A method and system for intelligent localization of traveling wave fault points in distribution networks based on deep temporal feature learning

Abstract

This invention provides an intelligent method and system for locating traveling wave faults in distribution networks based on deep temporal feature learning, belonging to the field of intelligent distribution detection technology based on deep learning. First, high-speed traveling wave sensors are deployed on the distribution lines to collect multi-dimensional three-phase voltage and current signals, and a dataset of fault type, phase, level, and location is constructed. Then, fault state features are extracted through topology-sensing normalization, symmetric component mapping, and dual-channel temporal modeling, and an attention mechanism is used to accurately identify the fault type, involved phase, and level. Further, a fault intelligent location model based on a topology graph is constructed, introducing tower sensing weights and line information carrying weights. Spatiotemporal embeddings are extracted through a graph attention network and an echo state network, thereby achieving line fault classification and precise location. Finally, by combining offline model training with online system deployment, real-time identification and location of distribution network faults are achieved, effectively improving the accuracy, robustness, and response speed of fault diagnosis.

Description

Technical Field

[0001] This invention belongs to the field of intelligent power distribution detection technology based on deep learning, and particularly relates to an intelligent location method and system for traveling wave fault points in power distribution networks based on deep time-series feature learning. Background Technology

[0002] In power systems, the distribution network, as a crucial link connecting the transmission and transformation system with end users, directly impacts the stable supply of electricity and the reliability of power consumption on the user side. With the increasing integration of new energy sources, enhanced load volatility, and the growing prominence of equipment aging, the distribution network faces increasing uncertainties and potential risks during operation, particularly the rising frequency and complexity of sudden faults. Traveling wave fault location, a key technology for fault diagnosis and rapid response in distribution networks, can determine the location and nature of faults in a very short time, providing core support for improving fault handling efficiency, shortening power restoration time, and ensuring system resilience. However, limited by the adaptability of traditional methods to noise interference, changes in signal propagation characteristics, and complex network structures, the accuracy and robustness of current traveling wave fault location methods still fall short of the operational requirements of high-reliability distribution systems. Therefore, there is an urgent need to develop new fault identification and location methods with high accuracy, strong adaptability, and real-time response capabilities to better support the safe operation and management of intelligent distribution systems.

[0003] Currently, the main methods for locating and identifying traveling wave faults in distribution networks include the following:

[0004] (1) Time-domain difference location method based on traveling wave signal: This method calculates the location of the fault point in the distribution line by accurately measuring the time difference of the traveling wave signal arriving at multiple measurement points after the fault occurs. Its principle is simple and intuitive and the hardware implementation threshold is low. However, this method is highly dependent on the accuracy of synchronous sampling and the clarity of the initial traveling wave of the fault. It is easily affected by noise interference at the measurement point, channel delay error and weak fault characteristics, resulting in large location error, especially in complex distributed power grids where the accuracy drops significantly.

[0005] (2) Time-frequency analysis method based on wavelet transform and energy change identification: This method uses wavelet transform to extract abrupt change features from traveling wave signals, uses energy jump points as fault features, and combines them with a physical model for localization. It has good time-frequency resolution at multiple scales and is suitable for processing non-stationary signals. However, this method is sensitive to the selection of wavelet basis, and the feature extraction parameters depend on manual setting; it does not respond significantly to weak disturbances and some high-impedance faults, which can easily lead to misidentification or missed identification; in addition, this method is not stable enough under high-frequency noise and is difficult to adapt to complex scenarios.

[0006] (3) End-to-end identification methods based on deep learning models such as LSTM and CNN: Deep learning methods can learn the nonlinear mapping relationship between traveling waves and fault types / locations through large-scale data. For example, LSTM is used to capture time-series features, and CNN is used to identify local mutations, forming an end-to-end fault identification framework. However, this type of method is highly dependent on large-scale labeled samples, and its generalization ability is limited by the distribution of training data. In addition, the model has poor interpretability and it is difficult to trace the source of key decision features, which limits its reliable deployment in high-safety-level power systems.

[0007] In summary, although current methods for identifying and locating traveling wave faults in power distribution networks have achieved application results in certain scenarios, they generally suffer from problems such as strong reliance on accuracy, poor anti-interference ability, and insufficient adaptability to complex faults. Summary of the Invention

[0008] To address the aforementioned problems, the first aspect of this invention provides an intelligent method for locating traveling wave fault points in distribution networks based on deep temporal feature learning, comprising the following steps:

[0009] S1, high-speed traveling wave sensors are deployed on each monitoring tower of the power distribution line to collect multi-dimensional complete traveling wave signals of three-phase voltage and current;

[0010] S2 inputs the complete traveling wave signal into the trained distribution network traveling wave fault state identification model. First, topology-aware normalization is performed to suppress amplitude drift. Then, symmetric component learnable mapping and structure-time dual masking mechanism are used to extract time-series features. Combining global attention enhancement and multi-task prediction branch, the fault type, involved phase and fault level are output.

[0011] S3 constructs an intelligent fault location model by building a graph structure data of the complete traveling wave signal and the physical connection relationship of the tower. The node feature is the traveling wave signal, and tower perception weight and line information carrying weight are introduced to reflect the node activity and channel strength. The weights are optimized using the fault state identification results output from S2. Based on the graph sequence structure, traveling wave-driven graph perception modeling is performed, and a multi-layer graph attention network extracts structural embedding features that fuse spatiotemporal information. Finally, a two-layer echo state network is used to output the fault classification results and fault location results.

[0012] Preferred methods for constructing the dataset used to train the model include:

[0013] High-speed traveling wave sensors are deployed on each monitoring tower of the power distribution line to collect continuous traveling wave signals before and after the fault occurs, thus obtaining the complete traveling wave signal for the duration of the fault. ,in, Indicates the first The traveling wave signals collected by each monitoring tower during this duration, and ; and the traveling wave signal collected by any monitoring tower ,in, Indicates the first One monitoring tower Traveling wave characteristic data acquired at specific time points, including three-phase voltage , , and three-phase current , , .

