A power distribution network fault section location method based on contrast topological time sequence graph neural network

Abstract

This invention discloses a method for locating fault sections in distribution networks based on a contrastive topological time-series graph neural network, comprising the following steps: acquiring time-series measurement data and nominal topology of distribution network nodes; dividing a complete fault recording sample into multiple patches; constructing a feature perturbation view and a topological perturbation view for each patch, and obtaining patch-level contrastive representations through a multi-kernel graph convolutional encoder; enhancing the robustness of local topology changes and measurement perturbations based on a topology consistency contrastive loss enhancement model; encoding unperturbed patches into node-level representations and constructing patch-level edge interaction sequences; utilizing a time-series graph neural network with a memory mechanism for dynamic message passing and node memory updates; and finally aggregating the high-level representations within the post-fault analysis window and outputting the fault section location result. This invention can collaboratively characterize local time-series features, topology consistency, and fault propagation dynamics, and is applicable to fault section location in distribution networks with distributed generation, slight topology mismatch, and switch reconfiguration conditions.

Description

Technical Field

[0001] This invention relates to the fields of power system fault location, distribution automation, graph neural networks and intelligent signal processing, specifically a method for locating fault sections in distribution networks based on a comparative topological time-series graph neural network. Background Technology

[0002] The power distribution network is a crucial component of the modern power system, and its operational safety and power supply reliability directly impact industrial production and residential life. With the widespread application of distribution automation, distributed power sources, and multi-source measurement devices, fault diagnosis in power distribution networks is gradually shifting from relying on manual inspections and protection action information to rapid location methods based on massive measurement data and intelligent algorithms.

[0003] Existing data-driven fault location methods typically utilize convolutional neural networks, recurrent neural networks, or traditional machine learning methods to extract fault features from recorded signals. While these methods can reflect changes in electrical quantities such as voltage and current after a fault to some extent, they often focus primarily on the temporal characteristics of a single measurement point or a local area, making it difficult to fully utilize the spatial dependencies inherent in the connection relationships of distribution network lines.

[0004] Graph neural networks can abstract a power distribution network as a graph structure composed of nodes and edges, and learn the spatial relationships between nodes through neighborhood aggregation, thus being used for power distribution network fault location tasks. However, most traditional graph neural networks rely on static adjacency matrices or single-moment data for inference, making it difficult to characterize the continuous dynamic process of fault disturbances propagating along the line.

[0005] In active distribution networks containing distributed generation sources, the integration of power sources such as photovoltaics, wind power, and energy storage alters local power flow distribution, short-circuit current levels, and fault response characteristics. Simultaneously, operations such as tie switch switching, feeder transfer, fault isolation, and network reconfiguration can cause deviations between the effective topology during operation and the nominal topology used in the modeling phase. If the model relies excessively on a fixed adjacency matrix during training, when minor changes in edge connections occur during testing or switching actions occur within the same fault analysis window, the information propagation path may differ from the actual physical propagation path, leading to a decrease in fault location accuracy.

[0006] Therefore, there is an urgent need for a fault section location method that can simultaneously utilize local transient time-series information, multi-scale topology information of the distribution network, and fault propagation memory information to improve the robustness and accuracy of fault section location in scenarios such as distributed power source access, slight topology mismatch, and switch reconfiguration. Summary of the Invention

[0007] The purpose of this invention is to provide a method for locating fault sections in distribution networks based on a comparative topology time-series neural network, which addresses the problems of existing methods being overly dependent on fixed topologies, insufficient dynamic characterization of fault propagation, and low robustness of location results under conditions of slight topology mismatch, switch reconfiguration, and distributed power source access in active distribution networks.

[0008] To achieve the above objectives, the present invention adopts the following technical solution, comprising the following steps:

[0009] 1) Obtain time-series measurement data of distribution network nodes, a set of candidate fault sections, and the nominal topology of the distribution network, and construct a nominal topology graph based on the connection relationships between distribution network nodes and lines;

[0010] 2) Take a complete fault recording sample as a fault sample unit, i.e., a fault episode, and divide the fault episode into multiple local time segments arranged in chronological order according to the preset patch length and step size.

[0011] 3) Construct a feature perturbation view and a topology perturbation view for each local time series segment. The feature perturbation view keeps the nominal topology unchanged and applies perturbation to the measurement features. The topology perturbation view keeps the measurement features unchanged and applies local edge connection perturbation to the nominal topology.

[0012] 4) Input the feature perturbation view and the topology perturbation view into the multi-kernel graph convolutional encoder to obtain patch-level contrastive representations under the two enhanced views, and constrain the cross-view representations of the same patch to maintain consistency based on the topology consistency contrastive loss;

[0013] 5) Input the unperturbed local time sequence into the multi-kernel graph convolutional encoder to obtain the node-level patch representation, and construct the patch-level edge interaction sequence according to the physical lines in the nominal topology;

[0014] 6) Input the Patch-level edge interaction sequence into a time graph neural network with a memory mechanism to perform interaction message generation, message aggregation, node memory update and node embedding readout to obtain node embeddings that reflect the dynamics of fault propagation;

[0015] 7) Aggregate the node embeddings within the post-fault analysis Patch window to obtain an episode-level global representation, and output the predicted probability of each candidate fault segment through a classifier. Take the candidate fault segment with the highest predicted probability as the fault segment localization result.

