Power distribution network fault positioning method and device fusing mechanism characteristics and deep learning network model

Abstract

The invention provides a power distribution network fault positioning method and device fusing mechanism characteristics and a deep learning network model, and the method comprises the steps: S01, collecting fault voltage traveling wave signals extracted by each detection device after a fault occurs, carrying out the recognition of an initial traveling wave head, and calculating a time residual error and a fault position; s02, constructing a sample library data set containing traveling wave signals, fault positions and time residual errors; step S03, constructing a mechanism-embedded deep learning network model, and embedding a fault position and a time residual error driven by knowledge into the model as mechanism features; training the model by using the sample library data set; and step S04, acquiring a fault voltage traveling wave signal after a real-time fault occurs, calculating a fault position and a time residual error, and inputting the fault position and the time residual error into the trained model to obtain a fault positioning result. According to the method, the positioning precision and reliability can be improved, the time synchronization error is reduced, and the generalization risk under the complex working condition is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of distribution network fault location, and in particular to a distribution network fault location method and device that integrates mechanism characteristics and a deep learning network model. Background Art

[0002] As a critical infrastructure of the power system, the distribution network undertakes the important task of supplying power to end users. However, the complex mesh topology, diverse factors affecting power supply security, and high operational failure rate of the distribution network make it a key and difficult link in power system reliability management.

[0003] In the field of distribution network fault location, there are two main technical approaches: knowledge-driven methods based on theoretical analysis and data-driven methods based on data mining. The knowledge-driven method achieves location by theoretically analyzing the differences in fault characteristic responses under different fault scenarios and establishing quantitative judgment criteria. Its typical implementation forms include impedance method, signal injection method, and traveling wave method. Among them, the traveling wave method has outstanding advantages such as rapid response, insensitivity to changes in grid topology, and minimal impact from distributed energy. This type of method can achieve meter-level positioning accuracy under the condition of accurate wave head calibration, but it is significantly sensitive to time synchronization errors. For example, under normal operating conditions, the time synchronization error can be controlled within 1μs, but in complex electromagnetic environments, it may produce deviations of several microseconds. When the time synchronization error exceeds a certain threshold, the positioning accuracy will drop significantly.

[0004] Data-driven methods, centered around machine learning, possess powerful pattern recognition capabilities and can automatically construct nonlinear correlation models between fault characteristics and fault locations from multidimensional data, providing a new approach to resolving time synchronization error issues. Existing data-driven methods mostly fuse multi-source, heterogeneous, wide-area measurement information to construct robust feature mapping relationships, demonstrating strong adaptability in scenarios involving data loss, noise interference, and weak faults. However, existing data-driven models generally face interpretability bottlenecks, and their decision-making logic lacks a clear physical explanation. When there is a distributional shift between the test samples and the training set, anomalous outputs that violate the physical laws of the power system may occur, posing a certain generalization risk. Summary of the Invention

[0005] The technical problem to be solved by the present invention is: in response to the technical problems existing in the prior art, the present invention provides a distribution network fault location method and device that integrates mechanism characteristics and deep learning network models. The positioning accuracy and reliability are high, the time synchronization error is small, and the generalization risk under complex working conditions can be reduced.

[0006] In order to solve the above technical problems, the technical solution proposed by the present invention is: A distribution network fault location method integrating mechanism characteristics and deep learning network model includes the following steps: Step S01: After a fault occurs, fault voltage traveling wave signals extracted by multiple detection devices installed on each branch of the distribution network are collected, and the initial voltage traveling wave head of the fault voltage traveling wave signal is identified; based on the arrival time of the initial voltage traveling wave head of the fault voltage traveling wave, the fault occurrence time, and the shortest distance from the fault point to the location of each detection device, the time residual of each detection device is calculated, and the fault location is determined based on the time residual of each detection device under knowledge-driven conditions; Step S02: constructing a sample library data set including the traveling wave signals of each detection device within a specified time period after the wave head arrives, the fault location driven by the knowledge, and the corresponding time residual; Step S03: Constructing a deep learning network model with embedded mechanism, in which the fault location and corresponding time residual driven by the knowledge are embedded as mechanism features into the traveling wave spatiotemporal features extracted by the deep learning network model to form a fusion feature embedded in the mechanism, and the fusion feature is mapped to the fault branch and fault location through a fully connected layer; using the sample library data set to train the deep learning network model with embedded mechanism, to obtain a trained deep learning network model with embedded mechanism; Step S04: Obtain the fault voltage traveling wave signal after the real-time fault occurs, identify the initial voltage traveling wave head, calculate the knowledge-driven fault location and the corresponding time residual, input the trained deep learning network model embedded in the mechanism, and obtain the positioning result output of the fault point location.

[0007] Furthermore, before step S01, a simulation model with the same line parameters and topology as the distribution network under test is built. In the simulation model, multiple fault points are set in each branch, and multiple fault conditions are set for each fault point. After the fault voltage traveling wave signal extracted by each detection device is collected in step S01, the fault voltage traveling wave signal is detected according to the change of the traveling wave head. , traveling wave head rise time t 1 and the amplitude of the traveling wave head The initial voltage traveling wave head of each fault voltage traveling wave signal is identified, and the corresponding wave head arrival time is obtained.

[0008] Furthermore, the calculation expression of the time residual of each detection device is:

[0009] in, For the i The time residual of each detection device, is the time when the fault occurred, Fault point FTo the current detection device i The shortest distance to the location, is the wave speed, Indicates the i The wave head arrival time of each detection device; Calculate the overall time residual of all detection devices: ,in n Indicates the number of detection devices, H Represents the overall time residual of the system; Solution The fault location driven by the knowledge is obtained.