[0014] Preferably, in the power distribution network traveling wave fault state identification model, topology-sensing normalization processing is performed to suppress amplitude drift. The specific process is as follows:

[0015] against The Middle One monitoring tower Traveling wave characteristic data acquired at specific time points First, determine the first line based on the line topology connection. The set of neighboring towers of a monitoring tower Calculate the neighborhood tower set in Local mean of traveling wave characteristic data at time points and standard deviation The Z-score normalization method was used to achieve local normalization based on topological neighborhood; topological sensing normalization was performed on all time-point data of all monitoring towers to obtain the processed data. .

[0016] Preferably, in the power distribution network traveling wave fault state identification model, the symmetric component learnable mapping is based on a linear mapping layer, which maps the three-phase signals [A,B,C] to [zero sequence, positive sequence, negative sequence] representations. This mapping layer uses the standard transformation matrix of the symmetric components as the initial weights for the linear transformation, preserves the original phase information through the residual structure, and allows the mapping weights to be optimized during training to obtain structured features. .

[0017] Preferably, in the power distribution network traveling wave fault state identification model, the structure-time dual masking mechanism is introduced to extract time-series features as follows:

[0018] Structural features Robust temporal features are obtained by applying random time-axis masks across time-mapping layers to mask certain time points. By applying a monitoring point mask across the monitoring point mask layer, a scenario with missing nodes is simulated, and structural mask features are obtained. ;after, and State-driven modeling and nonlinear transformation are performed through DSS layers respectively, and layer normalization and ReLU activation are combined to output the results. and ;

[0019] The above structure is repeatedly stacked five times, with each subsequent layer taking the input of the two channels of the previous layer as input, and finally spliced ​​together to form a multi-layer fusion feature. .

[0020] Preferably, the multi-task prediction branch in the power distribution network traveling wave fault state identification model specifically comprises:

[0021] For fusion semantic representation The three task branches are designed to make the following predictions:

[0022] Fault type prediction head: Consists of a fully connected layer, a Dropout layer, and a Softmax classification layer, which yields the fault type identification result. It includes six types: no fault, single-phase ground fault, two-phase short circuit fault, two-phase ground short circuit fault, three-phase short circuit fault, and three-phase ground short circuit fault.

[0023] Phase recognition prediction head: Consists of a fully connected layer, a Dropout layer, and a sigmoid activation function, which yields the phase recognition results. It includes three phases: A, B, and C.

[0024] Fault severity prediction head: Composed of a fully connected layer, a Dropout layer, and a Sigmoid activation function, it obtains the fault severity identification result. It is divided into Level I, Level II, and Level III.

[0025] Preferably, in the intelligent fault location model, a graph structure data is constructed by linking the complete traveling wave signal with the physical connection relationship of the tower, specifically as follows:

[0026] Based on complete traveling wave signal First, a basic diagram structure is constructed based on the physical connection relationships between the monitoring towers in the distribution network; specifically, each monitoring tower is used as a node, and the physical connection lines between the towers are used as edges; The traveling wave characteristics of each tower are used as nodal features to obtain a length of Graph sequence structure data ;

[0027] In addition, two types of key mapping parameters are introduced into the original graph structure:

[0028] Tower sensing weight It is used to measure the response intensity of different tower nodes to traveling wave disturbance signals, and can dynamically reflect the activity of local nodes under different fault types;

[0029] Line information carrying weight Used to measure the strength of graph edges in the propagation path of fault signals, enhancing the characterization of dominant propagation paths and key channels;

[0030] Two weight parameters and During initialization, all nodes and edges are assigned the same value and used as structural variables to be optimized, enabling the graph structure to dynamically adjust topology awareness under different fault modes.

[0031] Preferably, in the fault intelligent localization model, the weights are optimized using the fault state identification results output by S2, specifically as follows:

[0032] Define decision variables; tower sensing weights Dimensions are , To monitor the number of towers, among which Indicates the first The response strength weight of each tower to fault disturbance signals, with a value range of [value range missing]. A higher weight indicates a stronger ability of the tower to detect traveling wave signals under the current fault type; line information carrying weight. Dimensions are , The number of lines between towers, of which Indicates the first The carrying strength weight of each line for fault signal propagation, with a value range of [value range missing]. The higher the weight, the more likely the line is the dominant propagation path of the traveling wave signal under the current fault type.

[0033] Establish an objective function; the objective function aims to maximize the sparsity of fault types, the smoothness of tower sensing weights, and the weight of line information carrying capacity, while taking into account the constraints of fault type adaptability and structural smoothness:

[0034] ;

[0035] ;

[0036] ;

[0037] ;

[0038] in and Fault level identification results output by S2 Confirmed, details are as follows:

[0039] ;

[0040] ;

[0041] and Used to control sparsity constraints, indicating that the higher the fault level, the weaker the sparsity constraints; for Numerical representation of The values ​​1, 2, and 3 correspond to respectively Levels I, II, and III;

[0042] Used to constrain phase alignment under multi-phase faults, when When the node in the middle is a multi-phase fault , When the node in the middle is a single-phase fault ; The fault phase indication vector covers the phase encoding of each tower node, based on... The calculation shows that a fault in one phase is recorded as 1, and vice versa as 0. If the first phase is faulty... If all three phases of a node fail, then ; This is the inverse matrix of the route length, used to reinforce the weight of short-path routes; The weight of fault type complexity on path constraints is defined as follows: the more complex the fault type, the smaller the path constraint. The corresponding values ​​for no fault, single-phase ground fault, two-phase short-circuit fault, two-phase ground short-circuit fault, three-phase short-circuit fault, and three-phase ground short-circuit fault are defined. The values ​​are 0, 0.1, 0.2, 0.3, 0.4, and 0.5 respectively;

[0043] The objective function is solved using a physically constrained enhanced adaptive gradient descent algorithm to obtain the optimal tower sensing weights. and optimal line information carrying weight .

