A graph learning fault location method, device and equipment under the condition of topological uncertainty and distribution network bidirectional power flow, and a readable storage medium

Abstract

The present application relates to the field of electric power technology, in particular to a kind of topology uncertain and under the condition of distribution network bidirectional power flow graph learning fault location method, device, equipment and readable storage medium, the method is after obtaining distributed power supply state data, construct unified space-time graph representation fusing dynamic power flow direction characteristics, and generate a limited number of high possibility topology alternative set for uncertain switch state;According to the reliability of each topology, the parallel inference results are weighted and fused to generate the comprehensive probability ranking and Top-K candidate list of fault section;Finally, extract the graph attention weight and key feature contribution, form the explainability output containing key evidence path and decision basis, to improve the reliability, robustness and operation transparency of positioning result.The present application significantly improves the fault location accuracy and decision confidence of complex active distribution network under the condition of bidirectional power flow by explicitly modeling topology uncertainty and constraining multi-hypothesis consistency.

Description

Technical Field

[0001] This invention relates to a graph learning fault location method, apparatus, device, and readable storage medium under conditions of topological uncertainty and bidirectional power flow in distribution networks, belonging to the field of power technology. Background Technology

[0002] With the advancement of my country's dual-carbon goals and the deepening construction of new power systems, the penetration rate of distributed photovoltaic (PV) and energy storage (DER) systems in active distribution networks is continuously increasing. While this transformation injects clean energy into the power grid, it also brings challenges such as bidirectional power flow, variable power flow direction, and frequent uncertainty in network topology due to adjustments in operating modes or communication anomalies. Traditional fault location methods, such as rule-based reasoning, impedance methods, or traveling wave methods, heavily rely on accurate network topology and unidirectional power flow assumptions. Under conditions of topological uncertainty and bidirectional power flow, their location accuracy and robustness decrease significantly.

[0003] Graph neural networks (GNNs) are widely used in power system analysis due to their ability to model non-Euclidean data structures. Existing research has attempted to utilize graph attention networks for fault localization, learning the coupling relationships of the power grid through the features of nodes and edges. However, existing graph learning methods typically assume that the network topology is precisely known, failing to explicitly handle topological uncertainties caused by missing switch states, communication errors, or untimely manual reporting. When the actual topology deviates from the model assumptions, the localization performance deteriorates sharply. Furthermore, the bidirectional power flow introduced by DER (Device Regulator) integration makes the electrical characteristics of faults more complex, rendering traditional feature construction methods based on unidirectional power flow inapplicable. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a graph learning fault location method, apparatus, device, and readable storage medium under conditions of topological uncertainty and bidirectional power flow in distribution networks, aiming to solve the above problems.

[0005] The technical solution of this invention to solve the above-mentioned technical problems is as follows: A graph learning fault location method under topological uncertainty and bidirectional power flow conditions in distribution networks, comprising the following steps: When a distribution network fault trigger signal is detected, multi-source heterogeneous time-series data within the steady-state and transient time windows before and after the fault are extracted based on the fault event time. A unified spatiotemporal attribute map of the distribution network is constructed based on the time-series data. The edge features of the spatiotemporal attribute map incorporate bidirectional power flow direction indicators for quantifying the power flow directionality of the lines. Identify uncertain switches with questionable states in the distribution network, assess the confidence level of the opening and closing states of each uncertain switch, and generate multiple high-confidence candidate topologies that satisfy the radial constraints of the distribution network based on the confidence level. This transforms the topology uncertainty problem into a parallel analysis problem of multiple deterministic topology scenarios. The unified spatiotemporal attribute graph is matched with the adjacency matrix of each candidate topology, and then input into the pre-trained dynamic graph attention network. Fault probability inference is performed in parallel on each candidate topology, and the fault probability distribution of each line section of the distribution network under each candidate topology is output. Weights are assigned based on the credibility of each candidate topology, and the fault probability distributions corresponding to each candidate topology are weighted and fused to obtain the comprehensive fault probability of each line section of the distribution network. Fault candidate sections are output in order of comprehensive fault probability. At the same time, the graph attention weights and key feature contributions of the dynamic graph attention network are extracted to generate interpretable evidence for the location decision.

[0006] In a preferred embodiment of this application, when a distribution network fault trigger signal is detected, the step of extracting multi-source heterogeneous time-series data within the steady-state time window before the fault and the transient time window after the fault, based on the fault event time, includes: When the system detects a protection activation signal, a voltage drop exceeding a threshold, or a surge in AMI undervoltage alarms in a specific area, it automatically triggers a location process. The system uses the event time as the basis for this process. Based on the baseline, extract the time window. The multi-source time-series data within the system forms the raw input for this positioning.