[0010] Furthermore, the solution Obtaining the fault location driven by the knowledge includes: Step S101: Select one detection point as the reference point and all nodes on the branch close to the reference point as sub-reference points to generate a branch node matrix I : , in, Indicates a branch q The sub-reference point number on For branch q Non-sub reference point on the fault location ( q , d ) indicates that the fault is in the branch q Upper and lower reference points Point distance d Department; Step S102. Based on the distribution system topology, construct a network graph G = (V, E) with the busbar, feeder, and fault point F as nodes, the overhead lines or cables between the busbar and feeder fault points as edges, and the line length as the edge feature. V represents the vertex set of the network graph, and E represents the edge set of the network graph. The edge between nodes i and j in the network graph G is assigned a weight l. ij , otherwise the weight value is positive infinity, and the weight matrix is ​​obtained; Step S103: Obtain the shortest path matrix by solving the shortest distance between each node according to the weight matrix, and obtain the fault points through the shortest path matrix. F The shortest distance to each detection device .

[0011] Step S104: Use particle swarm algorithm to solve , and obtain the fault location driven by the knowledge.

[0012] Furthermore, in step S103, the Floyd algorithm is used to solve the shortest path matrix. In the solving process, starting from the weight matrix L, the path with the shortest length among all possible paths is selected to obtain the path from n i arrive n j The shortest path of , where possible paths include: If < n i , n j > exists, then the path exists< n i , n j > and the path does not contain other vertices, if < n i , n 1>、< n 1, n j > exists, then the path exists< n i , n 1, n j > and the vertex number contained in the path is not greater than 1, if < n i ,…, n 2>、< n 2,…, n j > exists, then there exists a path< n i ,…, n 2,…, n j >And the vertex numbers contained in the path are not greater than 2.

[0013] Furthermore, if it is an overhead line-cable hybrid line, it also includes building a normalized simplified model of the line: , in, Indicates the fault location, Indicates the cable length, v ol 、 v cable are the empirical values ​​of the speed of traveling waves in overhead lines and cables, respectively; After the fault location under the normalized line is solved, the original line location result is obtained by performing denormalization processing according to the following formula: , in, is the actual fault distance between the fault point and the branch sub-reference point. is the distance between the fault point and the sub-reference point of the branch obtained in the normalized line, for The length of overhead lines included in 、 are the empirical values ​​of the speed of traveling waves in overhead lines and cables, respectively.

[0014] Furthermore, in step S03, embedding the fault location and the corresponding time residual driven by the knowledge as mechanism features into the traveling wave spatiotemporal features extracted by the deep learning network model to form a mechanism-embedded fusion feature includes: The knowledge-driven location process for obtaining the fault location under the knowledge-driven approach is expressed as follows: , in, The fault characteristics selected by the knowledge-driven model for obtaining the fault location under the knowledge-driven method, is the fault location, K is the constraint between fault characteristics and fault location; The data-driven positioning process of the deep learning network model is expressed as: , in, D ( ) is the mining process of deep learning network model, is the fault traveling wave, G is a fully connected layer; The fault location obtained by knowledge-driven positioning The time residual corresponding to the fault location is used as the mechanism feature and embedded into the deep learning network model: , in, represents the fusion feature, Fault features mined by data-driven models, H represents the overall time residual of all detection devices.

[0015] Furthermore, in step S03, during the training process of the deep learning network model embedded in the mechanism, the loss function used is: , , , in, Indicates the distance between the predicted position driven by the deep learning network model data and the predicted position driven by knowledge. represents the fault location result obtained under knowledge-driven, represents the fault location result obtained based on the deep learning network model, dist() represents the shortest distance function, represents the physical inconsistency weight coefficient, Represents the distance between the actual position and the predicted position, Represents the overall loss function of the model, Indicates the actual fault location, N Indicates the number of samples.

[0016] Furthermore, in step S03, the deep learning network model is a BiGUR-Attention model, which is a model composed of BiGRU and self-attention mechanism. The input data of the BiGRU and the output data of the previous moment are respectively input into the reset gate and the update gate. The output of the reset gate is multiplied element-by-element by the output data of the previous moment, and the result is input into the hyperbolic tangent activation function with the input data to obtain a candidate output. The output of the update gate is subtracted to obtain a first output result and multiplied element-by-element with the output data of the previous moment to obtain a second output result. The first output result is multiplied element-by-element with the candidate output and then added to the second output result to obtain the final feature of the superposition of positive and negative direction features.

[0017] A distribution network fault locating device comprises a processor and a memory, wherein the memory is used to store a computer program, and the processor is used to execute the computer program to perform the distribution network fault locating method as described above.

[0018] Compared with the prior art, the present invention has the following beneficial effects: 1. The present invention collects the fault voltage traveling wave signal after the fault occurs, identifies the initial voltage traveling wave head, calculates the time residual of each detection device, and then obtains the fault location driven by knowledge. The time residual and the corresponding fault location are embedded as mechanism features into the spatiotemporal features extracted by data-driven deep learning network model, and a deep learning network model with embedded mechanism is constructed to obtain an enhanced positioning model, so that the model can consider the prior positioning results provided by the knowledge mechanism during the training process, improves the positioning accuracy of the model at the same training cost, and significantly reduces the generalization risk of the model under complex working conditions. At the same time, combined with the data-driven feature extraction capability, it can also enhance the adaptability of the positioning results to time synchronization errors, thereby effectively reducing time synchronization errors.