[0044] Preferably, in the fault intelligent localization model, traveling wave-driven graph perception modeling is performed based on graph sequence structure, and a multi-layer graph attention network extracts structural embedding features that fuse spatiotemporal information. The specific process is as follows:

[0045] For the first time point plot data A graph attention network enhancement layer is used to process the data, fusing the local structural relationships between nodes to extract the fault semantic embedding features at the first time point. ;

[0046] Subsequently, this feature With the graph structure input at the next time step The data is concatenated and input into the second layer, GATv2Conv, to achieve cross-time-lapse information transmission and collaborative modeling, and to extract the fault semantic embedding features at the second time point. Process all in sequence A time-point graph structure is generated, forming all fault semantic embedding features that encompass temporal evolution. ;

[0047] The above structure consists of four stacked GATv2Conv layers. Finally, the output results embedded in all time nodes are concatenated to form a complete line diagnostic sequence representation. .

[0048] The second aspect of the present invention provides an intelligent location system for traveling wave faults in a distribution network, characterized in that: it includes high-speed traveling wave sensors deployed on each monitoring tower of the distribution line, and a power grid monitoring center and edge computing nodes, wherein the power grid monitoring center and edge computing nodes are equipped with the traveling wave fault state identification model and the intelligent fault location model of the distribution network as described in the first aspect.

[0049] Compared with the prior art, the present invention has the following beneficial effects:

[0050] (1) Multi-task fault state identification mechanism for complex distribution network scenarios: In view of the problem that traveling wave signals are easily affected by branch reflection and noise interference in complex distribution networks, this invention proposes a feature extraction method that combines topology-aware normalization and symmetric component learnable mapping. By using dual-channel time-series modeling and attention aggregation, the joint identification of fault type, involved phase and fault level is achieved, thereby improving the diagnostic accuracy and stability in multi-interference scenarios.

[0051] (2) Topology graph modeling method with dynamic weights: In view of the differences in propagation path and tower response of different fault types, this invention introduces tower perception weight and line information carrying weight in topology graph modeling, and performs dynamic optimization in combination with fault identification results, so that the graph structure can adaptively reflect the key nodes and dominant paths under fault mode, thereby enhancing the adaptability of the positioning model to complex topology and multiple types of faults.

[0052] (3) Three-graph spatiotemporal joint driving fault location framework: In view of the task requirement of modeling the time evolution law of traveling wave and topological space dependence for fault location in distribution network, this invention uses graph attention network for structural relationship modeling and combines echo state network to characterize time dynamics, forming graph spatiotemporal joint feature representation, thereby achieving high-precision location of line fault while ensuring real-time performance. Attached Figure Description

[0053] Figure 1 This is a flowchart illustrating the overall technical route of the present invention.

[0054] Figure 2 This is a structural diagram of the power distribution network traveling wave fault state identification model of the present invention.

[0055] Figure 3 This is a network structure diagram of the fault intelligent location model based on topology graphs in this invention.

[0056] Figure 4 This is a comparison chart of the performance indicators for fault line identification in embodiments of the present invention.

[0057] Figure 5 This is a heat map for fault location in an embodiment of the present invention.

[0058] Figure 6 This is a heatmap of line information carrying weight in an embodiment of the present invention.

[0059] Figure 7 This is a distribution diagram of the characteristic response values ​​of the tower nodes in an embodiment of the present invention. Detailed Implementation

[0060] This invention proposes an intelligent method and system for locating traveling wave faults in distribution networks based on deep temporal feature learning. The core idea is to first identify the fault type and then locate the fault position to improve diagnostic accuracy and robustness. Specifically, high-speed traveling wave sensors are first deployed on the distribution lines to collect multi-dimensional three-phase voltage and current signals, and a dataset of fault type, phase, level, and location is constructed. Then, fault state features are extracted through topology-sensing normalization, symmetric component mapping, and dual-channel temporal modeling, and an attention mechanism is used to accurately identify the fault type, involved phase, and level. Further, a topology-based intelligent fault location model is constructed, introducing tower sensing weights and line information carrying weights. Spatiotemporal embeddings are extracted through graph attention networks and echo state networks, thereby achieving line fault classification and precise location. Finally, by combining offline model training with online system deployment, real-time identification and location of distribution network faults are achieved, effectively improving the accuracy, robustness, and response speed of fault diagnosis. The overall implementation process is as follows: Figure 1 As shown.

[0061] S1, Distribution Network Traveling Wave Signal Acquisition and Dataset Construction: Multiple monitoring towers are deployed in the distribution lines to collect three-phase voltage and current multi-dimensional traveling wave signals, forming complete traveling wave feature data; continuous signals before and after the fault are obtained through a high-speed sampling device, and a traveling wave fault identification dataset is constructed, labeling the fault type, involved phases, and fault level; on this basis, the fault location is further calibrated, and a traveling wave fault location dataset is constructed to provide structured data support for model training.

[0062] S2, Construction of a Fault State Identification Model for Distribution Networks: Addressing the issue of traveling wave signals being affected by reflection and interference, this invention designs a fault state identification module. First, topology-sensing normalization is performed to suppress amplitude drift. Then, a symmetric component learnable mapping and a structure-time dual-mask mechanism are used to extract temporal features. Combining global attention enhancement and multi-task prediction branches, accurate identification of fault type, involved phase, and fault level is achieved, providing stable input for subsequent localization.

[0063] S3, Construction of a fault intelligent location model based on topology graph: Constructing graph structure data by integrating complete traveling wave signals with the physical connection relationship of towers, with node features being traveling wave signals, and introducing tower sensing weights and line information carrying weights to reflect node activity and channel strength; optimizing weights using fault state identification results; performing traveling wave-driven graph perception modeling based on graph sequence structure, and extracting structural embedding features that fuse spatiotemporal information using a multi-layer graph attention network; finally, using a two-layer echo state network for fault classification and location, achieving accurate identification of faults in the line.

