Intelligent Monitoring and Fault Isolation Method and System for Branch Box

Abstract

The present invention belongs to the technical field of fault analysis and monitoring, and specifically relates to a method and system for intelligent monitoring and fault isolation of branch boxes. The method includes: performing multi-physical field data fusion acquisition on the branch boxes to obtain a spatio-temporal feature matrix; extracting dynamic features from the multi-physical field data of the branch boxes and outputting a fused dynamic feature vector of the branch boxes; obtaining the topological graph structure of the branch boxes to get a node fault probability matrix of the branch boxes; constructing a physically constrained fault propagation model based on the node fault probability matrix of the branch boxes to obtain a pre-isolation set and output a dynamic blocking instruction; using the pre-isolation set as an initial solution space constraint, outputting the fault type and obtaining an isolation instruction; setting branch box topological reconstruction constraints, and performing closed-loop control on each branch box node in combination with the isolation instruction to obtain the switch action sequence of each branch box node and the reconstructed power supply path. The present invention can achieve precise monitoring, rapid isolation and intelligent reconstruction of branch box faults, and significantly improve the operation and maintenance efficiency of the distribution network.

Description

Technical Field

[0001] The present invention belongs to the technical field of fault analysis and monitoring, and in particular relates to a method and system for intelligent monitoring and fault isolation of a branch box. Background Art

[0002] The distribution cable branch box is an indispensable part of the power distribution network. It can realize cable tapping and transfer. For example, when the line is outgoing to the power load, the main cable is often used to outgoing line. When close to the load, the branch box is used to branch into small area cables to access the load. When the line is long, the branch box can also be used for transfer. Its normal operation is crucial to the stability and security of the power system. With the development of smart grids, the requirements for intelligent monitoring and automated operation and maintenance of distribution network equipment are constantly increasing. Advanced technologies are needed to achieve remote monitoring of equipment operating status and rapid fault isolation to improve the reliability and efficiency of distribution network operation, improve power supply quality, and reduce labor intensity. Therefore, research on intelligent monitoring and fault isolation methods of branch boxes has emerged.

[0003] Traditional branch box intelligent monitoring and fault isolation methods include:

[0004] Single parameter threshold monitoring method: By setting thresholds for key parameters such as temperature and current, when the monitored parameters exceed the thresholds, it is determined that the branch box may be faulty and corresponding isolation measures are taken.

[0005] Monitoring isolation method based on simple relay protection: Use relay protection devices to detect fault signals such as short circuit and overload in the branch box circuit. Once the protection action value is reached, the fault line is quickly cut off to achieve isolation.

[0006] Traditional distribution cable branch box status monitoring only focuses on whether individual data meets maintenance indicators. It is unable to integrate internal distribution cable transmission data, connection information and internal environmental conditions to comprehensively evaluate the cable operation status and internal conditions of the branch box. It is difficult to detect faults in advance, resulting in low maintenance efficiency and poor safety, and limited fault isolation capabilities. Summary of the Invention

[0007] In view of the deficiencies in the prior art, the purpose of the present invention is to provide a method and system for intelligent monitoring and fault isolation of branch boxes, which can achieve accurate monitoring, rapid isolation and intelligent reconstruction of branch box faults, and significantly improve the operation and maintenance efficiency of the distribution network.

[0008] To achieve the above objectives, the present invention is implemented through the following technical solutions: a branch box intelligent monitoring and fault isolation method, comprising the following steps:

[0009] S1, perform branch box multi-physics field data fusion collection to obtain the spatiotemporal feature matrix;

[0010] S2. Based on the spatio-temporal feature matrix, perform dynamic feature extraction on the multi-physical field data of the branch box, and output the fused dynamic feature vector of the branch box;

[0011] S3. Obtain the topological graph structure of the branch box, combine it with the dynamic feature vector of the branch box, and perform spatio-temporal graph convolutional network analysis to obtain the node fault probability matrix of the branch box;

[0012] S4. Based on the node fault probability matrix of the branch box, construct a physically constrained fault propagation model, identify the critical path, obtain the pre-isolation set, analyze the nodes in the pre-isolation set, output the dynamic blocking instruction, and determine whether to take blocking measures for the nodes in the pre-isolation set;

[0013] S5. Perform quantum decision optimization based on the node fault probability matrix of the branch box, use the pre-isolation set as the initial solution space constraint, output the fault type and obtain the isolation instruction;

[0014] S6. Set the topological reconstruction constraint of the branch box, and perform closed-loop control on each branch box node in combination with the isolation instruction to obtain the switch action sequence of each branch box node and the reconstructed power supply path.

