S3 specifically includes: Using node feature vectors and edge feature vectors as initial embeddings, an initial embedding vector is assigned to each graph node based on the multi-node voltage transient samples and the electrical graph. The hidden representations of graph nodes and graph edges are obtained through graph message passing and attention updates. Each layer update includes: generating a message from the source node to the target node based on the physical prior parameters of the source node embedding, target node embedding, and edge configuration; aggregating all neighbor messages received by the target node, and inputting the aggregation result and the target node's current layer embedding into the update function to obtain the target node's next layer embedding. Based on the final node embedding obtained after multiple iterations, the confidence of the fault source node, the node physical consistency score, the edge fault score, and the confidence of the candidate propagation path are generated through multiple parallel output heads.

[0041] In this embodiment, firstly, based on the obtained multi-node voltage transient samples and the constructed electrical graph, an initial embedding vector is assigned to each graph node. The initial node embedding is processed by a learnable linear transformation layer, which maps the multi-scale voltage transient features of the corresponding node (such as instantaneous value, derivative, short-time energy, spectral entropy, kurtosis, bandpass energy ratio, etc.) to a high-dimensional hidden space to obtain the hidden representation of the graph node and graph edges. Simultaneously, an initial embedding vector is constructed for each graph edge. This initial embedding vector is obtained through a linear transformation based on the configured physical and geometric prior parameters, thereby encoding the inherent physical characteristics of the line into edge features.

[0042] Subsequently, multi-layer iterative message passing and node embedding updates are performed in the physical information graph neural network. Each layer's update process incorporates an attention mechanism to adaptively learn the importance weights of neighboring nodes. Specifically, for the current layer, the attention coefficients between the center node and each of its neighboring nodes are first calculated: the current-layer embeddings of the source node, the target node, and the edges connecting them are concatenated and input into a shared attention function (implemented by a single-layer feedforward neural network). The raw attention scores are then obtained through an activation function. Subsequently, the raw scores of all neighboring nodes are normalized using the Softmax function to obtain normalized attention weights. These attention weights reflect the importance of the information provided by each neighboring node when updating the center node representation.

[0043] Building upon this foundation, for each edge, a message vector from the source node to the target node is generated using a learnable message function based on the source node embedding, target node embedding, and edge embedding. This message function can be a multilayer perceptron, whose input is the concatenation of the three embeddings, and whose output is the message vector. Next, all neighbor messages received by the target node are weighted and aggregated, with the weights being the aforementioned normalized attention coefficients, thus obtaining the aggregated neighborhood information for that node. Finally, the aggregated neighborhood information is fused with the target node's current layer embedding, and an update function (such as a gated recurrent unit or a multilayer perceptron) is input to generate the target node's embedding in the next layer. This process of message generation, attention-weighted aggregation, and embedding update completes one layer of iteration. This process can be repeated multiple times, allowing information to propagate multi-hop along the electrical graph topology. Each layer further integrates the features of a wider range of neighbor nodes and the physical attributes of the path, ultimately obtaining the final embedding representation of each node. This representation includes the node's transient characteristics, local neighborhood topology, and physical constraints along the propagation path.

[0044] As a preferred embodiment of the present invention, step S3 further includes constructing a joint loss function to optimize the physical information graph neural network during the model training stage. The joint loss function includes a data fitting loss term and residual constraint terms constructed based on Kirchhoff's current law, Kirchhoff's voltage law and the fault transient propagation partial differential equation, respectively.

[0045] Specifically, both training and online inference incorporate soft constraints from the power grid's physical constraints, and a joint loss is used. Defined as:

[0046] in, For data fitting loss term, These are residual constraint terms of Kirchhoff's current law. This is a residual constraint term for Kirchhoff's voltage law. For the residual constraint terms of the partial differential equation, These are the weighting coefficients for the physical constraint terms.