[0064] S4, Model Training and System Deployment: Based on the dataset constructed in S1 and the models designed in S2 and S3, offline training is performed to optimize the prediction errors of fault type, involved phase, fault level, and fault location. After training, the fault state identification model and intelligent positioning model are modularly deployed to the monitoring center and edge nodes, and traveling wave signals are collected in real time through high-speed sensors. In the online stage, fault state identification is performed first, and then the physical constraint enhanced adaptive gradient descent algorithm is used to update the tower sensing weight and line information carrying weight in real time. Finally, the fault intelligent positioning model is input to output the specific location of the fault in the line, realizing fast and accurate online fault location.

[0065] The invention will be further described below with reference to specific embodiments.

[0066] I. Acquisition of Traveling Wave Signals and Dataset Construction in Distribution Networks

[0067] In engineering practice, different types of faults (such as single-phase grounding, two-phase short circuit, and three-phase short circuit) differ significantly in signal propagation path, disturbance amplitude, and duration characteristics. If a unified model is used, it can easily lead to feature mixing and ambiguous identification, thereby reducing location accuracy. To improve the overall performance of distribution network fault diagnosis, this invention adopts a phased modeling strategy of "first identifying the fault type, then locating the fault location." By constructing independent fault type identification and fault location datasets, the model training objectives are clear, and feature extraction is stable, thus effectively enhancing the accuracy and robustness of fault location. Therefore, it is necessary to first construct datasets containing two types of tasks: a traveling wave fault identification dataset and a traveling wave fault location dataset, specifically including the following steps:

[0068] Distribution network traveling wave signal feature selection: Multiple monitoring towers are pre-selected in the distribution line, and multi-dimensional features including voltage and current are selected as input signals. Specifically, the following six channels of traveling wave signals are collected at each monitoring point: three-phase voltage... , , and three-phase current , , Therefore, the complete traveling wave characteristic data of the distribution network traveling wave signal is: .

[0069] Traveling wave input data acquisition in power distribution network: High-speed traveling wave sampling devices are deployed on each monitoring tower of the power distribution line, and continuous traveling wave signals before and after the fault occur are acquired at high frequency to obtain the complete traveling wave signal during the duration of the fault. ,in, Indicates the first The traveling wave signals collected by each monitoring tower during this duration, and ; and the traveling wave signal collected by any monitoring tower ,in, Indicates the first One monitoring tower Traveling wave characteristic data (including three phase voltages and phase currents) collected at specific time points.

[0070] Construction of Traveling Wave Fault Identification Dataset: Based on the collected traveling wave input data of the distribution network, the fault type labels corresponding to the input data are labeled. (Including: no fault, single-phase ground fault, two-phase short-circuit fault, two-phase ground short-circuit fault, three-phase short-circuit fault, and three-phase ground short-circuit fault, a total of six types); The phases involved in the fault are also indicated. (Including three phases A, B, and C), and indicating the severity level of the fault. (According to GB / T 30137-2013, it is divided into three levels: Level I, Level II, and Level III, with Level I corresponding to the highest severity); therefore, traveling wave fault identification labels... ;

[0071] Furthermore, with complete traveling wave signals For input, traveling wave fault identification tag The output is a set of traveling wave fault identification datasets. Based on this method, extensive data collection is carried out to obtain a complete traveling wave fault identification dataset.

[0072] Construction of Traveling Wave Fault Location Dataset: After constructing the dataset for the fault identification task, further labeling of fault location information is performed; specifically, fault location labels are added to the lines between monitoring towers. and If a fault occurs in the wiring between the detection towers, then... Conversely, it is 0; This indicates the specific location of the fault in the line, with the starting point defined by the direction of the current. This indicates a fault occurred at the starting tower. This indicates a fault occurred at the end tower. If no fault occurs, then It is set to the preset maximum value;

[0073] Therefore, based on this method, fault locations are marked for all lines, and the final traveling wave fault location label is: ;

[0074] With complete traveling wave signal For input, traveling wave fault location tag The output is a set of traveling wave fault location datasets. Based on this method, extensive data collection is carried out to obtain a complete traveling wave fault location dataset.

[0075] II. Design of Traveling Wave Fault State Identification Model for Distribution Network

[0076] Because traveling wave signals in complex power distribution networks are often affected by branch reflections and environmental interference, direct fault location inference can easily lead to model location deviations and convergence instability. Therefore, this invention proposes introducing a traveling wave fault state identification module before fault location. By pre-identifying the global disturbance patterns of the traveling wave, a clearer input representation and a more stable optimization boundary are provided for subsequent location, ensuring that the overall method possesses the advantages of phased collaborative modeling.

[0077] The input data for this model is a complete traveling wave signal. The output includes fault type identification results, phase identification results, and fault level identification results. , structure as Figure 2 As shown.

[0078] Topology-sensing normalization processing: To suppress amplitude drift caused by load differences at different measurement points, this invention proposes a topology-sensing normalization method for traveling wave signals. Normalization is performed to suppress modeling bias caused by differences in amplitude at different measurement points;

[0079] Specifically, targeting The Middle One monitoring tower Traveling wave characteristic data acquired at specific time points The following processing is performed on it: First, determine the first line based on the line topology connection. The set of neighboring towers of a monitoring tower Calculate the neighborhood tower set in Local mean of traveling wave characteristic data at time points and standard deviation And the Z-score normalization method is used to achieve local normalization based on topological neighborhood;

[0080] Based on this method, topology-sensing normalization is performed on all time-point data from all monitoring towers to obtain the processed data. .