[0007] In a preferred embodiment of this application, the step of constructing a unified spatiotemporal attribute map of the distribution network based on the time-series data, wherein bidirectional power flow direction indicators for quantifying the directionality of line power flow are incorporated into the edge features of the spatiotemporal attribute map, includes: First, the captured spatiotemporal data is mapped into an attribute graph, where the vertex set V represents all bus nodes and line nodes equipped with measuring devices in the corresponding power grid; the edge set E represents the corresponding lines, transformers, and tie switches; and the vertex feature matrix X refers to the feature matrix X for each vertex. The statistical measures of the feature sequences within the time window are extracted to form the feature vector. eigenvectors Integrating three-phase voltage amplitude and phase angle Three-phase current amplitude Zero-sequence voltage and current The percentage of downstream users experiencing pressure loss at this node, derived from the Advanced Measurement System (AMI). Alarm action count from the protection information system ; and the currently available active and reactive power from the distributed energy monitoring system (DER). .

[0008] In a preferred embodiment of this application, the steps of identifying uncertain switches with questionable states in the distribution network, evaluating the confidence level of the opening and closing states of each uncertain switch, and generating multiple high-confidence candidate topologies that satisfy the radial constraints of the distribution network based on the confidence level, transform the topology uncertainty problem into a parallel analysis problem of multiple deterministic topology scenarios, include: First, identify all switches with suspicious states, and for each switch... Assess the confidence level of its state in open (0) and closed (1) states. Generate the top m most likely alternative topologies. This is obtained by solving the following optimization problem: ; in, It's a switch In topology The assumed state (0 or 1) in the context. Given radial constraints, the m topologies with the highest overall confidence are obtained through enumeration or heuristic search.

[0009] In a preferred embodiment of this application, the steps of matching the unified spatiotemporal attribute graph with the adjacency matrix of each candidate topology, inputting the graph into a pre-trained dynamic graph attention network, performing fault probability inference in parallel on each candidate topology, and outputting the fault probability distribution of each line segment of the distribution network under each candidate topology include: The constructed vertex features X and edge features U are compared with each candidate topology. adjacency matrix Combined, they are input into a pre-trained dynamic graph attention network. For topology The features of vertex V in the first layer of the network are updated as follows: ; in, For topology lower vertex The neighborhood group, For trainable weight matrix, For non-linear activation functions, the attention coefficients are... The calculation dynamically depends on the characteristics of vertices V and U, as well as the connecting edges. Features: ; in This represents vector concatenation. For attention vectors, An edge-level prediction layer is used to output a failure probability for each edge in the graph. The network output has one dimension. vector , representing in topology The probability of failure for each edge (segment) below. It can be represented as: .

[0010] In a preferred embodiment of this application, the step of setting weights based on the credibility of each candidate topology and performing weighted fusion on the fault probability distributions corresponding to each candidate topology to obtain the comprehensive fault probability of each line section of the distribution network includes: During inference, the inference results under m topologies are considered. The fusion is performed to obtain the final overall probability. for: ; Among them, weight It can be proportional to topological confidence or model determinism.

[0011] This application also provides a graph learning fault location device under conditions of topological uncertainty and bidirectional power flow in distribution networks, comprising: The analysis module is used to extract multi-source heterogeneous time-series data within the steady-state and transient time windows before and after the fault when a fault trigger signal is detected in the distribution network, based on the fault event time. A unified spatiotemporal attribute map of the distribution network is constructed based on the time-series data. The edge features of the spatiotemporal attribute map incorporate bidirectional power flow direction indicators for quantifying the power flow directionality of the lines. The evaluation module is used to identify uncertain switches with questionable status in the distribution network, evaluate the confidence level of the opening and closing status of each uncertain switch, and generate multiple high-confidence candidate topologies that meet the radial constraints of the distribution network based on the confidence level. This transforms the topology uncertainty problem into a parallel analysis problem of multiple deterministic topology scenarios. The inference module is used to match the unified spatiotemporal attribute graph with the adjacency matrix of each candidate topology, input the pre-trained dynamic graph attention network, perform fault probability inference in parallel on each candidate topology, and output the fault probability distribution of each line section of the distribution network under each candidate topology. The generation module is used to set weights based on the credibility of each candidate topology, perform weighted fusion on the fault probability distribution corresponding to each candidate topology, obtain the comprehensive fault probability of each line section of the distribution network, output the fault candidate sections according to the comprehensive fault probability, and extract the graph attention weight and key feature contribution of the dynamic graph attention network to generate interpretable evidence for the location decision.

[0012] In a preferred embodiment of this application, the analysis module includes: a detection unit, used to automatically trigger a location process when the system detects a protection start signal, a voltage drop exceeding a threshold, or a surge in AMI undervoltage alarms in a specific area, with the system using event time... Based on the baseline, extract the time window. The multi-source time-series data within the system forms the raw input for this positioning.

[0013] Furthermore, to achieve the above objectives, the present invention also provides a graph learning fault location device under conditions of topological uncertainty and bidirectional power flow in distribution networks. The graph learning fault location device under conditions of topological uncertainty and bidirectional power flow in distribution networks includes a processor, a memory, and a graph learning fault location program under conditions of topological uncertainty and bidirectional power flow in distribution networks stored in the memory and executable by the processor. When the graph learning fault location program under conditions of topological uncertainty and bidirectional power flow in distribution networks is executed by the processor, it implements the steps of the graph learning fault location method under conditions of topological uncertainty and bidirectional power flow in distribution networks as described above.