[0047] Specifically, the Kirchhoff current law residual constraint term adopts a discrete approximation of the nodal injection current, assuming the node... The adjacency set is With equivalent admittance Admittance of parallel branch Estimate the current, and denote the node voltage sample as... Then node At any moment Current residual for:

[0048]

[0049] Specifically, the Kirchhoff voltage law residual constraint term is defined as follows in each basic loop: Upper computational loop At any moment voltage residual The specific calculation formula is as follows:

[0050]

[0051] in, , Representing circuits respectively On the connection node, Indicates the connection node , At any moment voltage drop, The total number of sampling points within the time window.

[0052] Specifically, the residual constraint terms of the telegraph equations are discretized using finite difference discretization of propagated prior parameters, with the edge... The length is The sampling interval is Approximating the time derivative using backward difference, we obtain:

[0053] in, use approximate, Approximated by second-order difference An endpoint differential approximation is constructed using the voltages at both ends. Weights Optimization was performed using a validation set, with initial values ​​ranging from 0.1 to 1. To improve robustness in the field, noise enhancement, parameter perturbation, and missing value masks were added during the training phase, while temperature scaling and quantile calibration were applied to the output during the online phase. The graph inference latency for a single event window was optimally controlled between 50 and 100 ms, increasing linearly with node size. The final result is:

[0054] in, Output edge-level fault scores for the network to characterize the segment. The possibility of failure; It is the Sigmoid activation function. The linear mapping weights for the edge-level scores; This is the hidden representation of the edge.

[0055]

[0056] in, To output the node-level source probability, representing the node The possibility of a disturbance source, For learnable mappings, For the first Layer node embedding.

[0057]

[0058] in, For the edge Physical consistency score, The corresponding edge can be learned mapping.

[0059]

[0060] in, node Physical consistency score, The corresponding node can be learned mapping.

[0061] S4. Based on the sliding time window, perform online calculation of dynamic causal reasoning on multi-node transient sequences, obtain time-varying directional influence relationships and perturbation propagation directions, and provide candidate source nodes and arrival times.

[0062] As a preferred embodiment of the present invention, step S4 specifically includes: A sliding sub-window is set within the event window. The time delay cross-correlation function is calculated based on the voltage transient time series data of each pair of monitoring nodes within the sub-window. The lag time between nodes is obtained by finding the cross-correlation peak. After standardization, a directed time delay estimation matrix is ​​formed for each sub-window, which serves as a preliminary criterion for arrival time series. On the preset target frequency band, the coherence function and phase spectrum are calculated based on the time series data of each pair of monitoring nodes, and then the phase slope index is calculated as a directional index. The net propagation direction between nodes is determined according to the positive and negative signs and significance of the directional index, and the directional index matrix of each sub-window is obtained. Based on the directed time delay estimation matrix and the directional index matrix, weights are constructed by combining the signal-to-noise ratio and coherence of each node pair. The global time offset vector is estimated by weighted optimization and the original time series data is compensated to obtain the corrected time delay matrix. Based on the corrected time delay matrix and the directional index matrix, the directed influence matrix of each sub-window is constructed by fusing them together, and the directed influence matrix on the sliding sub-window sequence is subjected to exponential moving average to obtain the smoothed time-varying directed influence matrix. Based on the smoothed time-varying directed influence matrix, the net outflow intensity of each node is calculated. The net outflow intensity is the sum of the influence intensity from the current node to all other nodes minus the sum of the influence intensity from all other nodes to the current node. The candidate source node set is sorted and selected according to the magnitude of the net outflow intensity, and the evolution trajectory of the time-varying directed influence matrix, the candidate source node set, and the arrival time sequence of each candidate source are output.

[0063] In this embodiment, this step performs online analysis using a sliding sub-window within the event window, providing arrival order, propagation direction, and candidate source nodes. First, the time delay cross-correlation is calculated and standardized:

[0064] in, For nodes With nodes In time shift The standardized cross-correlation function under the given conditions.