[0081] Symmetric component learnable mapping: The symmetric component mapping layer, based on a linear mapping layer, maps the three-phase signal [A,B,C] to [zero-sequence, positive-sequence, negative-sequence] representation, thereby enhancing the structural decoupling of the electrophysical semantics. Furthermore, this mapping layer uses the standard transformation matrix of the symmetric components as the initial weights for the linear transformation, preserves the original phase information through the residual structure, and allows the mapping weights to be optimized during training to obtain structured features. .

[0082] Dual-channel temporal modeling based on DSS: In order to improve the robustness of the model to fault signals, this invention introduces a structure-time dual masking mechanism and integrates a deep separable state space (DSS) layer to extract temporal evolution features.

[0083] Specifically, structural features Robust temporal features are obtained by applying random time-axis masks across time-mapping layers to mask certain time points. By applying a monitoring point mask across the monitoring point mask layer, a scenario with missing nodes is simulated, and structural mask features are obtained. ;after, and State-driven modeling and nonlinear transformation are performed through DSS layers respectively, and layer normalization and ReLU activation are combined to output the results. and ;

[0084] This structure is repeatedly stacked five times, with each subsequent layer taking the input of two channels from the previous layer as input, and finally splicing them together to form a multi-layer fusion feature. It has both context-aware and structural consistency modeling capabilities.

[0085] Fusion of global perception and attention enhancement: The semantic features of the entire image are obtained through global average pooling and then fed into the fault perception attention aggregation layer. This layer dynamically assigns weights to key tower nodes through a multi-head attention mechanism, thereby strengthening the fault impact path modeling and outputting features. .

[0086] Subsequently The final fused semantic representation is generated by performing two fully connected layers and ReLU activation dimensionality reduction. .

[0087] Multi-task prediction output: for fused semantic representation The three task branches are designed to make the following predictions:

[0088] (1) Fault type prediction head: It consists of a fully connected layer, a Dropout layer and a Softmax classification layer, which obtains the fault type identification result. ;

[0089] (2) Phase recognition prediction head: It consists of a fully connected layer, a Dropout layer and a Sigmoid activation function, and obtains the phase recognition results. ;

[0090] (3) Fault severity level prediction head: It consists of a fully connected layer, a Dropout layer and a Sigmoid activation function, and obtains the fault level identification result. .

[0091] III. Design of a Fault Intelligent Location Model Based on Topology Graph

[0092] To address the uncertainty of traveling wave signal propagation paths in complex distribution networks and the impact of dynamic structural changes on accurate fault location, this invention designs a dynamic graph-driven intelligent fault location model. This model fully integrates distribution network topology connections, traveling wave disturbance characteristics, and fault type identification results to achieve accurate identification of fault locations and faulty lines. The network structure of the topology-based intelligent fault location model is as follows: Figure 3 As shown.

[0093] 1. Construction of graph structure data based on topological priors: based on complete traveling wave signals First, a basic diagram structure is constructed based on the physical connection relationships between the monitoring towers in the distribution network; specifically, each monitoring tower is used as a node, and the physical connection lines between the towers are used as edges; the traveling wave signal is then... The traveling wave characteristics of each tower are used as nodal features to obtain a length of Graph sequence structure data ;

[0094] Furthermore, considering the differences in tower response intensity and path dependence during the propagation of different fault types, this invention introduces two types of key mapping parameters into the original graph structure:

[0095] Tower sensing weight It is used to measure the response intensity of different tower nodes to traveling wave disturbance signals, and can dynamically reflect the activity of local nodes under different fault types;

[0096] Line information carrying weight Used to measure the strength of graph edges in the propagation path of fault signals, enhancing the characterization of dominant propagation paths and key channels;

[0097] The above two weight parameters and During initialization, all nodes and edges are assigned the same value and used as structural variables to be optimized. This allows the graph structure to dynamically adjust its topology awareness under different fault modes, thereby enhancing the model's adaptability to complex spatiotemporal disturbances.

[0098] 2. Optimization of tower sensing weight and line information carrying weight: To improve the sensing capability of the graph structure under different fault modes, reduce fault location errors, and enhance the model's adaptability to complex fault scenarios, the fault type identification results output by the distribution network traveling wave fault state identification model are utilized. Phase recognition results and fault level identification results Tower sensing weights for dynamic topology modeling Line information carrying weight The optimizations are as follows:

[0099] Define decision variables: tower perception weights Dimensions are , To monitor the number of towers, among which Indicates the first The response strength weight of each tower to fault disturbance signals, with a value range of [value range missing]. A higher weight indicates a stronger ability of the tower to detect traveling wave signals under the current fault type. Line information carrying weight. Dimensions are , The number of lines between towers, of which Indicates the first The carrying strength weight of each line for fault signal propagation, with a value range of [value range missing]. A higher weight indicates that the line is the dominant propagation path for traveling wave signals under the current fault type.

[0100] The objective function aims to maximize the sparsity of fault types, the smoothness of tower sensing weights, and the weight of line information carrying capacity. Considering both fault type adaptability and structural smoothness constraints, the expression is as follows:

[0101] ;

[0102] ;

[0103] ;

[0104] ;

[0105] in and Fault level identification results output by S2 Confirmed, details are as follows:

[0106] ;

[0107] ;

[0108] and Used to control sparsity constraints, indicating that the higher the fault level, the weaker the sparsity constraints; for Numerical representation of The values ​​1, 2, and 3 correspond to respectively Levels I, II, and III;

[0109] Used to constrain phase alignment under multi-phase faults, when When the node in the middle is a multi-phase fault , When the node in the middle is a single-phase fault ; The fault phase indication vector covers the phase encoding of each tower node, based on... The calculation shows that a fault in one phase is recorded as 1, and vice versa as 0. If the first phase is faulty... If all three phases of a node fail, then ; This is the inverse matrix of the route length, used to reinforce the weight of short-path routes; The weight of fault type complexity on path constraints is defined as follows: the more complex the fault type, the smaller the path constraint. The corresponding values ​​for no fault, single-phase ground fault, two-phase short-circuit fault, two-phase ground short-circuit fault, three-phase short-circuit fault, and three-phase ground short-circuit fault are defined. The values ​​are 0, 0.1, 0.2, 0.3, 0.4, and 0.5 respectively;