[0014] Furthermore, to achieve the above objectives, the present invention also provides a computer-readable storage medium storing a graph learning fault location program under conditions of topological uncertainty and bidirectional power flow in a distribution network. When the graph learning fault location program under conditions of topological uncertainty and bidirectional power flow in a distribution network is executed by a processor, it implements the steps of the graph learning fault location method under conditions of topological uncertainty and bidirectional power flow in a distribution network as described above.

[0015] The beneficial effects of this invention are as follows: When a distribution network fault trigger signal is detected, multi-source heterogeneous time-series data within the steady-state and transient time windows before and after the fault are extracted based on the fault event time. A unified spatiotemporal attribute graph of the distribution network is constructed based on the time-series data. The edge features of the spatiotemporal attribute graph incorporate bidirectional power flow direction indicators for quantifying the directionality of line power flow. Uncertain switches with questionable states in the distribution network are identified, and the confidence levels of the opening and closing states of each uncertain switch are evaluated. Based on the confidence levels, multiple high-confidence candidate topologies that satisfy the radial constraints of the distribution network are generated, transforming the topological uncertainty problem into multiple deterministic topological scenarios. Parallel analysis is employed; the unified spatiotemporal attribute graph is matched with the adjacency matrix of each candidate topology, input into a pre-trained dynamic graph attention network, and fault probability inference is performed in parallel on each candidate topology, outputting the fault probability distribution of each line segment of the distribution network under each candidate topology; weighted weights are set based on the credibility of each candidate topology, and the fault probability distributions corresponding to each candidate topology are weighted and fused to obtain the comprehensive fault probability of each line segment of the distribution network, and the fault candidate segments are output in order of comprehensive fault probability. At the same time, the graph attention weights and key feature contributions of the dynamic graph attention network are extracted to generate interpretable evidence for the location decision. In the fault triggering and data processing stages, this invention, by extracting multi-source time-series data before and after the fault and constructing a unified spatiotemporal attribute graph incorporating bidirectional power flow direction features, can adaptively match the bidirectional power flow scenario brought about by the high penetration of distributed power sources, effectively overcoming the defects of traditional methods that rely on the unidirectional power flow assumption, resulting in the failure of fault features and the inability to distinguish between upstream and downstream faults, thus improving the fault feature expression capability under complex power flow conditions. To address the topological uncertainty issues arising from missing switch states, communication anomalies, and operational mode adjustments in distribution networks, this invention transforms the uncertain topology problem into a parallel analysis of multiple deterministic scenarios by identifying uncertain switches, assessing state confidence, and generating a set of high-confidence candidate topologies. This fundamentally eliminates the strong dependence of existing methods on precise network topology, avoiding a sharp drop in location accuracy due to discrepancies between topology and model assumptions, and improving the stability and reliability of fault location under complex operating conditions. Employing multi-topology parallel inference and introducing cross-topology consistency constraint loss during training forces the model to learn deep fault characteristics unaffected by topology structure, enabling the model to maintain stable inference performance even in unseen topology change scenarios. This solves the problems of weak generalization ability and sensitivity to topology changes in traditional graph learning methods, further enhancing the robustness and versatility of the location method. Based on topology confidence and model determinism, multi-topology results are weighted and fused to obtain a high-confidence comprehensive fault probability. Interpretable decision-making criteria are generated through attention weights and key feature contributions, facilitating on-site maintenance verification and rapid fault troubleshooting.In summary, this invention can simultaneously adapt to bidirectional power flow and topological uncertainty scenarios, improving the accuracy of fault location while enhancing the anti-interference capability and interpretability of the method, effectively shortening fault handling time and reducing operation and maintenance costs. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the hardware structure of the graph learning fault location device under the conditions of topological uncertainty and bidirectional power flow in the distribution network involved in this invention. Figure 2 This is a flowchart illustrating the graph learning fault location method under topological uncertainty and bidirectional power flow conditions in distribution networks involved in this invention. Figure 3 This is a diagram of the attention network model under the conditions of topological uncertainty and bidirectional power flow in the distribution network involved in this invention. Figure 4 This is a diagram of the device architecture under the conditions of topological uncertainty and bidirectional power flow in the distribution network involved in this invention.

[0017] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0018] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0019] The graph learning fault location method under topological uncertainty and bidirectional power flow conditions of the distribution network involved in the embodiments of the present invention is mainly applied to the graph learning fault location device under topological uncertainty and bidirectional power flow conditions of the distribution network. The graph learning fault location device under topological uncertainty and bidirectional power flow conditions of the distribution network can be a PC, a portable computer, a mobile terminal or other device with display and processing functions.

[0020] Reference Figure 1 , Figure 1This is a schematic diagram of the hardware structure of a graph learning fault location device under topological uncertainty and bidirectional power flow conditions in a distribution network, as described in an embodiment of the present invention. In this embodiment, the graph learning fault location device under topological uncertainty and bidirectional power flow conditions may include a processor 1001 (e.g., a CPU), a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. The communication bus 1002 is used to establish communication between these components; the user interface 1003 may include a display screen or an input unit such as a keyboard; the network interface 1004 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface); the memory 1005 may be a high-speed RAM or a stable, non-volatile memory, such as a disk drive, and optionally, the memory 1005 may be a storage device independent of the aforementioned processor 1001.