[0019] 2. The present invention adopts a knowledge-driven model based on time residuals to locate faults and obtains the fault location under knowledge-driven approach. It can fully consider the intrinsic relationship between the wave head arrival time and the fault point, and realize accurate fault location by combining the overall time residuals of each detection device. It can make full use of the knowledge mechanism information of each measuring point to realize knowledge-driven fault location. It adopts a deep learning network model to perform preliminary feature mining on the traveling wave in the time dimension, which is conducive to capturing the change law between adjacent wave heads, and further mining the spatial relationship between the fault location and each detection device, enhancing the features of the measuring point with a greater correlation with the fault location, weakening the features of the measuring point with a smaller correlation with the fault location, reducing the impact of bad data, and fully extracting the deep spatiotemporal features of the fault traveling wave.

[0020] 3. The present invention achieves precise positioning of fault points by comprehensively utilizing the practical guidance ability of knowledge methods and the hidden feature mining ability of data methods. It can effectively avoid the excessive dependence of knowledge-driven methods based on wave head calibration on time synchronization errors, improve the interpretability of data-driven methods, and solve the generalization risk problem of data-driven methods that are overly dependent on the scale and quality of data. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 This is a flow chart of a distribution network fault location method according to an embodiment of the present invention.

[0022] Figure 2 Schematic diagram of the GRU unit structure in an embodiment of the present invention.

[0023] Figure 3 Schematic diagram of the self-attention mechanism structure in an embodiment of the present invention.

[0024] Figure 4 This is a topological diagram of a 10kV complex distribution network in an embodiment of the present invention.

[0025] Figure 5 Schematic diagram of the positioning effects of different models in an embodiment of the present invention. DETAILED DESCRIPTION

[0026] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0027] like Figure 1 As shown, this embodiment provides a distribution network fault location method, including the following steps: Step S01: After a fault occurs, fault voltage traveling wave signals extracted by multiple detection devices installed on each branch of the distribution network are collected, and the initial voltage traveling wave head of the fault voltage traveling wave signal is identified; the time residual of each detection device is calculated based on the arrival time of the initial voltage traveling wave head of the fault voltage traveling wave, the fault occurrence time, and the shortest distance from the fault point to the location of each detection device, and the fault location is determined based on the time residual of each detection device in a knowledge-driven manner; Step S02: Constructing a sample library data set including the traveling wave signals within a specified time period after the wave head of each detection device arrives, the fault location driven by knowledge, and the corresponding time residual; Step S03: Construct a deep learning network model with embedded mechanism. In the model, the fault location and the corresponding time residual driven by knowledge are embedded as mechanism features into the traveling wave spatiotemporal features extracted by the deep learning network model to form a fusion feature with embedded mechanism. The fusion feature is mapped to the fault branch and fault location through a fully connected layer. The deep learning network model with embedded mechanism is trained using the sample library dataset to obtain the trained deep learning network model with embedded mechanism. Step S04: Obtain the fault voltage traveling wave signal after the real-time fault occurs, identify the initial voltage traveling wave head, calculate the knowledge-driven fault location and the corresponding time residual, input the trained mechanism embedded BiGUR-Attention deep learning network model, and obtain the fault point location output.

[0028] In this embodiment, before step S01, a simulation model with the same line parameters and topology as the distribution network to be tested is also built. In the simulation model, multiple fault points are set in each branch, and multiple fault conditions are set for each fault point. After the fault voltage traveling wave signal extracted by each detection device is collected in step S01, the fault voltage traveling wave signal is detected according to the change of the traveling wave head. , traveling wave head rise time t 1 and the amplitude of the traveling wave head The initial voltage traveling wave head of each fault voltage traveling wave signal is identified, and the corresponding wave head arrival time is obtained.

[0029] In a specific application embodiment, a simulation model with the same line parameters and topology is built in PSCAD / EMTDC software for a certain actual distribution network. To improve the generalization ability of the trained model, multiple fault points are set for each branch, and multiple fault conditions are set for each fault point. Traveling wave detection devices are installed at the ends of the line to collect the voltage traveling wave signals extracted by the detection devices at each end after the fault occurs. The following Crombel transform is used to obtain Modulus component: (1) in, for Modulus component, 、 They are phase A voltage and phase C voltage respectively.

[0030] Based on simulation experiments, the traveling wave head changes , traveling wave head rise time t 1 and the amplitude of the traveling wave head Identify the initial traveling wave head. If the setting value satisfies the following formula (2), it is determined to be the initial traveling wave head of the fault. The head detection of each end detection device is independent, and the head calibration time matrix can be obtained: t , t The number of elements in the equation is the number of end points in the distribution system, that is, the number of traveling wave monitoring devices.

[0031] (2) in, is the per unit voltage value.

[0032] It can be seen from the traveling wave transmission characteristics that the traveling wave head calibration time of each detection point is determined by the fault location, traveling wave velocity and fault start time. In this embodiment, the traveling wave detection device is defined as i Time residual y i for: (3) in, For the i The time residual of each detection device, is the time when the fault occurred, Fault point F To the current detection device i The shortest distance to the location, is the wave speed, Indicates the i The arrival time of the wave head at each detection device. When time error is not considered, the time residual has a minimum value of 0. The smaller the time residual, the more consistent the solved fault location, wave velocity, and fault occurrence time are with the actual values, and the higher the positioning reliability.

[0033] Therefore, the problem of finding the fault location can be viewed as finding a set of variables in a high-dimensional space. In this embodiment, the mean square error of the time residuals of all monitoring points is used. H Represents the overall time residual: (4) Among them, when H When it is 0, the shortest distance from the fault point to each detection device l FTiis the shortest distance to the actual fault, and the fault location is the actual fault location. Therefore, the knowledge-driven fault location problem can be expressed as follows: (5) in: n is the total number of measurement points, S represents a set of route points, t min is the minimum value of the arrival time of the initial traveling wave head in the detection device. v max The speed of light is 299.79 km / s in this embodiment. The three constraints correspond to three variables. The fault location selection range is the entire line, and the wave speed is considered to be 0.9. 1 times the speed of light, the fault occurrence time is based on the earliest detected wave head time, and the transmission time and detection error of the traveling wave are comprehensively considered. 10us.