[0016] 8) During the model training phase, the supervised classification loss is calculated based on the true labels of the fault segments, and the model parameters are jointly optimized using the supervised classification loss and the topology consistency comparison loss. During the online localization phase, no feature perturbation view or topology perturbation view is constructed, and no topology consistency comparison loss is calculated. Instead, the trained model is used to infer the fault episode to be tested to obtain the fault segment localization result.

[0017] Furthermore, the timing measurement data of the distribution network nodes includes the three-phase voltage, three-phase current, zero-sequence voltage, and zero-sequence current of the distribution network measurement nodes.

[0018] Furthermore, the set of candidate fault segments is as follows: ,in The number of candidate fault sections, and each candidate fault section It corresponds to one physical line.

[0019] Furthermore, the nominal topology of the distribution network is ,in For a set of nodes, For the nominal line connection set, This is the corresponding adjacency matrix.

[0020] Furthermore, the effective topology diagram during the testing or operation phase is denoted as... , For a set of valid line connections; when and If there is an inconsistency, it is determined that there is a topology mismatch or a topology change caused by switch reconfiguration.

[0021] Furthermore, in step 2), the spatiotemporal graph tensor of the s-th fault episode is represented as: ,in Where N is the number of sampling points, C is the number of nodes, and P is the feature dimension of a single node. The number of patches is obtained by dividing the data according to the patch length P and stride S. .

[0022] Furthermore, the feature perturbation view and the topology perturbation view are represented as follows:

[0023] , in, These are the original node features. This indicates that random noise, masking, or amplitude perturbation is applied to the measurement features within the patch. This represents small-scale perturbations such as adding, deleting, or replacing local edge connections in the nominal topology.

[0024] Further, in step 4), the multi-kernel graph convolutional encoder is used to learn structural priors at different propagation scales. For a patch, the multi-kernel graph convolutional encoder uses multiple adjacency propagation kernels, diffusion kernels, or Chebyshev polynomial kernels of different orders to encode node features, and then obtains a patch-level representation through concatenation, linear mapping, and pooling.

[0025] Further, in step 4), the topology consistency comparison loss uses the feature perturbation view representation and topology perturbation view representation of the same patch in the same fault episode as positive sample pairs, and uses the representation of non-corresponding patches in the same fault episode and the patch representation of heterogeneous fault samples in mini-batch as negative samples.

[0026] Furthermore, in step 5), for any undirected physical path {u,v} in the nominal topology, construct respectively and Two-way edge interaction records; the features of the edge interaction records are obtained by concatenating the source node Patch representation, the target node Patch representation, and the difference features between the two.

[0027] Furthermore, the Patch-level edge interaction sequence is organized in Patch order, and each interaction record in the edge interaction sequence includes a source node, a target node, a Patch index, a physical timestamp, and edge interaction features.

[0028] Further, in step 6), the time-graph neural network with memory mechanism generates interactive messages based on the source node memory state, target node memory state, edge interaction features and time interval before the event occurs; after aggregating the messages received by the same node, the node memory state is updated and the node embedding is read out.

[0029] Furthermore, in step 7), the node embeddings within the post-fault analysis patch window are subjected to average pooling, weighted pooling, or attention pooling to obtain an episode-level global representation. The probability distribution of candidate fault segments is output through a Softmax classifier. .

[0030] Furthermore, in step 8), the model training phase uses supervised classification loss and topology consistency comparison loss to form a joint objective function; in the online localization phase, the enhanced view is no longer constructed and the topology consistency comparison loss is no longer calculated, but the trained model is directly used to infer the fault episode to be tested.

[0031] The technical effects of this invention are undeniable. This invention reduces the computational overhead of time-by-time dynamic graph inference for long sequences through patch-level local temporal representation; it enhances the model's robustness to measurement noise, local topology shifts, and minor topology mismatches by constructing topology consistency contrastive learning constraints through feature perturbation views and topology perturbation views; it characterizes the continuous evolution of fault perturbation propagation along distribution network lines through patch-level edge interaction sequences and time-graph neural networks with memory mechanisms; and it achieves accurate localization of fault sections in complex active distribution networks through episode-level aggregation and classification outputs. Attached Figure Description

[0032] Figure 1 This is the overall flowchart of the present invention;

[0033] Figure 2 A diagram illustrating the patch division of a fault episode;

[0034] Figure 3 A schematic diagram is constructed for the feature perturbation view and the topology perturbation view;

[0035] Figure 4 This is a schematic diagram of a time-graph neural network structure with a memory mechanism;

[0036] Figure 5 A schematic diagram for locating the faulty section is provided. Detailed Implementation

[0037] The present invention will be further described below with reference to embodiments, but it should not be construed that the scope of the present invention is limited to the following embodiments. Various substitutions and modifications made based on ordinary technical knowledge and common practices in the art without departing from the above-described technical concept of the present invention should be included within the scope of protection of the present invention.