[0015] Preferably, in S1, the process of obtaining the spatio-temporal feature matrix is as follows:

[0016] Perform multi-physical field data fusion acquisition of the branch box, and the acquisition content includes the temperature field of the branch box at time t , the current waveform of the branch box , the vibration signal of the branch box and the partial discharge amount of the branch box ;

[0017] Construct a three-dimensional data cube X:

[0018] ;

[0019] where N is the number of sensor nodes, L is the time window length, is the total number of multi-modal data channels, R is the set of real numbers, represents dimensional real space, Concatenate represents concatenation, and Stack represents stacking operation;

[0020] Reduce the dimension through tensor decomposition, and use Tucker decomposition to extract the core features:

[0021] ;

[0022] where G is the core tensor, is the transformation matrix on the node dimension, is the transformation matrix on the time dimension, is the transformation matrix in the sensor dimension; indicates transformation on the first dimension of the tensor, i.e., the node dimension; indicates transformation on the second dimension of the tensor, i.e., the time dimension; indicates transformation on the third dimension of the tensor, i.e., the sensor dimension;

[0023] to obtain a compact spatio-temporal feature matrix:

[0024] ;

[0025] In the formula, is the spatio-temporal feature matrix, is the total number of updated multi-modal data channels, represents a real number space of dimension, represents the mode-1 unfolding matrix of the core tensor G.

[0026] Preferably, in the above S2, the process of outputting the fused dynamic feature vector of the branch box is as follows:

[0027] Input the spatio-temporal feature matrix and the original high-frequency current signal ;

[0028] Conduct transient current analysis:

[0029] For perform 8-layer wavelet packet decomposition to obtain 256 sub-bands:

[0030] ;

[0031] In the formula, is the result of wavelet packet decomposition, is the coefficient of the k-th sub-band after decomposition, is the wavelet packet basis function of the k-th sub-band;

[0032] Calculate the energy entropy of each sub-band :

[0033] ;

[0034] Select the 3 sub-bands with the largest energy entropy, denoted as , , , and their coefficients , , constitute the transient feature :

[0035] ;

[0036] Perform temperature-vibration coupling analysis and calculate the correlation index between the temperature gradient and vibration energy :

[0037] ;

[0038] In the formula, is the gradient of temperature T in the x direction of space, is the calculation of the vibration signal energy over the time interval represents the time interval;

[0039] Obtain the partial discharge peak value of the branch box ;

[0040] Output the fused dynamic feature vector of the branch box :

[0041] ;

[0042] Among them, represents vector transpose.

[0043] Preferably, in the step S3, the process of obtaining the node fault probability matrix of the branch box is as follows:

[0044] Obtain the topological graph structure of the branch box , where V represents the set of nodes of the topological graph structure of the branch box, including the i-th node and the h-th node , and E represents the set of edges of the topological graph structure of the branch box;

[0045] Set the attribute of the i-th node , is the current of the i-th node, is the temperature of the i-th node, is the equivalent resistance of the i-th node;

[0046] The edge weight between the i-th node and the h-th node ;

[0047] In the formula, is the conductor resistance between the i-th node and the h-th node, is the imaginary unit, is the power grid angular frequency, is the line inductance between the i-th node and the h-th node;

[0048] Construct a spatio-temporal graph convolutional layer:

[0049] ;

[0050] In the formula, is the feature matrix of the (l + 1)-th layer of the output, is the feature matrix of the l-th layer, is the activation function, A is the angle matrix, and D is the normalized adjacency matrix, represents the negative half power of D, is the weight matrix of the l-th layer, is the hyperparameter for balancing spatial convolution and temporal convolution, is the temporal convolution operation;

[0051] Input the dynamic feature vector of the branch box and the attributes of each node;

[0052] Output the node failure probability:

[0053] ;

[0054] In the formula, is the failure probability of the i-th node, is the final hidden layer feature of the node after spatio-temporal graph convolution processing, is the classifier weight vector, and Sigmoid is the activation function;

[0055] Based on the failure probabilities of each node, output the node failure probability matrix of the branch box .