[0034] To solve the above model (5), it is necessary to determine how to represent the fault location and how to obtain the shortest path from the fault point to each measuring point, and then solve the model.

[0035] In step S01 of this embodiment, the step of solving the model to obtain the fault location under knowledge drive includes: Step S101: Select one detection point as the reference point, and all detection points on the branch close to the reference point as sub-reference points to generate a branch node matrix I : (6) in, Indicates a branch q The sub-reference point number on For branch q If the fault is not a sub-reference point on the branch, then the position of the fault on any branch is represented by the sub-reference point of the corresponding branch. q Upper and lower reference points Point distance d The fault location at is expressed as ( q , d ); Step S102. Based on the distribution system topology, construct a network graph G = (V, E) with the busbar, feeder, and fault point F as nodes, the overhead lines or cables between the busbar and feeder fault points as edges, and the line length as the edge feature. V represents the vertex set of the network graph, and E represents the edge set of the network graph. The edge between nodes i and j in the network graph G is assigned a weight l. ij , otherwise the weight value is positive infinity, and the weight matrix is ​​obtained; Step S103. Obtain the shortest path matrix by solving the shortest distance between each node according to the weight matrix, and obtain the fault points through the shortest path matrix. F The shortest distance to each detection device .

[0036] Step S104: Use particle swarm algorithm to solve , and obtain the fault location driven by knowledge.

[0037] In this embodiment, for the overhead line-cable hybrid line, a normalized simplified model of the line can be constructed first: (7) in, Indicates the fault location, Indicates the cable length, v ol 、 v cable They are the empirical values ​​of the speed of traveling waves in overhead lines and cables, for example, ,It is understandable that the wave velocity of the traveling wave in overhead ,lines and cables can be selected from different values ​​according to ,the actual situation; After solving the fault location under the normalized line, the original line location result can be obtained by performing denormalization processing using the following formula: (8) in, is the actual fault distance between the fault point and the branch sub-reference point. is the distance between the fault point and the sub-reference point of the branch obtained in the normalized line, for The length of overhead lines included in 、 are the empirical values ​​of the speed of traveling waves in overhead lines and cables, respectively.

[0038] In this embodiment, the weight is an indicator of the number of edges assigned to the network graph. The weight matrix L for Matrix, get the weight matrix L Then, in step S103, the Floyd algorithm can be used to solve the shortest path matrix. Specifically, the weight matrix is ​​obtained. L Then, the shortest distance between nodes is obtained by Floyd algorithm. The Floyd algorithm is used from the weight matrix L Departure from n i arrive n j The shortest path is the shortest of the following possible paths: If ni , n j > exists, then the path exists< n i , n j >(the path does not contain other vertices); If n i , n 1>、< n 1, n j > exists, then the path exists< n i , n 1, n j >(the vertex number contained in the path is not greater than 1); If n i ,…, n 2>、< n 2,…, n j > exists, then there exists a path< n i ,…, n 2,…, n j >(The vertex number contained in the path is not greater than 2).

[0039] By analogy, n i arrive n j The shortest path should be the one with the shortest path length among the above paths. After traversing the weight matrix, the shortest path matrix D is obtained. Through the matrix D, the point F The shortest distance to each measuring point.

[0040] In this embodiment, the particle swarm (PSO) algorithm is used to solve the model. In the PSO algorithm, each particle is a solution in the solution space. In fact, it adjusts its flight according to its own flight experience and the flight experience of its companions. The best position each particle has experienced during the flight process is the optimal solution found by the particle itself. The best position experienced by the entire group is the optimal solution currently found by the entire group. The former is called the individual extreme value ( pBest ), the latter is called the global extreme value ( gBest ) In actual operation, the overall time deviation is used to evaluate the "goodness" of particles. Each particle continuously updates itself through the above two extreme values ​​to produce a new generation of population.

[0041] If the particle group size is N , then i (i =1,2,…, N The position of each particle can be expressed as X i , the “best” position it has ever experienced is recorded as pBest [ i ], its speed is V i Indicates that X i、 V i、 pBest [ i ] 、 gBest are defined as three-dimensional variables that characterize the traveling wave velocity, fault location, and fault onset time. The index number of the position of the "best" particle in the group is g So the particle i Will update its own speed and position according to the following formula: (9) in c 1, c 2 is a constant, called the learning factor; rand ()and Rand () is a random number on [0,1], w is the inertia weight. The above formula consists of three parts. The first part is the previous speed of the particle, which describes the current state of the particle; the second part is the cognitive part (cognition modal), which represents the thinking of the particle itself; the third part is the social part (social modal). The three parts together determine the spatial search ability of the particle. The first part plays a role in balancing the ability of global and local searches. The second part enables the particle to have a strong enough global search capability to avoid local minima. The third part reflects the information sharing between particles. Only with the joint action of these three parts can the particle effectively reach the best position. In addition, when the particle constantly adjusts its position according to the speed, it is also subject to the maximum speed. V max The restrictions when V i Exceed V max will be limited to V max .

[0042] In this embodiment, the traveling wave signal within 100us after the arrival of the wave head is intercepted. A sample consists of the traveling wave signal intercepted by the detection device at each end, the fault location obtained by the knowledge method, and its corresponding time residual. To reduce the influence of the fault transition resistance and the initial fault phase angle, the intercepted traveling wave signal can also be normalized by maximum and minimum per sample to establish a data sample library. The normalization formula is: (10) in, x is the original data, For normalized data, and are the maximum and minimum values ​​in the original data, respectively.