[0065] Peak lag was subsequently observed: And as a preliminary judgment of arrival order. In the target frequency band The coherence function and phase slope index are calculated above, assuming the autospectrum is... Mutual spectrum Phase coherence Obtain directional indicators .in, When the value is positive and significant, the net propagation direction is considered to be trending towards... Conversely .

[0066] This invention addresses the systematic deviations caused by minor asynchrony in the link by adjusting the global time offset. The estimation and compensation are performed using the following specific formula:

[0067] in, Weighted by signal-to-noise ratio and coherence.

[0068] Specifically, to enhance robustness, this invention avoids power frequency and its strong harmonics in frequency band selection, employs percentile scaling for spectral quantities and cross-correlation peaks, and uses median slope estimation for time series within the window to suppress occasional spikes. The above results are then subjected to exponential moving average over a sliding sub-window to form a time-varying directed influence matrix. , Indicates at time nodes For nodes The influence intensity is determined, and a candidate source node set is selected based on the ratio of "total outflow intensity minus total inflow intensity". The final output includes... The time trajectory, candidate source set, and arrival time sequence are used for subsequent fusion localization.

[0069] S5. Integrate the scoring results of physical consistency with the results of dynamic causal inference. When the two are consistent, locate directly; when the two are inconsistent, first distinguish between fault and normal fluctuation, and make segment judgment based on the confidence level and the weight of the physical residual.

[0070] As a preferred embodiment of the present invention, step S5 specifically includes: Based on the topological connection relationship of the electrical diagram, the line branches between adjacent monitoring nodes are constructed as a set of candidate segments, and the physical consistency score of each graph node and the edge fault score of each graph edge are obtained, as well as the set of candidate source nodes and the arrival time sequence of each node are obtained. For each candidate segment, adaptive fusion weights are calculated based on the physical consistency scores of the nodes at both ends of the segment and the physical residuals that characterize the deviation between the observation time difference and the physical model. Based on the edge fault score, the adaptive fusion weight, the candidate source node indication, and the observation arrival time difference of the nodes at both ends of the segment, the basic fusion score of each candidate segment is calculated. Based on the configured line physical prior parameters, the theoretical time difference of arrival for the fault disturbance to propagate along the candidate segment is calculated. The theoretical time difference is compared with the observed time difference of arrival to obtain the time difference error. A reward decay function is constructed based on the time difference error, and the basic fusion score is corrected using the reward decay function to obtain the final fusion score for each candidate segment. When determining the faulty segment based on the final fusion score, a hysteresis comparison mechanism is introduced. When the segment corresponding to the maximum final fusion score at the current moment is different from the faulty segment maintained at the previous moment, and the maximum final fusion score exceeds the sum of the final fusion score of the faulty segment maintained at the previous moment and the preset hysteresis boundary value, the faulty segment is switched to the segment corresponding to the current maximum score; otherwise, the faulty segment at the previous moment is maintained.

[0071] In this embodiment, this step fuses graph learning evidence and causal evidence on the candidate segment set composed of adjacent monitoring nodes, and introduces a time difference benchmark of propagation prior parameters for verification. (Segment definition) The fusion score is calculated using the following formula:

[0072]

[0073] in, Indicates a section The strength of causal evidence Indicates from node To the node The intensity of the directional influence, Indicates from node To the node The intensity of the directional influence, This indicates adaptive weights.

[0074] Its adaptive weights Adjustments are made based on physical consistency and residuals, using the following formula:

[0075] in, The physical residuals are based on the telegraph equations and endpoint measurements. These represent the learnable weight coefficients.