[0110] A physical constraint-enhanced adaptive gradient descent algorithm is designed to obtain the optimal tower sensing weights. and optimal line information carrying weight The details are as follows:

[0111] 1) Initialize tower sensing weights for As the current sensing weight of the tower Each dimension in the equation is a random number between 0 and 1, used to initialize the weights of the line information. for As the current line information carrying weight Each dimension in the expression is a random number between 0 and 1, and a convergence threshold is set. ;

[0112] 2) Fault type identification results output by the power grid traveling wave fault state identification model Phase recognition results and fault level identification results calculate , , , and ;

[0113] 3) Calculate using the current tower sensing weight, the current line information carrying weight, and the calculated parameters. , and Then calculate the current objective function value. ;

[0114] 4) Perturb the values ​​of the current tower sensing weight and the current line information carrying weight, and randomly mutate some values ​​of the tower sensing weight and the current line information carrying weight to obtain the mutated tower sensing weight and the mutated line information carrying weight.

[0115] 5) Interchange the values ​​of random points in the sensing weight of sudden towers and the information carrying weight of sudden lines to obtain the sensing weight of cross towers and the information carrying weight of cross lines.

[0116] 6) Input the sensing weight of the cross tower and the carrying weight of the cross line information, and calculate the objective function value at the next moment. ;

[0117] 7) If Then the output of the cross tower sensing weight and the cross line information carrying weight are used as the optimal tower sensing weight and the optimal line information carrying weight, respectively; otherwise, repeat steps 3) to 7).

[0118] 8) Utilize the optimal tower sensing weight and optimal line information carrying weight configuration diagram structure.

[0119] 3. Traveling wave driven graph perception modeling: based on graph sequence structure data Perform traveling wave-driven graph perception modeling to extract structural embedding features that fuse spatiotemporal information;

[0120] Specifically, regarding the plot data at the first time point... The graph attention network enhancement layer (GATv2Conv) is used for processing to fuse the local structural relationships between nodes and extract the fault semantic embedding features at the first time point. ;

[0121] Subsequently, this feature With the graph structure input at the next time step The data is concatenated and input into the second layer, GATv2Conv, to achieve cross-time-lapse information transmission and collaborative modeling, and to extract the fault semantic embedding features at the second time point. Based on this method, all are processed sequentially. A time-point graph structure is generated, forming all fault semantic embedding features that encompass temporal evolution. ;

[0122] To enhance the model's high-order structure awareness capability, the structure consists of four stacked GATv2Conv layers. Finally, the output results embedded from all time points are concatenated to form a complete line diagnostic sequence representation. This is used for reasoning in subsequent tasks.

[0123] 4. Line Fault Classification and Location: Based on Line Diagnostic Sequence Characteristics This invention utilizes a two-layer echo state network (ESN) to process the sequence features, thereby modeling the dynamic law of fault signal evolution over time and obtaining fault time-series processing features. Then, fault diagnosis and location are performed using the following two task heads:

[0124] (1) Fault timing processing characteristics The Softmax classification layer is used to identify the fault type of the line between towers, and the line fault classification results are obtained. ;

[0125] (2) Fault timing processing characteristics By calculating the fault probability of each line using the ReLU activation function, precise spatial location can be achieved, yielding the fault location result. ;

[0126] Therefore, the topology-based intelligent fault location model combines the actual topology of the power system with the dynamic traveling wave characteristics, and ultimately improves the accuracy of fault location for different fault modes by identifying whether a specific tower line has a fault and the location of the fault.

[0127] IV. Model Training and System Deployment

[0128] This invention involves model training and systematic deployment, achieving a closed-loop process from offline training to online inference, thereby ensuring the availability and stability of the method in actual power distribution networks. Specifically, it includes the following steps:

[0129] Training of the traveling wave fault state identification model for distribution networks: Based on the constructed traveling wave fault identification dataset and the constructed traveling wave fault state identification module for distribution networks, a joint loss function is used to simultaneously optimize the three tasks of fault type, involved phase and fault level; training is terminated when the loss function converges, and the final traveling wave fault state identification model for distribution networks is obtained.

[0130] Training of the intelligent fault location model based on topology graph: Based on the constructed traveling wave fault location dataset and the constructed intelligent fault location model based on topology graph, the structural parameters of the model, such as the graph attention network layer and the echo state network layer, are trained to minimize the line fault classification error and fault location deviation. The training is terminated when the loss function converges, and the final intelligent fault location model is obtained.

[0131] Model system deployment: The trained distribution network traveling wave fault state identification model and fault intelligent location model are deployed in a modular manner to the power grid monitoring center and edge computing nodes, and high-speed traveling wave sensors are deployed on each monitoring tower of the distribution line to realize the real-time acquisition of complete traveling wave signal data.

[0132] Online fault state identification: The collected complete traveling wave signal data is input into the trained distribution network traveling wave fault state identification model to obtain fault type identification results, involved phase identification results and fault level identification results, thereby completing the real-time diagnosis of the current fault state of the distribution network.

[0133] Online weight optimization: Based on the obtained fault type identification results, phase identification results, and fault level identification results, the physical constraint-enhanced adaptive gradient descent algorithm designed by the fault intelligent location model is used to update the tower sensing weight and line information carrying weight in the fault intelligent location model in real time, so as to obtain the optimal weight configuration under this working condition.

[0134] Fault location: After completing the weight update, the graph structure data is constructed by combining the complete traveling wave signal data and input into the trained intelligent fault location model. Finally, the specific location of the current fault in the line is output, so as to realize accurate location and rapid response to the fault in the distribution network.