[0043] In this embodiment, the deep learning network model specifically adopts the BiGUR-Attention model, which is a model composed of BiGRU and self-attention mechanism. That is, the BiGUR-Attention deep learning network model with built-in mechanism is composed of three parts: bidirectional gate recurrent unit (BiGRU), self-attention mechanism (self-Attention) and built-in mechanism. After obtaining the sample library data set, the sample library data set is divided into a training set and a test set in proportion, and the BiGUR-Attention deep learning network model with built-in mechanism is trained using the training set. The model learning rate, batch size, and number of learning rounds are debugged according to the performance of the trained model on the test set.

[0044] In this embodiment, the input data of the bidirectional gated recurrent unit BiGRU and the output data at the previous moment are respectively input into the reset gate and the update gate, the output of the reset gate is multiplied element-by-element by the output data at the previous moment, and the result and the input data are input into the hyperbolic tangent activation function to obtain the candidate output; the output of the update gate is subtracted to obtain the first output result and multiplied element-by-element with the output data at the previous moment to obtain the second output result; the first output result is multiplied element-by-element with the candidate output and then added to the second output result to obtain the final feature of the superposition of positive and negative direction features.

[0045] Specifically, the gated recurrent unit (GRU) is a type of recurrent neural network (RNN) that can fully exploit the temporal information in traveling wave signals. The unit structure of GRU is as follows: Figure 2 As shown. Among them, is the Sigmoid function, represents the Hadamard product.

[0046] The core of GRU's work is to adjust the reset gate rt and update gate z t , learning the changing pattern of the sequence. t , in producing real output h t Before, a candidate output must be calculated n t The candidate output is composed of the output at the previous moment h t-1 and the current input x t In this process, the reset gate will weaken the information far away from the current time step, thereby enhancing the sensitivity of GRU to short-term rules and facilitating the capture of the changing rules between adjacent wave heads. Subsequently, the update gate will control the real output to h t-1 and n t If t When there is important information at all times, such as a sudden change in the waveform, the update gate will increase the influence of the candidate output; if the current information is invalid, the output will completely depend on the previous state, which enables GRU to better capture the overall change pattern in long sequences. The specific calculation process is as follows: (11) (12) (13) (14) in, W ir 、 W hr 、 W xz 、 W hz 、 W in 、 W hn is the weight parameter in each step, b r 、 b z 、 b n is the bias parameter in each step, is a fully connected layer with sigmoid activation function, is a fully connected layer with a hyperbolic tangent activation function. Since the traveling wave signal at each time point is affected by both past signals and future signals, it is necessary to comprehensively consider past and future information. A bidirectional gated recurrent unit is used to superimpose the features in the positive and negative directions through Equation (15) to obtain the final feature.

[0047] (15) Where, is the output of the sequence transmitted in the positive direction, The output of the sequence is transmitted in the reverse direction.

[0048] The BIGRU-Attention hybrid neural network model of this embodiment captures the dynamic temporal characteristics of traveling wave propagation through a bidirectional gated recurrent unit (BiGRU), and combines the self-attention mechanism to analyze the spatial correlation characteristics of multi-point signals, thereby achieving efficient characterization of the spatiotemporal coupling characteristics of fault traveling waves.

[0049] In this embodiment, in order to further mine fault features from the spatial dimension, the self-attention mechanism can be used to enhance the features of the measurement points that are more closely related to the fault location, while weakening the features of the measurement points that are less closely related to the fault location, thereby reducing the impact of bad data. The specific structure of the self-attention mechanism is as follows: Figure 3 As shown. The self-attention mechanism is divided into three steps: mapping, weighting and summation. First, use the matrix W Q 、 W K 、 W V Enter x Mapping to three spaces, we get Q 、 K 、 V Then, calculate Q and K The similarity between each element is obtained by the softmax activation function, and the weight is combined with V Multiply and sum the corresponding elements and output .

[0050] In this embodiment, the self-attention mechanism of the BiGUR-Attention deep learning network model is embedded in the mechanism. The calculation process of the self-attention mechanism is as follows: (16) (17) (18) in, x i 、 x j are different elements in the input, After linear transformation x i , is the weight coefficient, For elementsi With elements j The similarity between 、 、 is a learnable linear transformation matrix, is the length of the input sequence.

[0051] In this embodiment, the knowledge-driven fault location and the corresponding time residual are embedded as mechanism features into the traveling wave spatiotemporal features extracted by the deep learning network model to form the mechanism-embedded fusion features including: Step S301: The knowledge-driven location process of the fault location is expressed as follows: (19) in, Fault features selected for the knowledge-driven model used to obtain fault locations under knowledge-driven conditions, is the fault location, K is the constraint between fault characteristics and fault location; Step S302: The data-driven positioning process based on BiGRU-Attention is expressed as: (20) in, The fault features mined by the data-driven framework based on BiGRU-Attention are: D ( ) is the mining process, is the fault traveling wave, G is a fully connected layer; therefore, formula (20) can also be expressed as: (twenty one) Step S303: The fault location obtained by knowledge-driven positioning The time residual corresponding to the fault location is used as a mechanism feature and embedded in the data-driven framework positioning process based on BiGRU-Attention: , in, represents the fusion feature, K represents the constraints between fault characteristics and fault locations, D ( ) represents the mining process of data-driven models, Fault features mined by data-driven models, is the fault traveling wave, G Usually a fully connected layer, H Represents the overall time residual of the system.