[0076] Specifically, the propagation prior model verification uses representative frequency bands. Using the phase velocity as a reference, calculate its theoretical time difference of arrival. The specific formula is as follows:

[0077] The time difference error was obtained by comparing it with the observed time difference of arrival. :

[0078] Time difference error Not exceeding tolerance Time with reward decay function enlarge Otherwise press attenuation Reward decay function The mathematical expression is as follows:

[0079] in, To avoid output jitter between adjacent segments, this invention introduces a hysteresis mechanism, which maintains the current positioning if the newly added maximum score does not significantly exceed the current segment score plus the safety boundary. Ultimately, the result is... Identify the faulty section and output its start and end nodes or sub-segments, event type, and overall confidence level. Simultaneously record key evidence indicators, including... The decision path is used for alarm linkage and post-event verification.

[0080] As can be seen, when both are consistent in segment and direction, the positioning between adjacent monitoring nodes is directly completed; when there is a discrepancy, fault / normal fluctuation is first judged, and then adaptive weighting is performed based on confidence level and physical residual, and consistency verification is performed on arrival time difference in combination with propagation prior, forming a robust segment-level decision. The output includes specific fault segments, event types, comprehensive confidence levels and key evidence indicators containing start and end nodes / sub-segments, providing a linkage interface with scheduling / protection / alarm systems; it achieves low-latency and highly selective online positioning without requiring full and accurate parameter identification, and supports multi-line parallel and incremental updates, meeting the needs of large-scale operation and maintenance and rapid response of new energy power stations.

[0081] The method of this invention is applicable to the online operation of the collection lines (radial / ring network, cable / overhead / mixed laying) of new energy power plants such as wind power and photovoltaics under the condition of having nanosecond-level synchronous voltage measurement. It aims to integrate physical information graph neural network and dynamic causal reasoning to intelligently distinguish between normal fluctuations and fault disturbances, accurately locate the fault section between adjacent monitoring nodes, and output the location, type and confidence level for linkage alarm.

[0082] In a second aspect, the present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements any of the above-described methods for identifying fault zones in a collector wire based on a physical information graph neural network and dynamic causal reasoning.

[0083] Thirdly, the present invention also provides an electronic device, including a processor and a memory, the memory storing a plurality of instructions; the processor loads instructions from the memory to execute the steps in any of the above-described methods for identifying fault zones in a collector wire based on a physical information graph neural network and dynamic causal reasoning.

[0084] In this embodiment, the computer-readable storage medium may be a non-transitory computer-readable storage medium, such as a read-only memory (ROM), random access memory (RAM), CD-ROM, magnetic tape, floppy disk, and optical data storage device.

[0085] Fourthly, the present invention also provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps in the method for identifying fault zones in a collector wire based on a physical information graph neural network and dynamic causal reasoning as described in any one of the claims.

[0086] In this embodiment, those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be accomplished by instructions, or by instructions controlling related hardware. These instructions can be stored in a computer-readable storage medium and loaded and executed by a processor.

[0087] Therefore, embodiments of this application provide a computer-readable storage medium storing a plurality of instructions that can be loaded by a processor to execute steps in any of the methods for identifying fault zones in a collector wire based on physical information graph neural networks and dynamic causal reasoning provided in embodiments of this application.

[0088] It should be noted that, through the above description of the implementation methods, those skilled in the art can clearly understand that, for the sake of convenience and brevity, only the division of the above functional modules is used as an example. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above.

[0089] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. The scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for identifying fault zones in a collector wire based on a physical information graph neural network and dynamic causal reasoning, characterized in that, The method includes the following steps: S1. Real-time acquisition of multi-node voltage transient data of the collector line, and preprocessing. The preprocessed data is used to form multi-node voltage transient samples based on threshold or sudden change detection trigger event window. S2. Based on the principle of primary wiring, the monitoring points are modeled as graph nodes and the line branches are graph edges. The physical and geometric prior parameters are configured for the graph edges to obtain an electrical diagram with physical priors. S3. Construct a physical information graph neural network, taking multi-node voltage transient samples and electrical graph as input, and outputting the physical consistency score between the graph nodes and graph edges in the electrical graph, the candidate propagation path and the confidence of the source node; S4. Online calculation of dynamic causal reasoning for multi-node transient sequences based on sliding time windows, obtaining time-varying directional influence relationships and perturbation propagation directions, and providing candidate source nodes and arrival time sequences; S5. Integrate the scoring results of physical consistency with the results of dynamic causal inference. When the two are consistent, locate directly; when the two are inconsistent, first distinguish between fault and normal fluctuation, and make segment judgment based on the confidence level and the weight of the physical residual.