[0038] Example 1

[0039] See Figures 1 to 5 A method for locating fault sections in a distribution network based on a comparative topology time-series neural network includes the following steps:

[0040] 1) Obtain time-series measurement data of distribution network nodes, a set of candidate fault sections, and the nominal topology of the distribution network, and construct a nominal topology graph based on the connection relationships between distribution network nodes and lines;

[0041] 2) Take a complete fault recording sample as a fault episode, and divide the fault episode into multiple local time segments arranged in chronological order according to the preset patch length and step size;

[0042] 3) Construct a feature perturbation view and a topology perturbation view for each local time series segment. The feature perturbation view keeps the nominal topology unchanged and applies perturbation to the measurement features. The topology perturbation view keeps the measurement features unchanged and applies local edge connection perturbation to the nominal topology.

[0043] 4) Input the feature perturbation view and the topology perturbation view into the multi-kernel graph convolutional encoder to obtain patch-level contrastive representations under the two enhanced views, and constrain the cross-view representations of the same patch to remain consistent based on the topology consistency contrastive loss.

[0044] 5) Input the unperturbed local time sequence into the multi-kernel graph convolutional encoder to obtain the node-level patch representation, and construct the patch-level edge interaction sequence according to the physical lines in the nominal topology;

[0045] 6) Input the Patch-level edge interaction sequence into a time graph neural network with a memory mechanism to perform interaction message generation, message aggregation, node memory update and node embedding readout to obtain node embeddings that reflect the dynamics of fault propagation.

[0046] 7) Aggregate the node embeddings within the post-fault analysis Patch window to obtain an episode-level global representation, and output the predicted probability of each candidate fault segment through a classifier. Take the candidate fault segment with the highest probability as the fault segment localization result.

[0047] 8) During the model training phase, the supervised classification loss is calculated based on the true labels of the fault segments, and the model parameters are jointly optimized using the supervised classification loss and the topology consistency comparison loss. During the online localization phase, the feature perturbation view and the topology perturbation view are not constructed, and the topology consistency comparison loss is not calculated. Instead, the trained model is used to execute steps 1), 2), and 5) to 7) of the fault episode to be tested to obtain the fault segment localization result.

[0048] The timing measurement data of the distribution network nodes include the three-phase voltage, three-phase current, zero-sequence voltage, and zero-sequence current of the distribution network measurement nodes.

[0049] The set of candidate fault sections is as follows: Where K is the number of candidate fault sections, and each candidate fault section It corresponds to one physical line.

[0050] The nominal topology of the distribution network is Where V is the set of nodes, For the nominal line connection set, This is the corresponding adjacency matrix.

[0051] A valid topology diagram during the testing or runtime phase is ,in, This is the set of valid line connections. When... and If there is an inconsistency, it is determined that there is a topology mismatch or a topology change caused by switch reconfiguration.

[0052] In step 2), the original spatiotemporal graph tensor of the fault episode is: Where T is the number of sampling points, N is the number of nodes, and C is the feature dimension of a single node. The number of patches is obtained by dividing the area according to the patch length P and the stride S. And form a patch sequence according to the time position of each patch in the original fault record.

[0053] In step 3), the feature perturbation view is represented as The topology perturbation view is represented as .

[0054] in, These are the original node features. This indicates a disturbance in the measurement characteristics. This represents the perturbation caused by adding or deleting local topological edge connections. The feature perturbation view maintains the nominal topology unchanged, and the topological perturbation view maintains the measurement features unchanged.

[0055] In step 4), the multi-kernel graph convolutional encoder includes at least two sets of graph convolutional kernels with different propagation orders or different receptive fields, used to extract structural propagation features at different scales on the nominal topology and topological perturbation view. The multi-kernel outputs are then concatenated, mapped, or pooled to obtain a patch-level contrastive representation.

[0056] For the p-th patch of the s-th fault episode, its feature perturbation view representation and topology perturbation view representation are used as positive sample pairs, and the representations of non-corresponding patches in the same fault episode and the patch representations of heterogeneous fault samples in the current mini-batch are used as negative samples. The topology consistency contrast loss is used for training.

[0057] In step 5), for any undirected physical path {u,v} in the nominal topology, construct... and The interaction records of edges in two directions.

[0058] The features of the edge interaction record are obtained by concatenating the source node patch representation, the target node patch representation, and the difference features between the two. The patch-level edge interaction sequence is organized in patch order, and each interaction record includes the source node, target node, patch index, physical timestamp, and edge interaction features.