[0052] The main difference between different data-driven models is D process, which is much more complex than G The stronger the task's demand for model fitting ability, the deeper the model will be. D The more complex the process, the worse the interpretability, making it difficult to strike a balance between the two. This embodiment uses feature fusion to ensure that the positioning results are determined by both the knowledge mechanism and the spatial and temporal characteristics of the traveling wave. Compared to data-driven approaches, embedding knowledge mechanisms can narrow the data-driven decision space, enabling better fitting of the nonlinear relationship between fault features and fault locations at the same training cost. At the same time, it can reduce the degree of deviation from the knowledge mechanism, avoid producing positioning results that seriously violate common sense, and improve interpretability.

[0053] To further enhance the constraints of the knowledge mechanism on model training, this embodiment introduces a physical inconsistency loss function into the loss function of the BiGUR-Attention deep learning network model training embedded in the mechanism. The physical inconsistency loss function is: (twenty three) in, represents the distance between the data-driven prediction position and the knowledge-driven prediction position, represents the positioning result based on knowledge-driven, represents the positioning result driven by BiGRU-Attention data, and dist( ) represents the shortest distance between the positioning result driven by knowledge and the positioning result driven by BiGRU-Attention data; Definition of the overall loss function for training the BiGUR-Attention deep learning network model embedded in the mechanism : (twenty four) (25) in, represents the distance between the data-driven prediction position and the knowledge-driven prediction position, represents the physical inconsistency weight coefficient, dist( ) represents the shortest distance between two locations, represents the positioning result based on knowledge-driven, Indicates the positioning result based on data driving, Represents the distance between the actual position and the predicted position, represents the actual fault location, and N represents the number of samples.

[0054] In this embodiment, the multi-terminal traveling wave signals are first used to preliminarily determine the fault location by solving the minimum time deviation, generating knowledge mechanism features. Simultaneously, the spatial features of the multidimensional traveling waves are extracted using the BiGRU-Attention data-driven framework. The knowledge mechanism features are then embedded into the traveling wave spatiotemporal features through matrix concatenation to form a fused feature representation. Finally, a fully connected layer maps the fused features to the fault branch and fault location. The current model parameters are evaluated using a loss function that considers physical inconsistencies, and the optimal model parameters are determined through backpropagation. The dataset is split into training and test sets in an 8:2 ratio. Finally, the model is trained offline using the Adam solver. The batch size and training rate are adjusted based on the localization performance of the test set, and the trained model is saved. After a fault occurs, the raw data is processed through steps S01 and S02 and input into the trained model. The model outputs the probability of each branch being a fault branch and the corresponding fault location. The branch with the highest fault probability is the fault branch predicted by the model.

[0055] In summary, the knowledge-driven model based on time residual proposed in this embodiment locates the fault by comparing the residual size. This model not only considers the intrinsic relationship between the wave head calibration time and the fault point, but also realizes accurate fault location by minimizing the overall time residual, and can make full use of the knowledge mechanism information of each measuring point. The proposed BiGUR-Attention deep learning network model performs preliminary feature mining on the traveling wave in the time dimension, which is conducive to capturing the change law between adjacent wave heads, and further explores the spatial relationship between the fault location and each detection device, enhances the features of the measuring point with a large correlation with the fault location, weakens the features of the measuring point with a small correlation with the fault location, reduces the impact of bad data, and can fully extract the deep spatiotemporal features of the fault traveling wave. The knowledge-driven model based on the optimization problem of the difference is combined with the data-driven model of the bidirectional gated recurrent unit integrated with the self-attention mechanism. The knowledge-driven positioning results and the corresponding time deviation are used as mechanism features and embedded in the spatiotemporal features extracted by data-driven to form a mechanism-embedded knowledge-data feature fusion. Then, the physical inconsistency of data-driven is considered in the loss function. The practical guidance ability of the knowledge method and the hidden feature mining ability of the data method are comprehensively utilized to achieve precise positioning of the fault point. This effectively avoids the excessive dependence of the knowledge-driven method based on wave head calibration on time synchronization error, improves the interpretability of the data-driven method, and solves the problem that the data-driven method is overly dependent on the scale and quality of data, has the risk of generalization, and is difficult to promote in actual engineering.

[0056] like Figure 4As shown, in this embodiment, a corresponding distribution network model is built on the PSCAD / EMTDC software, and frequency-variable parameter models are used for both overhead lines and cables. Traveling wave detection devices are installed at the ends of the lines, with a total of 16 measuring points and a sampling rate of 10Mhz. The hybrid model proposed in this invention is compared with a single model based on knowledge and a single model based on data. The knowledge model uses the method proposed in the literature [1] (Qiao Jian, Yin Xianggen, Wang Yikai, et al. A multi-terminal traveling wave fault location method for active distribution networkbased on residual clustering [J]. International Journal of Electrical Power&Energy Systems, 2021, 131: 107070.) and the method proposed in this article, which are model 1 and model 2 respectively. The data model uses the model proposed in this invention and the output layer structure remains unchanged, which is model 3. The sample library is generated by traversing Table 1. It is divided into a training set and a test set in a ratio of 8:2.

[0057] Table 1 Fault parameter traversal table of the total sample set

[0058] Considering the time error and the fact that the error point and the fault point are on the same line, we select the test set samples that can meet this condition and randomly add a 3~6us time synchronization error to the measurement point where the fault line is located. The positioning effects of different models are shown in Table 2 and Figure 5 shown.