[0135] V. Experimental Results and Analysis

[0136] To verify the effectiveness of the proposed method for accurate location of traveling wave faults in distribution networks, this invention constructs a multi-type fault sample set based on a typical distribution network simulation scenario, covering different topologies, disturbance types, and fault levels. The experimental objectives include: (1) evaluating the predictive performance of the model in fault line identification and fault location; and (2) verifying the applicability and modeling advantages of this method in dynamic graph structure modeling and topology-aware feature extraction, highlighting its robustness and accuracy under complex spatiotemporal disturbance conditions.

[0137] 1. Fault classification and location prediction effect

[0138] First, the line identification and location positioning capabilities of the model of the present invention under different fault scenarios were compared and analyzed. Figure 4 The comparison of metrics such as accuracy, macro-F1 score, and recall for the fault line identification task is presented. Model A is a graph convolutional neural network model, and model B is an LSTM model. Figure 5 The thermal distribution map shows the location of the corresponding fault.

[0139] First, the performance of the proposed model in line identification and location positioning under different fault scenarios was compared and analyzed. Figure 4 The results show the comparison of three key performance indicators in the fault line identification task, including accuracy, macro-F1 score and recall. Model A is a structure-aware model based on graph convolutional neural network (GCN), and Model B is a typical time-series model based on long short-term memory network (LSTM). Figure 5 This presents a heat map of fault location in typical fault scenarios, which can intuitively reflect the differences in the spatial recognition accuracy of different models for fault points.

[0140] from Figure 4 As can be seen, the method of the present invention outperforms the comparative model in all evaluation metrics, with an accuracy of up to 97.3% and an F1 score of over 96%. This indicates that the collaborative architecture of "topology enhancement mapping + GATv2Conv graph modeling + echo state network (ESN) time series processing" constructed in the present invention can effectively extract key disturbance path features and accurately map them to fault types. Figure 5 The model further demonstrates the thermal distribution of fault location on the physical topology. It effectively focuses on the faulty tower and its adjacent nodes, showcasing the advantages of "line diagnostic sequence characteristics" in fine-grained localization. Overall results indicate that this model can achieve high-precision identification of fault types and spatial locations in distribution networks under complex disturbance propagation conditions.

[0141] 2. Interpretability Analysis of Graph Modeling Structure and Node Features

[0142] To analyze the contribution of key parameters to model performance and physical semantics during graph structure modeling, this paper further examines the learning process and spatial distribution of two types of structural variables: tower sensing weight and line information carrying weight. Figure 6 The heatmap of the edge weight space obtained after training is shown; Figure 7 The characteristic response values ​​of different tower nodes are distributed.

[0143] Depend on Figure 6 It is evident that, under different combinations of fault types and paths, the model can dynamically adjust the edge weight distribution, concentrating the disturbance intensity on the actual propagation path, thus improving the physical consistency and structural flexibility of graph modeling. Figure 7 The data shows that nodes with high perception weights are concentrated in multi-path convergence areas and topological boundaries. This indicates that "tower perception weights" can effectively measure the response capability of nodes to fault disturbances and are a key mediating feature for achieving interpretable fault location.

[0144] Experimental results show that the fault intelligent location method proposed in this invention achieves a unified processing path for fault type identification and spatial location inference through topology enhancement, weight optimization perception, and multi-channel graph time-series modeling. It has both high-precision prediction and structural interpretation capabilities, providing a solid algorithmic foundation and technical support for rapid fault diagnosis and fine control of intelligent power distribution systems.

[0145] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0146] While the specific embodiments of the present invention have been described above, they are not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A distribution network traveling wave fault point intelligent positioning method based on deep timing feature learning, characterized in that, The method comprises the following steps: S1, high-speed traveling wave sensors are arranged on each monitoring tower of a power distribution line to collect multi-dimensional complete traveling wave signals of three-phase voltage and current; S2, the complete traveling wave signals are input into a trained power distribution network traveling wave fault state identification model, first, topological perception normalization processing is performed to suppress amplitude drift, then symmetric component learnable mapping and structure-time double masking mechanisms are used to extract time sequence features; combined with global attention enhancement and multi-task prediction branches, the fault type, involved phase and fault level are output; S3, a fault intelligent positioning model is constructed, graph structure data is constructed from complete traveling wave signals and tower physical connection relationships, node features are traveling wave signals, tower perception weights and line information bearing weights are introduced to reflect node activity and channel strength; the fault state identification results output by S2 are used to optimize the weights; based on graph sequence structure, traveling wave driven graph perception modeling is performed, and multi-layer graph attention networks are used to extract and fuse structure embedding features of time and space information; finally, a double-layer echo state network is used to output fault classification results and fault positioning results.

2. The method of claim 1, wherein the method is based on deep time-series feature learning for intelligent locating of a fault point of a distribution network. The data set for training the model is constructed in the following manner: High-speed traveling wave sensors are arranged at each monitoring tower of the power distribution line to collect continuous traveling wave signals before and after the fault occurs, and complete traveling wave signals of the fault occurrence duration period are obtained wherein, represents the traveling wave signals collected by the i-th monitoring tower in the duration period, and ; and the traveling wave signals collected by any monitoring tower ; and the traveling wave signals collected by any monitoring tower wherein, represents the traveling wave feature data collected by the i-th monitoring tower at the time point t, including three-phase voltages , , , and three-phase currents , , , , .

3. The power distribution network traveling wave fault point intelligent positioning method based on deep time sequence feature learning according to claim 1, characterized in that: in the power distribution network traveling wave fault state identification model, topological perception normalization processing is performed to suppress amplitude drift, and the specific process is as follows: against The Middle One monitoring tower Traveling wave characteristic data acquired at specific time points First, determine the first line based on the line topology connection. The set of neighboring towers of a monitoring tower Calculate the neighborhood tower set in Local mean of traveling wave characteristic data at time points and standard deviation The Z-score normalization method was used to achieve local normalization based on topological neighborhood; topological sensing normalization was performed on all time-point data of all monitoring towers to obtain the processed data. .