[0059] In step 6), the time-graph neural network with memory mechanism generates interactive messages based on the memory states of the source node and the target node before the interactive event occurs, the edge interaction features and the time interval, aggregates the messages received by the same node, and updates the node memory state through a gated recurrent unit, a long short-term memory unit or a multilayer perceptron, and then reads out the node embedding.

[0060] In step 7), the node embeddings within the post-fault analysis patch window are subjected to average pooling, weighted pooling, or attention pooling to obtain an episode-level global representation. The probability distribution of candidate fault segments is output through a Softmax classifier. .

[0061] In step 8), the model training phase uses supervised classification loss and topology consistency comparison loss to form a joint objective function; in the online localization phase, the enhanced view is no longer constructed and the topology consistency comparison loss is no longer calculated, but the trained model is directly used to infer the fault episode to be tested.

[0062] Example 2

[0063] A method for locating fault sections in a distribution network based on a contrastive topology time-series neural network includes the following steps:

[0064] 1) Obtain time-series measurement data of distribution network nodes, a set of candidate fault sections, and the nominal topology of the distribution network, and construct a nominal topology graph based on the connection relationships between distribution network nodes and lines;

[0065] 2) Take a complete fault recording sample as a fault episode, and divide the fault episode into multiple local time segments arranged in chronological order according to the preset patch length and step size;

[0066] 3) Construct a feature perturbation view and a topology perturbation view for each local time segment;

[0067] 4) Input the feature perturbation view and topology perturbation view into the multi-kernel graph convolutional encoder to obtain patch-level contrastive representations under the two enhanced views, and calculate the topology consistency contrastive loss;

[0068] 5) Input the undisturbed local temporal segments into the multi-kernel graph convolutional encoder to obtain node-level patch representations and construct patch-level edge interaction sequences;

[0069] 6) Input the Patch-level edge interaction sequence into a time-graph neural network with a memory mechanism to obtain node embeddings that reflect the dynamics of fault propagation;

[0070] 7) Aggregate the node embeddings within the post-fault analysis Patch window, output the predicted probability of each candidate fault segment through the classifier, and determine the fault segment location result.

[0071] 8) During the training phase, supervised classification loss and topology consistency comparison loss are used to jointly optimize the model parameters. During the online localization phase, the trained model is used to output the fault segment localization results of the episode to be tested.

[0072] Example 3

[0073] A method for locating fault sections in a distribution network based on a contrastive topology time-series neural network, with the same technical content as Embodiment 2, but further:

[0074] The timing measurement data of the distribution network nodes are the three-phase voltage, three-phase current, zero-sequence voltage, and zero-sequence current of the distribution network measurement nodes.

[0075] The set of candidate fault sections is as follows: Where K is the number of candidate fault sections, and each candidate fault section It corresponds to a single physical line, or a set of continuous physical lines between adjacent switch nodes, measurement nodes, or busbar nodes.

[0076] The nominal topology of the distribution network is Where V is the set of nodes, For the nominal line connection set, This is the corresponding adjacency matrix.

[0077] The effective topology diagram for the testing or operational phase is as follows: ,in, This is the set of valid line connections. When... and If there is an inconsistency, it is determined that there is a topology mismatch or a topology change caused by switch reconfiguration.

[0078] Example 4

[0079] A method for locating fault sections in a distribution network based on a contrastive topology time-series neural network, with the same technical content as any one of Embodiments 2 to 3, further comprising:

[0080] In step 2), a complete fault recording sample is taken as a fault episode.

[0081] The original spacetime graph tensor of the fault episode is: Where T is the number of sampling points, N is the number of nodes, and C is the feature dimension of a single node.

[0082] The fault episode is divided according to the patch length P and step size S to obtain the number of patches. Each patch is arranged sequentially according to its temporal position in the complete fault record, forming a local time sequence.

[0083] Example 5

[0084] A method for locating fault sections in a distribution network based on a contrastive topology time-series neural network, with the same technical content as any one of Embodiments 2 to 4, further comprising:

[0085] In step 3), a feature perturbation view and a topology perturbation view are constructed for each original local time series segment.

[0086] The feature perturbation view is represented as follows:

[0087] .

[0088] The topology perturbation view is represented as follows:

[0089] .

[0090] in, These are the original node features. This indicates a disturbance in the measurement characteristics. This represents the perturbation caused by adding or deleting edges in the local topology.

[0091] The feature perturbation view keeps the nominal topology unchanged and applies perturbation to the measurement features within the patch; the topology perturbation view keeps the measurement features within the patch unchanged and applies local edge connection perturbation to the nominal topology.

[0092] Example 6

[0093] A method for locating fault sections in a distribution network based on a contrastive topology time-series neural network, with the same technical content as any one of Embodiments 2 to 5, further comprising:

[0094] In step 4), the multi-kernel graph convolution encoder includes at least two sets of graph convolution kernels with different propagation orders or different receptive fields.

[0095] Each graph convolution kernel is used to extract the distribution network structure features at different propagation scales, and to encode the node features on the nominal topology and topology perturbation views.