[0021] Those skilled in the art will understand that Figure 1 The hardware structure shown does not constitute a limitation on graph learning fault location equipment under topological uncertainty and bidirectional power flow conditions in distribution networks. It may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0022] Continue to refer to Figure 1 , Figure 1 The memory 1005, which is a computer-readable storage medium, may include an operating system, a network communication module, and a graph learning fault location program under conditions of topological uncertainty and bidirectional power flow in the distribution network.

[0023] exist Figure 1 In this embodiment, the network communication module is mainly used to connect to the server and communicate with the server for data; while the processor 1001 can call the graph learning fault location program under topological uncertainty and bidirectional power flow conditions of the distribution network stored in the memory 1005, and execute the graph learning fault location method under topological uncertainty and bidirectional power flow conditions of the distribution network provided in this embodiment of the invention.

[0024] This invention provides a graph learning-based fault location method under conditions of topological uncertainty and bidirectional power flow in distribution networks, referencing... Figures 2-4 The steps include: S10, When a distribution network fault trigger signal is detected, multi-source heterogeneous time-series data within the steady-state and transient time windows before and after the fault are extracted based on the fault event time. A unified spatiotemporal attribute map of the distribution network is constructed based on the time-series data. The bidirectional power flow direction indicator for quantifying the power flow directionality of the line is incorporated into the edge features of the spatiotemporal attribute map. S20 identifies uncertain switches with questionable states in the distribution network, evaluates the confidence level of the opening and closing states of each uncertain switch, and generates multiple high-confidence candidate topologies that satisfy the radial constraints of the distribution network based on the confidence level, thus transforming the topology uncertainty problem into a parallel analysis problem of multiple deterministic topology scenarios. S30, the unified spatiotemporal attribute graph is matched with the adjacency matrix of each candidate topology, and the pre-trained dynamic graph attention network is input. Fault probability inference is performed in parallel on each candidate topology, and the fault probability distribution of each line section of the distribution network under each candidate topology is output. S40, weighted weights are set based on the credibility of each candidate topology, and the fault probability distributions corresponding to each candidate topology are weighted and fused to obtain the comprehensive fault probability of each line section of the distribution network. The fault candidate sections are sorted and output according to the comprehensive fault probability. At the same time, the graph attention weights and key feature contributions of the dynamic graph attention network are extracted to generate interpretable evidence for the location decision.

[0025] Specifically, in this embodiment, to address the reduced reliability of fault location caused by the increased complexity of bidirectional power flow and frequent uncertainty in network topology resulting from the high penetration of distributed energy sources in active distribution networks, a method is adopted that includes a unified construction method for spatiotemporal graph features of the power grid and a topology uncertainty modeling method, a multi-topology consistency constraint graph learning and inference method, and a multi-hypothesis result fusion and interpretable decision output method. Taking into account multi-source heterogeneous data such as node electrical measurements, advanced measurement system alarms, protection signals, and distributed power source status, a unified graph representation containing dynamic power flow direction features is constructed. For the topology uncertainty caused by missing or conflicting switch states, a limited set of high-probability topology candidates is generated based on logical verification and confidence assessment, transforming the uncertainty problem into a parallel analysis of multiple deterministic scenarios.

[0026] In some embodiments, when a distribution network fault trigger signal is detected, multi-source heterogeneous time-series data within the steady-state and transient time windows before and after the fault are extracted based on the fault event time. A unified spatiotemporal attribute graph of the distribution network is constructed based on the time-series data. The edge features of the spatiotemporal attribute graph incorporate bidirectional power flow direction indicators for quantifying the directionality of line power flow. The steps include: When the system detects a protection activation signal, a voltage drop exceeding a threshold, or a surge in AMI undervoltage alarms in a specific area, it automatically triggers a location process. The system uses the event time as the basis for this process. Based on the baseline, extract the time window. The multi-source time-series data within the system forms the raw input for this positioning.

[0027] When the system detects a protection activation signal, a voltage drop exceeding a threshold, or a surge in AMI undervoltage alarms in a specific area, it automatically triggers the location process. The system uses event timelines... Based on the baseline, extract the time window. Multi-source time-series data within the system forms the raw input for this localization. Typically, Use 2-5 cycles to capture the steady state before the fault. Use 10-20 cycles to capture transient characteristics after a fault.

[0028] First, the captured spatiotemporal data is mapped into an attribute graph. Where the vertex set V represents all bus nodes and line nodes equipped with measuring devices in the corresponding power grid; the edge set E represents the corresponding lines, transformers, and tie switches; and the vertex feature matrix X refers to the feature matrix X for each vertex. The feature vector is constructed by extracting the statistics (mean, variance, and mutation rate) of the feature sequences within the time window. This vector integrates the three-phase voltage amplitude and phase angle. Three-phase current amplitude Zero-sequence voltage and current The percentage of downstream users experiencing pressure loss at this node, derived from the Advanced Measurement System (AMI). Alarm action count from the protection information system ; and the currently available active and reactive power from the distributed energy monitoring system (DER). Taking time t as an example, its basic characteristics include: ; in, This represents the percentage of AMI (Automatic Machine Access) failure users within the power supply area of ​​that vertex. This is used to count the actions of associated protection devices.