2. The method for identifying fault zones in a collector wire based on a physical information graph neural network and dynamic causal reasoning according to claim 1, characterized in that, Step S1 specifically includes: High-speed synchronous voltage sensors are deployed at key nodes such as the wind turbine outlet, the low-voltage side of the box-type transformer, and the busbar junction of the collection line. Multi-node voltage transient sequences are obtained by sampling at the millisecond level, and data preprocessing is performed. The voltage change rate is calculated based on the obtained multi-node voltage transient sequence, and the maximum value is taken within the node set to characterize the voltage change intensity. An event is triggered when the maximum rate of change is not less than the rate of change threshold. At the same time, the square of the preprocessed voltage transient sequence of each node is integrated within the energy statistics window and summed over all nodes to characterize the cumulative energy of the transient intensity. An event is also triggered when the sum of this cumulative energy is not less than the energy threshold. When any triggering condition is met, an event time window is constructed by extending the event window forward and backward with the triggering time as the center. Within the event window, the preprocessed voltage transient sequence of each node is extracted to form a multi-node event dataset and output.

3. The method for identifying fault zones in a collector wire based on a physical information graph neural network and dynamic causal reasoning according to claim 1, characterized in that, The specific steps in step S2 for configuring the physical and geometric prior parameters for the graph edges include: Each graph edge is configured with prior physical parameters and geometric information such as impedance, inductance, capacitance, conductance, and geometric length, and an edge feature vector is formed to construct multi-scale temporal features for graph nodes.

4. The method for identifying fault zones in a collector wire based on a physical information graph neural network and dynamic causal reasoning according to claim 3, characterized in that, The electrical diagram with physical prior knowledge obtained in step S2 specifically includes: Based on the preprocessed voltage transient sequence of nodes within the event window, the feature components of five types of prior physical parameters are calculated on the same event window and then concatenated to form a node feature vector. The telegraph equations with distributed parameters are used as physical consistency constraints for the prior physical model of electromagnetic propagation. The propagation constant, phase velocity, and attenuation coefficient of the prior physical parameters in the frequency domain are calculated based on the electromagnetic propagation physical prior model. The propagation constant, phase velocity, and attenuation coefficient are then analytically mapped to the physical propagation index. This mapping result is then injected as a boundary-level physical prior into the graph data construction and model constraints to obtain an electrical graph with physical priors.

5. The method for identifying fault zones in a collector wire based on a physical information graph neural network and dynamic causal reasoning according to claim 4, characterized in that, Step S3 specifically includes: Using node feature vectors and edge feature vectors as initial embeddings, an initial embedding vector is assigned to each graph node based on the multi-node voltage transient samples and the electrical graph. The hidden representations of graph nodes and graph edges are obtained through graph message passing and attention updates. Each layer update includes: generating a message from the source node to the target node based on the physical prior parameters of the source node embedding, target node embedding, and edge configuration; aggregating all neighbor messages received by the target node, and inputting the aggregation result and the target node's current layer embedding into the update function to obtain the target node's next layer embedding. Based on the final node embedding obtained after multiple iterations, the confidence of the fault source node, the node physical consistency score, the edge fault score, and the confidence of the candidate propagation path are generated through multiple parallel output heads.

6. The method for identifying fault zones in a collector wire based on a physical information graph neural network and dynamic causal reasoning according to claim 1, characterized in that, Step S3 further includes constructing a joint loss function to optimize the physical information graph neural network during the model training phase. The joint loss function includes a data fitting loss term and residual constraint terms constructed based on Kirchhoff's current law, Kirchhoff's voltage law, and fault transient propagation partial differential equations, respectively.