[0096] The outputs of the convolution kernels of each graph are concatenated, mapped, or pooled to obtain patch-level comparative representations under the two enhanced views.

[0097] Example 7

[0098] A method for locating fault sections in a distribution network based on a contrastive topology time-series neural network, with the same technical content as any one of Embodiments 2 to 6, further comprising:

[0099] In step 4), for the p-th patch of the s-th fault episode, the feature perturbation view representation and the topology perturbation view representation corresponding to the patch are taken as positive sample pairs.

[0100] Other patches in the same fault episode that do not correspond to the p-th patch are taken as negative samples, and the patch representations corresponding to fault episodes with different fault segment categories in the current mini-batch are taken as negative samples.

[0101] The topology consistency comparison loss is calculated using the positive and negative sample pairs mentioned above, so that the representation of the same patch remains consistent under different perturbation conditions, and the patch representation of different propagation stages or outlier fault samples is distinguished.

[0102] Example 8

[0103] A method for locating fault sections in a distribution network based on a contrastive topology time-series neural network, with the same technical content as any one of Embodiments 2 to 7, further comprising:

[0104] In step 5), the original local temporal segments without perturbation are input into the multi-kernel graph convolutional encoder to obtain the node-level patch representation of each node under the corresponding patch.

[0105] For any undirected physical path {u,v} in the nominal topology, construct respectively and The interaction records of edges in two directions.

[0106] The features of the edge interaction record are obtained by concatenating the source node patch representation, the target node patch representation, and the difference features between the source node patch representation and the target node patch representation.

[0107] The patch-level edge interaction sequence is organized according to the temporal order of the patches. Each interaction record in the edge interaction sequence includes a source node, a target node, a patch index, a physical timestamp, and edge interaction features.

[0108] Example 9

[0109] A method for locating fault sections in a distribution network based on a contrastive topology time-series neural network, with the same technical content as any one of Embodiments 2 to 8, further comprising:

[0110] In step 6), the edge interaction sequence organized according to the Patch order is input into the time graph neural network with a memory mechanism.

[0111] For each edge interaction record, the temporal graph neural network generates an interaction message based on the memory states of the source node and the target node before the interaction event occurs, the current edge interaction characteristics, and the time interval.

[0112] When a node receives multiple interaction messages, it aggregates the interaction messages and updates the node's memory state based on the aggregated interaction messages.

[0113] The node memory state is updated through a gated cyclic unit, a long short-term memory unit, or a multilayer perceptron. After the node memory update is completed, the node embedding of each node under the current patch is read out.

[0114] Example 10

[0115] A method for locating fault sections in a distribution network based on a contrastive topology time-series neural network, with the same technical content as any one of Embodiments 2 to 9, further comprising:

[0116] In step 7), select the post-fault analysis patch window based on the time the fault occurred.

[0117] The node embeddings within the post-fault analysis patch window are subjected to average pooling, weighted pooling, or attention pooling to obtain the episode-level global representation corresponding to the fault episode. .

[0118] The episode-level global representation Input a Softmax classifier and output the probability distribution of the fault episode belonging to each candidate fault segment. .

[0119] The candidate fault segment with the highest predicted probability is taken as the fault segment location result.

[0120] Example 11

[0121] A method for locating fault sections in a distribution network based on a contrastive topology time-series neural network, with the same technical content as any one of Embodiments 2 to 10, further comprising:

[0122] In step 8), during the model training phase, the supervised classification loss is calculated based on the true labels of the fault segments in each fault episode, and the topological consistency comparison loss between the feature perturbation view and the topological perturbation view is calculated.

[0123] The supervised classification loss and the topology consistency comparison loss are used to form a joint objective function, and the joint objective function is used to optimize the model parameters of the multi-kernel graph convolutional encoder, the time graph neural network with memory mechanism, and the classifier.

[0124] During the online localization phase, no feature perturbation view or topology perturbation view is constructed, and no topology consistency contrast loss is calculated.

[0125] The fault episode to be tested is divided into segments with the same patch length and stride as the training phase. The unperturbed local temporal segments are input into the trained multi-kernel graph convolutional encoder. Patch-level edge interaction sequences are constructed, node memories are updated, node embeddings are read out, and the probability of candidate fault segments is calculated to obtain the fault segment localization result.

[0126] Example 12

[0127] This embodiment takes the IEEE 30-node distribution network as an example to illustrate the method for locating fault sections in a distribution network based on a comparative topology time sequence graph neural network. The specific steps are as follows.

[0128] 1) Construct the IEEE 30-node distribution network model.

[0129] An IEEE 30-node distribution network model was built based on the PSCAD simulation platform. The system power frequency was set to 50 Hz, the total system load was set to 3.715 MW, and photovoltaic power generation units were configured at the preset distributed power access nodes.