[0059] Table 2 Positioning effects of different models

[0060] It can be seen that the overall positioning performance of the knowledge-based model has declined, with positioning accuracy exceeding 600m, the minimum positioning error exceeding 300m, and a branch identification accuracy of less than 56%, making it difficult to meet practical requirements. This is mainly because the knowledge-driven model can only locate based on traveling wave signals in one direction. The impact on knowledge-driven methods is similar to the impact of time synchronization error in dual-terminal positioning, resulting in a significant reduction in positioning accuracy. In contrast, when time synchronization error exists at a single measurement point, the positioning accuracy and branch identification accuracy of the data-driven model 3 remain largely unchanged, maintaining high positioning accuracy for most fault scenarios. However, the maximum positioning error violates the knowledge mechanism to a high degree. Although the average positioning error derived from the knowledge mechanism reaches 618.9m in this case, significantly reducing its knowledge constraint capability, the data-driven approach can still narrow the model's decision space from the entire line network to a certain range of the fault point. Therefore, the proposed method achieves the highest positioning accuracy and branch identification accuracy on the same test set, reducing the maximum positioning error by 64% compared to the data-driven approach, improving interpretability to a certain extent.

[0061] The following experiments are conducted on the positioning results of the present invention under different fault conditions: like Figure 4 As shown in Figure 1, a corresponding distribution network model was constructed using PSCAD / EMTDC software. Frequency-variable parameter models were used for both overhead lines and cables. Traveling wave detection devices were installed at each line end, with a total of 16 measurement points and a sampling rate of 10 MHz. Different fault conditions were set to test the location results. A 5 µs time synchronization error was added to measurement points A and H. Measurement point A was set as the reference point, and the fault distance was expressed as the distance from the fault point to the sub-reference point on the faulty branch line. The simulation structure is shown in Table 3.

[0062] Table 3 Positioning results under different fault conditions

[0063] Analysis of the positioning results shows that the method of the present invention can still operate reliably under extreme working conditions and is not affected by fault distance, fault type, fault initial phase angle and fault transition resistance. However, the positioning accuracy will be significantly reduced after the fault transition resistance reaches 1500Ω, but relatively accurate positioning can still be performed.

[0064] This embodiment further provides a distribution network fault location system, including a processor and a memory, wherein the memory is used to store a computer program, and the processor is used to execute the computer program to perform a distribution network fault location method.

[0065] Those skilled in the art will appreciate that the above-mentioned embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application may take the form of a computer program product implemented on one or more computer-readable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of processes and / or boxes in the flowchart and / or block diagram, may be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the steps in the process. Figure 1 a process or multiple processes and / or boxes Figure 1 These computer program instructions can also be stored in a computer-readable memory that can guide a computer or other programmable data processing device to work in a specific way, so that the instructions stored in the computer-readable memory produce a product including the instruction device, which implements the function specified in the process. Figure 1 a process or multiple processes and / or boxes Figure 1 These computer program instructions can also be loaded onto a computer or other programmable data processing device, so that a series of operation steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process in the process. Figure 1 a process or multiple processes and / or boxes Figure 1 The above is only a preferred embodiment of the present invention and does not limit the present invention in any form. Although the present invention has been disclosed as above with preferred embodiments, it is not intended to limit the present invention. Therefore, any simple modification, equivalent change and modification made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall fall within the scope of protection of the technical solution of the present invention.

[0066] The above description is merely a preferred embodiment of the present invention and does not constitute any form of limitation to the present invention. Although the present invention has been disclosed above with reference to the preferred embodiment, it is not intended to limit the present invention. Therefore, any simple modifications, equivalent variations, and modifications to the above embodiment that do not depart from the technical solution of the present invention and are based on the technical essence of the present invention shall fall within the scope of protection of the technical solution of the present invention.

Claims

1. A distribution network fault location method integrating mechanism characteristics and deep learning network model, characterized in that: The following steps are involved: Step S01: collecting fault voltage traveling wave signals extracted by multiple detection devices installed on each branch of the distribution network after a fault occurs, and identifying the initial voltage traveling wave head of the fault voltage traveling wave signal; The time residual of each detection device is calculated based on the arrival time of the initial voltage traveling wave head of the fault voltage traveling wave, the fault occurrence time, and the shortest distance from the fault point to the location of each detection device. The fault location under knowledge-driven conditions is determined based on the time residual of each detection device. Step S02: constructing a sample library data set including the traveling wave signals of each detection device within a specified time period after the wave head arrives, the fault location driven by the knowledge, and the corresponding time residual; Step S03: Constructing a deep learning network model with embedded mechanism, in which the fault location and corresponding time residual driven by the knowledge are embedded as mechanism features into the traveling wave spatiotemporal features extracted by the deep learning network model to form a fusion feature embedded in the mechanism, and the fusion feature is mapped to the fault branch and fault location through a fully connected layer; using the sample library data set to train the deep learning network model with embedded mechanism, to obtain a trained deep learning network model with embedded mechanism; Step S04: Obtain the fault voltage traveling wave signal after the real-time fault occurs, identify the initial voltage traveling wave head, calculate the knowledge-driven fault location and the corresponding time residual, input the trained deep learning network model embedded in the mechanism, and obtain the positioning result output of the fault point location.

2. The distribution network fault location method integrating mechanism characteristics and deep learning network model according to claim 1 is characterized in that: Before step S01, it also includes building a simulation model with the same line parameters and topology as the distribution network under test. In the simulation model, multiple fault points are set in each branch, and multiple fault conditions are set for each fault point. After the fault voltage traveling wave signal extracted by each detection device is collected in step S01, the fault voltage traveling wave signal is detected according to the change of the traveling wave head. , traveling wave head rise time t 1 and the amplitude of the traveling wave head The initial voltage traveling wave head of each fault voltage traveling wave signal is identified, and the corresponding wave head arrival time is obtained.