4. The method of claim 1, wherein the method is based on deep time-series feature learning for intelligent locating of a fault point of a distribution network. The symmetrical component learnable mapping in the power distribution network traveling wave fault state identification model is based on a linear mapping layer, maps the three-phase signals [A, B, C] into [zero sequence, positive sequence, negative sequence] expressions, and the mapping layer takes the symmetrical component standard transformation matrix as the initial weight of linear transformation, retains the original phase information through a residual structure, and allows the mapping weight to be optimized in training to obtain structured features .

5. The method of claim 4, wherein the method is based on deep time-series feature learning for intelligent locating of a fault point of a distribution network traveling wave. in the power distribution network traveling wave fault state identification model, the structure-time double masking mechanism is introduced to extract time sequence features, and the specific process is as follows: structured features robust temporal features are obtained by applying time-axis random masks through time-mask layers structured mask features are obtained by applying monitor-point masks through monitor-point-mask layers to simulate node missing scenarios ; then, and state-driven modeling and nonlinear transformation are performed through DSS layers, combined with layer normalization and ReLU activation, to respectively output and ; The above structure is repeated and stacked for five layers, and each layer of the rear layer is inputted with two-channel output results of the upper layer, so as to finally splice to form multi-layer fusion features .

6. The method of claim 1, wherein the method is based on deep time-series feature learning for intelligent locating of a fault point of a distribution network traveling wave. the multi-task prediction branch in the power distribution network traveling wave fault state identification model is as follows: For fusing semantic representations , three task branches are designed to make the following predictions: Fault type prediction head: composed of a fully connected layer, a Dropout layer and a Softmax classification layer, to obtain fault type identification results , including no fault, single-phase ground fault, two-phase short circuit fault, two-phase ground short circuit fault, three-phase short circuit fault and three-phase ground short circuit fault, a total of six; Phase identification prediction head: composed of a fully connected layer, a Dropout layer, and a Sigmoid activation function, to obtain a phase identification result , including three phases A, B, and C; The fault severity grade prediction head is composed of a full connection layer, a Dropout layer and a Sigmoid activation function, and obtains a fault grade identification result , and is divided into grade I, grade II and grade III.

7. The method of claim 1, wherein the method is based on deep time-series feature learning for intelligent locating of a fault point of a distribution network traveling wave. in the fault intelligent positioning model, graph structure data is constructed from complete traveling wave signals and tower physical connection relationships, and the specific process is as follows: Based on complete traveling wave signals , first, a basic graph structure is constructed according to the physical connection relationship between each monitoring tower in the power distribution network; specifically, each monitoring tower is taken as a node, and the physical connection line between the towers is taken as an edge; the traveling wave characteristics of each tower in the are taken as node characteristics to obtain a graph sequence structure data with a length of ; In addition, two types of key mapping parameters are introduced in the original graph structure: Tower perception weight : for measuring the response strength of different tower nodes to the traveling wave disturbance signal, dynamically reflecting the local node activity under the difference of fault types; Line information carrying weight : for measuring the strength of the graph edge pairs in the fault signal propagation path, strengthening the characterization ability of the dominant propagation path and the critical path Two weight parameters And Assign the same value to all nodes and edges at initialization, and as a structure variable to be optimized, so that the graph structure can dynamically adjust the topology awareness under different failure modes.

8. The method of claim 7, wherein the method is based on deep time-series feature learning for intelligent locating of a fault point of a distribution network traveling wave. in the fault intelligent positioning model, the fault state identification results output by S2 are used to optimize the weights, and the specific process is as follows: Define decision variables; tower perception weight : dimension is , The number of towers is monitored, wherein represents the response strength weight of the th tower to the fault disturbance signal, and the value range is ; The higher the weight, the stronger the sensing ability of the tower to the traveling wave signal under the current fault type; the line information bearing weight : the dimension is , the number of lines between towers, wherein represents the bearing strength weight of the th line to the fault signal propagation, and the value range is The higher the weight, the stronger the sensing ability of the tower to the traveling wave signal under the current fault type; the line information bearing weight a target function is established, the target function aims to maximize the fault type sparsity, the tower perception weight smoothness and the line information bearing weight, and comprehensively considers the fault type adaptability and the structure smoothness constraint; The target function is solved based on a physical constraint enhanced adaptive gradient descent algorithm to obtain optimal tower perception weights and optimal line information bearing weights .

9. The method of claim 8, wherein the method is based on deep time-series feature learning for intelligent locating of a fault point of a distribution network traveling wave. in the fault intelligent positioning model, based on the graph sequence structure, traveling wave driven graph perception modeling is performed, and multi-layer graph attention networks are used to extract and fuse structure embedding features of time and space information, and the specific process is as follows: For the first time point graph data , a graph attention network enhancement layer is used for processing, local structure relationship between nodes is fused, and fault semantic embedding features of the first time point are extracted ; Subsequently, the feature is concatenated with the graph structure input of the next time point , input into the second layer GATv2Conv, to realize cross-time information transmission and linkage modeling, and to extract the fault semantic embedding features of the second time point ; all graph structures of time points are processed in turn to form all fault semantic embedding features containing time evolution ; The above structure co-stacks 4 layers of GATv2Conv, and finally the output results of all time node embeddings are spliced to form a complete line diagnosis sequence representation .

10. A distribution network traveling wave fault point intelligent positioning system based on deep timing feature learning, characterized in that: high-speed traveling wave sensors arranged on each monitoring tower of a power distribution line and a power grid monitoring center and an edge computing node, the power grid monitoring center and the edge computing node are deployed with the power distribution network traveling wave fault state identification model and the fault intelligent positioning model according to any one of claims 1 to 9.

Images