[0130] The bus nodes and measurement nodes in the IEEE 30-node distribution network are abstracted as graph nodes, and the physical line connections between nodes are abstracted as graph edges. The nominal topology graph of the IEEE 30-node distribution network is constructed based on the node and line connection relationships.

[0131] Each physical line in the nominal topology is considered a candidate fault segment. A group of continuous physical lines between adjacent switch nodes, measurement nodes, or bus nodes can also be classified as a candidate fault segment.

[0132] Each measurement node acquires three-phase voltage, three-phase current, zero-sequence voltage, and zero-sequence current. Each node includes a total of 8 measurement channels.

[0133] 2) Generate a fault episode.

[0134] Faults were set in each candidate fault section of the IEEE 30-node distribution network, with the fault locations set at 0.3, 0.5, and 0.7 of the total line length, respectively.

[0135] The fault types include 11 types: AG, BG, CG, ABG, ACG, BCG, AB, AC, BC, ABCG, and ABC.

[0136] The fault resistances for ground-related faults are set to 0.01 Ω, 1 Ω, 10 Ω, 100 Ω and 1000 Ω, respectively.

[0137] The fault resistances for phase-to-phase short-circuit related faults are set to 0.01 Ω, 0.1 Ω, 1 Ω, 5 Ω and 20 Ω, respectively.

[0138] The output levels of the distributed power sources are set to 30% and 70%, respectively, and the load levels are set to 80% and 120%, respectively.

[0139] By changing the candidate fault section, fault location, fault type, fault resistance, distributed power supply output level, and load level, fault waveform samples under different operating conditions can be obtained.

[0140] The simulation sampling rate was set to 2 kHz, the fault initiation time was set to 0.40 s, and transient waveform data within the range of 0.35 s to 0.55 s were used as model input. Each waveform sample contained 400 sampling points, and each complete fault waveform sample was considered as a fault episode.

[0141] In the switch reconfiguration test scenario, the switch reconfiguration time is set to 0.50 s, so that the effective topology of the IEEE 30-node distribution network changes within the same fault episode.

[0142] 3) Divide the local time sequence into segments.

[0143] Each fault episode is divided according to the preset patch length and step size.

[0144] The patch length is set to 25 sampling points, and the step size is also set to 25 sampling points. Therefore, each fault episode containing 400 sampling points is divided into 16 patches arranged in chronological order.

[0145] Patch partitioning is performed only within a complete fault episode, and the training, validation, and test sets still use the complete fault episode as the basic partitioning unit.

[0146] 4) Construct two enhanced views.

[0147] During the model training phase, a feature perturbation view and a topological perturbation view are constructed for each patch.

[0148] The feature perturbation view maintains the nominal topology of the IEEE 30-node system and applies random noise, feature masking, or amplitude perturbation to the node measurement features within the patch.

[0149] The topology perturbation view preserves the original node measurement characteristics within the patch, while adding, deleting, or replacing a small number of local edge connections in the nominal topology.

[0150] For the same patch in the same fault episode, the feature perturbation view representation and the topology perturbation view representation are used as positive sample pairs.

[0151] The non-corresponding patch representations within the same fault episode and the patch representations of out-of-type fault samples in the current mini-batch are used as negative samples.

[0152] 5) Perform multi-kernel graph convolutional encoding.

[0153] The feature perturbation view and the topology perturbation view are respectively input into the multi-kernel graph convolution encoder.

[0154] The multi-core graph convolutional encoder in this embodiment uses a 1-hop propagation kernel, a 2-hop propagation kernel, and a PPR diffusion kernel.

[0155] Among them, the 1-hop propagation kernel is used to extract local structural associations between directly adjacent nodes, the 2-hop propagation kernel is used to extract structural associations within a larger neighborhood, and the PPR diffusion kernel is used to extract the diffusion characteristics of fault disturbances within a larger structural range.

[0156] The restart factor for the PPR diffusion core is set to 0.15.

[0157] Features obtained from different propagation kernels are fused and pooled to obtain patch-level contrastive representations of feature perturbation views and topology perturbation views, respectively.

[0158] The topology consistency contrast loss constraint ensures that the representation of the same patch remains consistent in both views, and the contrast learning temperature parameter is set to 0.18.

[0159] 6) Construct a Patch-level edge interaction sequence.

[0160] The original, undisturbed patch and the nominal topology of the IEEE 30-node distribution network are input into a multi-core graph convolutional encoder to obtain the node-level patch representation of each node under the corresponding patch.

[0161] For each undirected physical line in the nominal topology of the IEEE 30-node distribution network, construct edge interaction records in two directions.

[0162] The features of each edge interaction record are obtained by concatenating the source node patch representation, the target node patch representation, and the difference features between the source node patch representation and the target node patch representation.

[0163] Each edge interaction record includes the source node, target node, patch index, physical timestamp, and edge interaction characteristics.

[0164] All edge interaction records are organized according to the Patch order to obtain the Patch-level edge interaction sequence corresponding to each fault episode.

[0165] 7) Perform time-map neural network inference.