7. The method for identifying fault zones in a collector wire based on a physical information graph neural network and dynamic causal reasoning according to claim 6, characterized in that, Step S4 specifically includes: A sliding sub-window is set within the event window. The time delay cross-correlation function is calculated based on the voltage transient time series data of each pair of monitoring nodes within the sub-window. The lag time between nodes is obtained by finding the cross-correlation peak. After standardization, a directed time delay estimation matrix is ​​formed for each sub-window, which serves as a preliminary criterion for arrival time series. On the preset target frequency band, the coherence function and phase spectrum are calculated based on the time series data of each pair of monitoring nodes, and then the phase slope index is calculated as a directional index. The net propagation direction between nodes is determined according to the positive and negative signs and significance of the directional index, and the directional index matrix of each sub-window is obtained. Based on the directed time delay estimation matrix and the directional index matrix, weights are constructed by combining the signal-to-noise ratio and coherence of each node pair. The global time offset vector is estimated by weighted optimization and the original time series data is compensated to obtain the corrected time delay matrix. Based on the corrected time delay matrix and the directional index matrix, the directed influence matrix of each sub-window is constructed by fusing them together, and the directed influence matrix on the sliding sub-window sequence is subjected to exponential moving average to obtain the smoothed time-varying directed influence matrix. Based on the smoothed time-varying directed influence matrix, the net outflow intensity of each node is calculated. The net outflow intensity is the sum of the influence intensity from the current node to all other nodes minus the sum of the influence intensity from all other nodes to the current node. The candidate source node set is sorted and selected according to the magnitude of the net outflow intensity, and the evolution trajectory of the time-varying directed influence matrix, the candidate source node set, and the arrival time sequence of each candidate source are output.

8. The method for identifying fault zones in a collector wire based on a physical information graph neural network and dynamic causal reasoning according to claim 7, characterized in that, Step S5 specifically includes: Based on the topological connection relationship of the electrical diagram, the line branches between adjacent monitoring nodes are constructed as a set of candidate segments, and the physical consistency score of each graph node and the edge fault score of each graph edge are obtained, as well as the set of candidate source nodes and the arrival time sequence of each node are obtained. For each candidate segment, adaptive fusion weights are calculated based on the physical consistency scores of the nodes at both ends of the candidate segment and the physical residuals that characterize the deviation between the observation time difference and the physical model. Based on the edge fault score, the adaptive fusion weight, the candidate source node indication, and the observation arrival time difference of the nodes at both ends of the segment, the basic fusion score of each candidate segment is calculated. Based on the configured line physical prior parameters, the theoretical time difference of arrival for the fault disturbance to propagate along the candidate segment is calculated. The theoretical time difference is compared with the observed time difference of arrival to obtain the time difference error. A reward decay function is constructed based on the time difference error, and the basic fusion score is corrected using the reward decay function to obtain the final fusion score for each candidate segment. When determining the faulty segment based on the final fusion score, a hysteresis comparison mechanism is introduced. When the segment corresponding to the maximum final fusion score at the current moment is different from the faulty segment maintained at the previous moment, and the maximum final fusion score exceeds the sum of the final fusion score of the faulty segment maintained at the previous moment and the preset hysteresis boundary value, the faulty segment is switched to the segment corresponding to the current maximum score; otherwise, the faulty segment at the previous moment is maintained.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements a method for identifying fault zones in a collector wire based on a physical information graph neural network and dynamic causal reasoning as described in any one of claims 1 to 8.

10. An electronic device comprising a processor and a memory, characterized in that, The memory stores multiple instructions, and the processor loads the instructions from the memory to execute the steps in the method for identifying collector fault zones based on physical information graph neural networks and dynamic causal reasoning as described in any one of claims 1 to 8.

Images