3. The distribution network fault location method integrating mechanism characteristics and deep learning network model according to claim 1 is characterized in that: The calculation expression of the time residual of each detection device is: in, For the i The time residual of each detection device, is the time when the fault occurred, Fault point F To the current detection device i The shortest distance to the location, is the wave speed, Indicates the i The wave head arrival time of each detection device; Calculate the overall time residual of all detection devices: ,in n Indicates the number of detection devices, H Represents the overall time residual of the system; Solution The fault location driven by the knowledge is obtained.

4. The distribution network fault location method integrating mechanism characteristics and deep learning network model according to claim 3 is characterized in that: The solution Obtaining the fault location driven by the knowledge includes: Step S101: Select one detection point as the reference point and all nodes on the branch close to the reference point as sub-reference points to generate a branch node matrix I : in, Indicates a branch q The sub-reference point number on For branch q Non-sub reference point on the fault location ( q , d ) indicates that the fault is in the branch q Upper and lower reference points Point distance d Department; Step S102. Based on the distribution system topology, construct a network graph G = (V, E) with the busbar, feeder, and fault point F as nodes, the overhead lines or cables between the busbar and feeder fault points as edges, and the line length as the edge feature. V represents the vertex set of the network graph, and E represents the edge set of the network graph. The edge between nodes i and j in the network graph G is assigned a weight l. ij , otherwise the weight value is positive infinity, and the weight matrix is ​​obtained; Step S103: Obtain the shortest path matrix by solving the shortest distance between each node according to the weight matrix, and obtain the fault points through the shortest path matrix. F The shortest distance to each detection device ; Step S104. Solve using particle swarm algorithm , and obtain the fault location driven by the knowledge.

5. The distribution network fault location method integrating mechanism characteristics and deep learning network model according to claim 4 is characterized in that: In step S103, the Floyd algorithm is used to solve the shortest path matrix. In the solution process, starting from the weight matrix L, the path with the shortest length among all possible paths is selected to obtain the path from n i arrive n j The shortest path of , where possible paths include: If < n i , n j > exists, then the path exists< n i , n j > and the path does not contain other vertices, if < n i , n 1>、< n 1, n j > exists, then the path exists< n i , n 1, n j > and the vertex number contained in the path is not greater than 1, if < n i ,…, n 2>、< n 2,…, n j > exists, then there exists a path< n i ,…, n 2,…, n j >And the vertex numbers contained in the path are not greater than 2.

6. The distribution network fault location method integrating mechanism characteristics and deep learning network model according to claim 4 is characterized in that: If it is an overhead line-cable hybrid line, it also includes building a normalized simplified model of the line: in, Indicates the fault location, Indicates the cable length, v ol 、 v cable are the empirical values ​​of the speed of traveling waves in overhead lines and cables, respectively; After the fault location under the normalized line is solved, the original line location result is obtained by performing denormalization processing according to the following formula: in, is the actual fault distance between the fault point and the branch sub-reference point. is the distance between the fault point and the sub-reference point of the branch obtained in the normalized line, for The length of overhead lines included in 、 are the empirical values ​​of the speed of traveling waves in overhead lines and cables, respectively.

7. The distribution network fault location method integrating mechanism characteristics and deep learning network model according to any one of claims 1 to 6, characterized in that: In step S03, the process of embedding the knowledge-driven fault location and the corresponding time residual as mechanism features into the traveling wave spatiotemporal features extracted by the deep learning network model to form a mechanism-embedded fusion feature includes: The knowledge-driven location process for obtaining the fault location under the knowledge-driven approach is expressed as follows: in, The fault characteristics selected by the knowledge-driven model for obtaining the fault location under the knowledge-driven method, is the fault location, K is the constraint between fault characteristics and fault location; The data-driven positioning process of the deep learning network model is expressed as: in, D ( ) is the mining process of deep learning network model, is the fault traveling wave, G is a fully connected layer; The fault location obtained by knowledge-driven positioning The time residual corresponding to the fault location is used as the mechanism feature and embedded into the deep learning network model: in, represents the fusion feature, Fault features mined by data-driven models, H represents the overall time residual of all detection devices.

8. The distribution network fault location method integrating mechanism characteristics and deep learning network model according to any one of claims 1 to 6, characterized in that: In step S03, during the training process of the deep learning network model embedded in the mechanism, the loss function used is: in, Indicates the distance between the predicted position driven by the deep learning network model data and the predicted position driven by knowledge. represents the fault location result obtained under knowledge-driven, represents the fault location result obtained based on the deep learning network model, dist() represents the shortest distance function, represents the physical inconsistency weight coefficient, Represents the distance between the actual position and the predicted position, Represents the overall loss function of the model, Indicates the actual fault location, N Indicates the number of samples.

9. The distribution network fault location method integrating mechanism characteristics and deep learning network model according to any one of claims 1 to 6, characterized in that: In step S03, the deep learning network model is a BiGUR-Attention model, which is a model composed of BiGRU and a self-attention mechanism. The input data of the BiGRU and the output data of the previous moment are respectively input into a reset gate and an update gate. The output of the reset gate is element-wise multiplied by the output data of the previous moment, and the result is input into a hyperbolic tangent activation function with the input data to obtain a candidate output. The output of the update gate is subtracted to obtain a first output result and element-wise multiplied with the output data of the previous moment to obtain a second output result. The first output result is element-wise multiplied with the candidate output and then added to the second output result to obtain a final feature of the superposition of positive and negative direction features.

10. A distribution network fault location device, comprising a processor and a memory, wherein the memory is used to store a computer program, characterized in that: The processor is configured to execute the computer program to perform the distribution network fault location method according to any one of claims 1 to 9.

Images