[0166] Input the Patch-level edge interaction sequence into a time-graph neural network with a memory mechanism.

[0167] Interaction messages are generated based on the memory states of the source and target nodes before the interaction event occurs, edge interaction characteristics, and time intervals.

[0168] Aggregate messages received by the same node, update the node's memory state based on the aggregated messages, and read out the node embedding.

[0169] In this embodiment, the structural hidden layer dimension of the model is set to 64, and the node memory dimension is set to 64.

[0170] The interaction records of each side are processed sequentially according to the Patch time order, so that the node memory state continuously accumulates historical information of the fault disturbance propagation process along the IEEE 30-node distribution network.

[0171] 8) Classify faulty sections and train the model.

[0172] Aggregate the node embeddings within the post-fault analysis patch window to obtain the global representation corresponding to the fault episode.

[0173] The global representation is input into the Softmax classifier, which outputs the predicted probability of each candidate fault segment.

[0174] During the model training phase, the supervised classification loss is calculated based on the true labels of the fault sections and jointly optimized with the topology consistency comparison loss. The loss balance coefficient is set to 0.24.

[0175] The dataset is divided into training, validation, and test sets according to the fault episode, with a ratio of 6:2:2.

[0176] All node features and edge interaction features are processed using standardized parameters determined solely from training set statistics.

[0177] The training batch size was set to 32 fault episodes, the maximum number of training rounds was set to 200, and an early stopping strategy was adopted based on the validation set loss.

[0178] 9) Perform online fault location.

[0179] During the online localization phase, the three-phase voltage, three-phase current, zero-sequence voltage, and zero-sequence current corresponding to the fault episode under test are obtained, and the fault episode under test is divided into 16 patches using the same patch length and step size as in the training phase.

[0180] During the online localization phase, no feature perturbation view or topology perturbation view is constructed, nor is the topology consistency comparison loss calculated.

[0181] The original, unperturbed patch is input into the trained multi-kernel graph convolutional encoder to obtain a node-level patch representation, and a patch-level edge interaction sequence is constructed based on the nominal topology of the IEEE 30-node distribution network.

[0182] We utilize a time-graph neural network with a memory mechanism for message generation, message aggregation, node memory update, and node embedding readout.

[0183] The node embeddings within the post-fault analysis patch window are aggregated, and the predicted probability of each candidate fault segment is output through the Softmax classifier. The candidate fault segment with the highest probability is taken as the fault segment localization result.

[0184] In the slight topology mismatch test, scenarios were set up where the effective topology differed from the nominal topology by one edge connection and two edge connections.

[0185] In the switch reconfiguration test, the effective topology changes every 0.50 s. The waveform data to be tested is generated from the corresponding effective topology, and the model outputs the fault segment location result according to the online location process.

[0186] 10) Fault section location results.

[0187] In the nominal static topology scenario of the IEEE 30-node distribution network, the fault segment location accuracy of the method of this invention is 90.74%, the macro average F1 value is 91.46%, and the Top-2 accuracy is 95.82%.

[0188] When the effective topology changes by one edge relative to the nominal topology, the fault segment location accuracy of the method of the present invention is 89.42%.

[0189] When the effective topology changes by two edges relative to the nominal topology, the fault segment location accuracy of the method of the present invention is 87.73%.

[0190] In scenarios where switch reconfiguration occurs within a fault episode, the fault segment localization accuracy of the method of this invention is 86.24%, the macro average F1 value is 86.98%, and the Top-2 accuracy is 92.31%.

[0191] As can be seen from the above results, this invention extracts structural features at different propagation scales through a multi-kernel graph convolutional encoder, improves the model's adaptability to local topology changes through topology consistency comparison learning between feature perturbation views and topology perturbation views, and depicts the continuous dynamic process of fault perturbation propagation along the IEEE 30-node distribution network lines through a time-graph neural network with a memory mechanism. It can realize fault section location in nominal static topology, slight topology mismatch and switch reconfiguration scenarios.