[0029] The edge feature matrix U refers to the matrix U for each edge. (Connect vertices i and j) to construct feature vectors : ; in, It is in the switch state (0 / 1). As a key indicator of bidirectional power flow direction, it is used to quantitatively characterize the directionality of power flow in active distribution networks. A positive value indicates that power flows from node i to node j, while a negative value indicates the opposite. This feature is crucial for the model to adapt to bidirectional power flow and distinguish between upstream and downstream faults. Its calculation is based on the estimation of power difference between adjacent vertices. ; in, The active power estimated for vertices i and j. To avoid division by zero for small constants, sign is the sign function, and the result takes the values ​​of +1, -1, or 0.

[0030] In some of the descriptions, the steps of identifying uncertain switches with questionable states in the distribution network, evaluating the confidence level of the opening and closing states of each uncertain switch, and generating multiple high-confidence candidate topologies that satisfy the radial constraints of the distribution network based on the confidence level, transforming the topology uncertainty problem into a parallel analysis problem of multiple deterministic topology scenarios, include: The system first identifies all switches with questionable states. These switches may have missing remote signaling values ​​or logical conflicts with their associated electrical measurements (e.g., extremely small voltage difference across the line but zero current). For n uncertain switches, theoretically there are... This invention identifies several possible topologies. To control computational complexity, a confidence-based filtering strategy is employed, whereby the system reads a snapshot of the current topology and identifies an uncertain set of switches. For each switch Assess the confidence level of its state in open (0) and closed (1) states. Confidence level calculation can integrate electrical measurement logic and communication status.

[0031] Subsequently, the top m most likely alternative topologies are generated. This is achieved by solving the following optimization problem: ; in, It's a switch In topology The assumed state (0 or 1) in the context. The constraints are radial. The m topologies with the highest overall confidence are obtained through enumeration or heuristic search.

[0032] In some embodiments, the unified spatiotemporal attribute graph is matched with the adjacency matrix of each candidate topology, input into a pre-trained dynamic graph attention network, and fault probability inference is performed in parallel on each candidate topology to output the fault probability distribution of each line segment of the distribution network under each candidate topology.

[0033] In this embodiment, the attention mechanism in the GAT architecture endows the model with a unique ability: assigning differentiated weights to the neighboring nodes of each node. This design breaks the limitation of indiscriminate aggregation of neighbor information in graph convolutional networks, enabling the central node to focus on truly key features based on the strength of the correlation between itself and its neighbors when integrating neighborhood information, thereby exhibiting higher efficiency in capturing complex interactions between nodes.

[0034] As a core component of the GAT model, the graph attention layer is responsible for selectively aggregating local neighborhood information using an adaptive attention mechanism. It no longer uses the fixed averaging aggregation method of traditional graph convolution, but instead assigns differentiated importance to different neighbors based on the correlation between nodes, thereby achieving more effective information integration. The computational process of GAL can be divided into two key stages: the calculation of attention coefficients and the normalization of weights. Its overall structure is as follows: Figure 3 As shown.

[0035] Attention coefficient During the computation phase, for the central node and one of its neighboring nodes The model first maps its original features to a latent space specifically for attention computation through a learnable linear transformation. Then, it calculates the correlation score between the two features using operations such as dot product. This score essentially characterizes the degree of influence of the neighboring node on the central node. The higher the score, the closer the two are at the feature level, and the more important the neighboring information should be when aggregating features.

[0036] During the weight normalization phase, the Softmax function is used to calculate the attention coefficients of all neighboring nodes. This is transformed into a probability distribution, thus obtaining the attention weights of each neighbor node. After this processing, the weights of all neighbors are... The sum is normalized to 1, thus significantly amplifying the contributions of highly correlated neighbors during subsequent aggregation, while effectively suppressing interfering information from low-correlation neighbors. Finally, the features of each neighbor node are weighted by attention... The weighted summation yields the central node. GAL layer output features .

[0037] To further enhance the model's ability to model complex graph structures and its training stability, GAT introduces a multi-head attention mechanism. Under this mechanism, multiple attention heads run in parallel, each with an independent parameter matrix, mapping and aggregating node features from different semantic perspectives. For example, some attention heads may focus more on the attribute similarity between nodes, while others may emphasize topological proximity relationships. The outputs from multiple heads are then fused through feature concatenation or feature averaging, enabling the model to adaptively integrate information from multiple dimensions when dealing with graph data with diverse structures and complex relationships. It is this multi-level, multi-perspective attention aggregation mechanism that allows GAT to demonstrate powerful representation capabilities in tasks such as social network analysis and molecular graph property prediction.

[0038] During the implementation phase, the constructed vertex features X and edge features U are first compared with each generated candidate topology. adjacency matrix Combined, they are input into a pre-trained dynamic graph attention network. This network learns graph representations through multi-layered information passing and aggregation. For the topology... The features of vertex V in the first layer of the network are updated as follows: ; in, For topology lower vertex The neighborhood group, For trainable weight matrix, It is a non-linear activation function. Attention coefficients. The calculation dynamically depends on the characteristics of vertices V and U, as well as the connecting edges. Features: ; in This represents vector concatenation. This is the attention vector.