Claims

1. A method for locating fault sections in a distribution network based on a comparative topological time-series neural network, characterized in that, Includes the following steps: 1) Obtain time-series measurement data of distribution network nodes, a set of candidate fault sections, and the nominal topology of the distribution network, and construct a nominal topology graph based on the connection relationships between distribution network nodes and lines; 2) Take a complete fault recording sample as a fault episode, and divide the fault episode into multiple local time segments arranged in chronological order according to the preset patch length and step size; 3) Construct a feature perturbation view and a topology perturbation view for each local time series segment. The feature perturbation view keeps the nominal topology unchanged and applies perturbation to the measurement features. The topology perturbation view keeps the measurement features unchanged and applies local edge connection perturbation to the nominal topology. 4) Input the feature perturbation view and the topology perturbation view into the multi-kernel graph convolutional encoder to obtain patch-level contrastive representations under the two enhanced views, and constrain the cross-view representations of the same patch to maintain consistency based on the topology consistency contrastive loss; 5) Input the unperturbed local time sequence into the multi-kernel graph convolutional encoder to obtain the node-level patch representation, and construct the patch-level edge interaction sequence according to the physical lines in the nominal topology; 6) Input the Patch-level edge interaction sequence into a time graph neural network with a memory mechanism to perform interaction message generation, message aggregation, node memory update and node embedding readout to obtain node embeddings that reflect the dynamics of fault propagation; 7) Aggregate the node embeddings within the post-fault analysis Patch window to obtain an episode-level global representation, and output the predicted probability of each candidate fault segment through a classifier. Take the candidate fault segment with the highest predicted probability as the fault segment localization result. 8) During the model training phase, the supervised classification loss is calculated based on the true labels of the faulty sections, and the model parameters are jointly optimized using the supervised classification loss and the topology consistency comparison loss. During the online localization phase, the feature perturbation view and the topology perturbation view are not constructed, and the topology consistency comparison loss is not calculated. Instead, the trained model is used to execute steps 1), 2), and 5) to 7) on the faulty sample unit to be tested to obtain the faulty section localization result.

2. The method for locating fault sections in a distribution network based on a comparative topology time-series neural network according to claim 1, characterized in that, In step 1), 1) The timing measurement data of the distribution network nodes are the three-phase voltage, three-phase current, zero-sequence voltage and zero-sequence current of the distribution network measurement nodes; 2) The set of candidate fault sections is as follows: ,in The number of candidate fault sections, and each candidate fault section Corresponding to one physical line; 3) The nominal topology of the distribution network is ,in For a set of nodes, For the nominal line connection set, This is the corresponding adjacency matrix; 4) The valid topology diagram for the testing or operation phase is denoted as... , For a set of valid line connections; when and If there is an inconsistency, it is determined that there is a topology mismatch or a topology change caused by switch reconfiguration.

3. The method for locating fault sections in a distribution network based on a comparative topology time-series neural network according to claim 1, characterized in that, In step 2), the original spatiotemporal graph tensor of the fault episode is: ,in The number of sampling points. For the number of nodes, The feature dimension of a single node; The number of patches is obtained by dividing the patch into sections based on its length P and step size S. .

4. The method for locating fault sections in a distribution network based on a comparative topology time-series neural network according to claim 1, characterized in that, In step 3), the feature perturbation view is represented as The topology perturbation view is represented as ;in Original node features, This indicates a disturbance in the measurement characteristics. This represents the perturbation caused by adding or deleting edges in the local topology.

5. The method for locating fault sections in a distribution network based on a comparative topology time-series neural network according to claim 1, characterized in that, In step 4), the multi-kernel graph convolutional encoder includes at least two sets of graph convolutional kernels with different propagation orders or different receptive fields, used to extract structural propagation features at different scales on the nominal topology and topological perturbation view; After concatenating, mapping, or pooling the multi-core outputs, a patch-level contrast representation is obtained.

6. The method for locating fault sections in a distribution network based on a comparative topology time-series neural network according to claim 1, characterized in that, In step 4), for the p-th patch of the s-th fault episode, its feature perturbation view representation and topology perturbation view representation are used as positive sample pairs, and the representations of non-corresponding patches in the same fault episode and the patch representations of heterogeneous fault samples in the current mini-batch are used as negative samples. The contrastive loss is used for training.

7. The method for locating fault sections in a distribution network based on a comparative topology time-series neural network according to claim 1, characterized in that, In step 5), for any undirected physical path in the nominal topology... Construct separately and Two-way edge interaction records; the features of the edge interaction records are obtained by concatenating the source node Patch representation, the target node Patch representation, and the difference features between the two. The Patch-level edge interaction sequence is organized in Patch order. Each interaction record in the edge interaction sequence includes a source node, a target node, a Patch index, a physical timestamp, and edge interaction features.

8. The method for locating fault sections in a distribution network based on a comparative topology time-series neural network according to claim 1, characterized in that, In step 6), the time-graph neural network with memory mechanism generates interactive messages based on the memory states of the source node and the target node before the interactive event occurs, the edge interaction features and the time interval, aggregates the messages received by the same node, and updates the node memory state through a gated recurrent unit, a long short-term memory unit or a multilayer perceptron.

9. The method for locating fault sections in a distribution network based on a comparative topology time-series neural network according to claim 1, characterized in that, In step 7), the node embeddings within the post-fault analysis patch window are subjected to average pooling, weighted pooling, or attention pooling to obtain an episode-level global representation. The probability distribution of candidate fault segments is output through a Softmax classifier. .

10. A method for locating fault sections in a distribution network based on a comparative topology time-series neural network according to claim 1, characterized in that, In step 8), the model training phase uses supervised classification loss and topological consistency contrast loss to form a joint objective function; in the online localization phase, the enhanced view is no longer constructed and the contrast loss is no longer calculated, but the trained model is directly used to infer the fault episode to be tested.

Images