[0039] This allows the model to adaptively focus on the electrical paths most relevant to the fault. The network ultimately outputs a fault probability for each edge in the graph (i.e., a possible fault segment) through an edge-level prediction layer. Finally, the network will output a dimension of vector , representing in topology The probability of failure for each edge (segment) below. It can be represented as: ; A real-world fault event, under several reasonable topological assumptions, should exhibit consistent patterns of essential electrical anomalies and spatial correlations. To ensure model robustness, in addition to minimizing the standard cross-entropy localization loss between model predictions and actual fault labels... In addition, this invention introduces an extra consistency constraint loss. .

[0040] Therefore, during model training, the loss function is determined by the localization loss. And consistency constraint loss constitute: ; in, For hyperparameters, Let KL divergence be the constraint. This constraint forces the output probability distributions to be similar across different topologies. The model minimizes this constraint simultaneously through backpropagation. and It forces the learning of deep fault feature representations that do not change with the topology connection method, thus maintaining stable inference performance when facing topological uncertainty scenarios that have not appeared in training.

[0041] In some embodiments, weighted weights are set based on the credibility of each candidate topology, and the fault probability distributions corresponding to each candidate topology are weighted and fused to obtain the comprehensive fault probability of each line segment of the distribution network. Fault candidate segments are then sorted and output according to the comprehensive fault probability. Simultaneously, the graph attention weights and key feature contributions of the dynamic graph attention network are extracted to generate interpretable evidence for the location decision. The steps include: During inference, the inference results under m topologies are considered. The fusion process is performed. The final overall probability is... for: ; Among them, weight It can be proportional to topological confidence or model determinism, such as , For probability distribution The smaller the entropy, the higher the certainty.

[0042] Finally, according to Sort all segments and output. (e.g., K=3) candidate fault segments and their probability values. Simultaneously, interpretable evidence is generated, first by combining all topological data... Average attention weight The top few edges and nodes are visualized to form suspected fault paths, and then the faulty paths are listed. The vertex feature that contributes the most to the candidate segment probability is given in the end. Candidates in different topologies The probability ranking standard deviation is calculated below; the smaller the standard deviation, the more robust the result.

[0043] To systematically evaluate the overall performance of the model proposed in this invention, a multi-dimensional quantitative evaluation system is constructed, mainly covering four core indicators: accuracy, precision, recall, and F-score. These indicators together form the basis for a comprehensive measurement of the model's prediction results.

[0044] Before detailing the various indicators, it is necessary to first clarify the four basic correspondences that may occur between the model's predictions and the actual situation, as shown in Table 1. Here, True Positive Examples (TP) represent the number of samples correctly predicted as positive by the model, True Negative Examples (TN) represent the number of samples correctly predicted as negative by the model, False Positive Examples (FP) characterize the error type where the model misclassifies a negative class as a positive class, and False Negative Examples (FN) reflect the bias situation where the model misses a positive class as a negative class.

[0045] Table 1 Evaluation Indicators: Based on the above classification results, the definitions and significance of each evaluation indicator are as follows: Accuracy: This metric characterizes the proportion of correctly predicted samples out of all samples. It is calculated as the ratio of correctly predicted samples to the total number of samples, using the following formula: ; Accuracy reflects a model’s ability to distinguish between different samples from a global perspective. The higher the value, the better the model can distinguish between different samples.

[0046] Precision: The proportion of samples that are predicted to be TP out of all samples that are predicted to be positive. It is used to measure the accuracy of the model.

[0047] ; This refers to the percentage of samples that the model classifies as faults that are actually faults. This metric reflects the reliability of the model's predictions; a higher value indicates a lower false alarm rate.

[0048] Recall: The proportion of samples that are predicted to be positive (TP) out of all samples that are actually positive. It is used to measure the model's ability to identify positive samples.

[0049] ; In other words, recall reflects how many of the actual failures the model actually captured. The higher the recall rate, the lower the false negative rate.

[0050] F-score: Used to comprehensively evaluate the model's performance while balancing precision and recall.

[0051] ; The higher the value, the better the model achieves a balance between accuracy and completeness, maintaining high prediction accuracy without sacrificing too much recall, or controlling the false positive level while maintaining a high recall.

[0052] By comprehensively applying the above four indicators, the prediction accuracy, false alarm control capability, false alarm suppression capability, and overall performance balance of the model in fault location tasks can be systematically quantified, thereby providing a reliable basis for the objective evaluation of model performance.

[0053] This application provides a graph learning fault location device under conditions of topological uncertainty and bidirectional power flow in a distribution network, comprising: The analysis module is used to extract multi-source heterogeneous time-series data within the steady-state and transient time windows before and after the fault when a fault trigger signal is detected in the distribution network, based on the fault event time. A unified spatiotemporal attribute map of the distribution network is constructed based on the time-series data. The edge features of the spatiotemporal attribute map incorporate bidirectional power flow direction indicators for quantifying the power flow directionality of the lines. The evaluation module is used to identify uncertain switches with questionable status in the distribution network, evaluate the confidence level of the opening and closing status of each uncertain switch, and generate multiple high-confidence candidate topologies that meet the radial constraints of the distribution network based on the confidence level. This transforms the topology uncertainty problem into a parallel analysis problem of multiple deterministic topology scenarios. The inference module is used to match the unified spatiotemporal attribute graph with the adjacency matrix of each candidate topology, input the pre-trained dynamic graph attention network, perform fault probability inference in parallel on each candidate topology, and output the fault probability distribution of each line section of the distribution network under each candidate topology. The generation module is used to set weights based on the credibility of each candidate topology, perform weighted fusion on the fault probability distribution corresponding to each candidate topology, obtain the comprehensive fault probability of each line section of the distribution network, output the fault candidate sections according to the comprehensive fault probability, and extract the graph attention weight and key feature contribution of the dynamic graph attention network to generate interpretable evidence for the location decision.

[0054] The analysis module includes: The detection unit is used to automatically trigger the location process when the system detects a protection start signal, a voltage drop exceeding a threshold, or a surge in AMI undervoltage alarms in a specific area. The system uses the event time as the basis for the location process. Based on the baseline, extract the time window. The multi-source time-series data within the system forms the raw input for this positioning.

[0055] In addition, embodiments of the present invention also provide a computer-readable storage medium.

[0056] The present invention stores a graph learning fault location program under topological uncertainty and bidirectional power flow conditions in a computer-readable storage medium. When the graph learning fault location program under topological uncertainty and bidirectional power flow conditions is executed by a processor, it implements the steps of the graph learning fault location method under topological uncertainty and bidirectional power flow conditions as described above.

[0057] The method implemented when the graph learning fault location program under topological uncertainty and bidirectional power flow conditions of the distribution network is executed can be referred to in various embodiments of the graph learning fault location method under topological uncertainty and bidirectional power flow conditions of the present invention, and will not be repeated here.

[0058] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0059] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0060] This application can be used in a wide variety of general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices. This application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform specific tasks or implement specific abstract data types. This application can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0061] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0062] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are exhaustively listed. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0063] For those skilled in the art, various modifications and improvements can be made without departing from the concept of the present invention, and these modifications and improvements are all within the scope of protection of the present invention. The scope of protection of the present invention is defined by the appended claims.

Claims

1. A graph learning-based fault location method under conditions of topological uncertainty and bidirectional power flow in distribution networks, applied to active distribution networks containing distributed generation sources, comprising the following steps: When a distribution network fault trigger signal is detected, multi-source heterogeneous time-series data within the steady-state and transient time windows before and after the fault are extracted based on the fault event time. A unified spatiotemporal attribute map of the distribution network is constructed based on the time-series data. The edge features of the spatiotemporal attribute map incorporate bidirectional power flow direction indicators for quantifying the power flow directionality of the lines. Identify uncertain switches with questionable states in the distribution network, assess the confidence level of the opening and closing states of each uncertain switch, and generate multiple high-confidence candidate topologies that satisfy the radial constraints of the distribution network based on the confidence level. This transforms the topology uncertainty problem into a parallel analysis problem of multiple deterministic topology scenarios. The unified spatiotemporal attribute graph is matched with the adjacency matrix of each candidate topology, and then input into the pre-trained dynamic graph attention network. Fault probability inference is performed in parallel on each candidate topology, and the fault probability distribution of each line section of the distribution network under each candidate topology is output. Weights are assigned based on the credibility of each candidate topology, and the fault probability distributions corresponding to each candidate topology are weighted and fused to obtain the comprehensive fault probability of each line section of the distribution network. Fault candidate sections are output in order of comprehensive fault probability. At the same time, the graph attention weights and key feature contributions of the dynamic graph attention network are extracted to generate interpretable evidence for the location decision.

2. The graph learning fault location method under topological uncertainty and bidirectional power flow conditions in distribution networks according to claim 1, characterized in that, When a distribution network fault trigger signal is detected, multi-source heterogeneous time-series data is extracted within the steady-state and transient time windows before and after the fault, based on the fault event time. The steps include: When the system detects a protection activation signal, a voltage drop exceeding a threshold, or a surge in AMI undervoltage alarms in a specific area, it automatically triggers a location process. The system uses the event time as the basis for this process. Based on the baseline, extract the time window. The multi-source time-series data within the system forms the raw input for this positioning.

3. The graph learning fault location method under topological uncertainty and bidirectional power flow conditions in distribution networks according to claim 2, characterized in that, The steps of constructing a unified spatiotemporal attribute graph of the distribution network based on the time-series data, wherein bidirectional power flow direction indicators for quantifying the directionality of line power flow are incorporated into the edge features of the spatiotemporal attribute graph, include: First, the captured spatiotemporal data is mapped into an attribute graph, where the vertex set V represents all bus nodes and line nodes equipped with measuring devices in the corresponding power grid; the edge set E represents the corresponding lines, transformers, and tie switches; and the vertex feature matrix X refers to the feature matrix X for each vertex. The statistical measures of the feature sequences within the time window are extracted to form the feature vector. eigenvectors Integrating three-phase voltage amplitude and phase angle Three-phase current amplitude Zero-sequence voltage and current The percentage of downstream users experiencing pressure loss at this node, derived from the Advanced Measurement System (AMI). Alarm action count from the protection information system ; and the currently available active and reactive power from the distributed energy monitoring system (DER). .

4. The graph learning fault location method under topological uncertainty and bidirectional power flow conditions in distribution networks according to claim 3, characterized in that, The process involves identifying uncertain switches with questionable states in the distribution network, evaluating the confidence level of the opening and closing states of each uncertain switch, and generating multiple high-confidence candidate topologies that satisfy the radial constraints of the distribution network based on the confidence level. This transforms the topology uncertainty problem into a parallel analysis problem of multiple deterministic topology scenarios. The steps include: First, identify all switches with suspicious states, and for each switch... Assess the confidence level of its state in open (0) and closed (1) states. Generate the top m most likely alternative topologies. This is obtained by solving the following optimization problem: ； in, It's a switch In topology The assumed state (0 or 1) in the context. Given radial constraints, the m topologies with the highest overall confidence are obtained through enumeration or heuristic search.

5. The graph learning fault location method under topological uncertainty and bidirectional power flow conditions in distribution networks according to claim 4, characterized in that, The steps include: matching the unified spatiotemporal attribute graph with the adjacency matrix of each candidate topology, inputting the pre-trained dynamic graph attention network, performing fault probability inference in parallel on each candidate topology, and outputting the fault probability distribution of each line segment of the distribution network under each candidate topology. The constructed vertex features X and edge features U are compared with each candidate topology. adjacency matrix Combined, they are input into a pre-trained dynamic graph attention network. For topology The features of vertex V in the first layer of the network are updated as follows: ； in, For topology lower vertex The neighborhood group, For trainable weight matrix, For non-linear activation functions, the attention coefficients are... The calculation dynamically depends on the characteristics of vertices V and U, as well as the connecting edges. Features: ； in This represents vector concatenation. For attention vectors, An edge-level prediction layer is used to output a failure probability for each edge in the graph. The network output has one dimension. vector , representing in topology The probability of failure for each edge (segment) below. It can be represented as: 。 6. The graph learning fault location method under topological uncertainty and bidirectional power flow conditions in distribution networks according to claim 1, characterized in that, The steps include: setting weights based on the credibility of each candidate topology, weighting and fusing the fault probability distributions corresponding to each candidate topology to obtain the comprehensive fault probability of each line section of the distribution network; During inference, the inference results under m topologies are considered. The fusion is performed to obtain the final overall probability. for: ； Among them, weight It can be proportional to topological confidence or model determinism.

7. A graph learning fault location device under conditions of topological uncertainty and bidirectional power flow in a distribution network, characterized in that, include: The analysis module is used to extract multi-source heterogeneous time-series data within the steady-state and transient time windows before and after the fault when a fault trigger signal is detected in the distribution network, based on the fault event time. A unified spatiotemporal attribute map of the distribution network is constructed based on the time-series data. The edge features of the spatiotemporal attribute map incorporate bidirectional power flow direction indicators for quantifying the power flow directionality of the lines. The evaluation module is used to identify uncertain switches with questionable status in the distribution network, evaluate the confidence level of the opening and closing status of each uncertain switch, and generate multiple high-confidence candidate topologies that meet the radial constraints of the distribution network based on the confidence level. This transforms the topology uncertainty problem into a parallel analysis problem of multiple deterministic topology scenarios. The inference module is used to match the unified spatiotemporal attribute graph with the adjacency matrix of each candidate topology, input the pre-trained dynamic graph attention network, perform fault probability inference in parallel on each candidate topology, and output the fault probability distribution of each line section of the distribution network under each candidate topology. The generation module is used to set weights based on the credibility of each candidate topology, perform weighted fusion on the fault probability distribution corresponding to each candidate topology, obtain the comprehensive fault probability of each line section of the distribution network, output the fault candidate sections according to the comprehensive fault probability, and extract the graph attention weight and key feature contribution of the dynamic graph attention network to generate interpretable evidence for the location decision.

8. The graph learning fault location device under topological uncertainty and bidirectional power flow conditions in distribution networks according to claim 7, characterized in that, The analysis module includes: The detection unit is used to automatically trigger the location process when the system detects a protection start signal, a voltage drop exceeding a threshold, or a surge in AMI undervoltage alarms in a specific area. The system uses the event time as the basis for the location process. Based on the baseline, extract the time window. The multi-source time-series data within the system forms the raw input for this positioning.

9. A graph learning fault location device under conditions of topological uncertainty and bidirectional power flow in distribution networks, characterized in that, The system includes a processor, a memory, and a computer program stored in the memory and executable by the processor, wherein when the computer program is executed by the processor, it implements the steps of the graph learning fault location method under topological uncertainty and bidirectional power flow conditions of distribution networks as described in any one of claims 1 to 6.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, wherein when the computer program is executed by a processor, it implements the steps of the graph learning fault location method under topological uncertainty and bidirectional power flow conditions of distribution network as described in any one of claims 1 to 6.

Images