[0056] Preferably, in S4, the process of obtaining the pre-isolation set is as follows:

[0057] Construct a failure propagation model with physical constraints:

[0058] ;

[0059] In the formula, is the failure probability of the i-th node at time t, represents the rate of change with respect to time t, that is, the speed of failure propagation; is the learnable parameter of the topological propagation term, is the learnable parameter of the physical strengthening term, is the weight of the adjacency matrix, is the failure probability of the h-th node, is the summation operation for all neighbor nodes of the i-th node, is the set of neighbor nodes of the i-th node, is the maximum value of, is the topological propagation term, is the physical strengthening term;

[0060] Solving the propagation range within the future time interval based on the graph diffusion equation within:

[0061] ;

[0062] In the formula, is the failure propagation probability vector of the node at time , is the failure propagation probability vector of the node at time t, is the matrix exponential operation, is the Laplacian matrix, and e is the natural constant;

[0063] Performing critical path identification:

[0064] Defining the propagation risk index:

[0065] ;

[0066] In the formula, is the propagation risk index of the i-th node, is the failure probability of the g-th node , represents the summation of the failure propagation probabilities of all nodes on the propagation path of the i-th node, and g represents the node index on the propagation path of the i-th node;

[0067] Screening the propagation risk index higher than the risk index threshold stored in the database corresponding nodes to form a pre-isolation set .

[0068] Preferably, in the step S4, the process of determining whether to take blocking measures for the nodes in the pre-isolation set is as follows:

[0069] Setting a dynamic blocking instruction :

[0070] ;

[0071] In the formula, 1 indicates that blocking measures need to be taken for this node, and 0 indicates that no blocking is required;

[0072] Outputting the dynamic blocking instruction .

[0073] Preferably, in the step S5, the process of outputting the failure type and obtaining the isolation instruction is as follows:

[0074] Based on the failure probability of the branch box node, performing quantum decision optimization and first classifying the failures:

[0075] ;

[0076] In the formula, is the quantum potential function value corresponding to the y-th type of fault, u is the upper limit of the number of summation terms, representing the number of characteristic dimensions involved in the calculation, is the r-th weight parameter related to the y-th type of fault, is the dynamic feature vector is the r-th eigenvalue in is the central value of the y-th type of fault in the r-th characteristic dimension, represents the quantum phase similarity function;

[0077] Then calculate the classification probability:

[0078] ;

[0079] In the formula, y is the index of the fault category, Y is the total number of fault categories, represents the sample belongs to the probability of the y-th type of fault, is the temperature parameter;

[0080] Output the fault type of the branch box node;

[0081] Isolation strategy optimization, quantum annealing to solve the optimal switch combination, and use the pre-isolation set as the initial solution space constraint:

[0082] ;

[0083] In the formula, is the optimal switch combination, argmin represents finding the parameter value that minimizes the objective function, is the set of all possible switch combination vectors, n is the total number of switches, is the fault classification loss weight coefficient, is the switch operation loss weight coefficient, is the fault classification loss function, is the switch operation loss function;

[0084] Output the optimal switch combination as the isolation instruction.

[0085] Preferably, the process of using the pre-isolation set as the initial solution space constraint is:

[0086] Modify the loss function to:

[0087] ;

[0088] In the formula, is the modified loss function, is the original loss function, which includes the fault classification loss function and the switch operation loss function. is the weight coefficient. is the i-th node in the set of nodes representing the branch box topology structure. corresponding switch state. is the pre-isolation set.

[0089] Preferably, in the above S6, the process of obtaining the switch action sequence of each branch box node and the reconstructed power supply path is as follows:

[0090] Set the branch box topology reconstruction constraints, including the line rated current and the allowable voltage range.

[0091] Input the isolation instruction, and generate the switch control signal according to the isolation instruction:

[0092] ;

[0093] In the formula, is the action signal of switch d at time t. is the isolation instruction of switch d. is the time point to trigger the switch action.

[0094] When , that is, when the isolation instruction requires the switch to be disconnected and , the trigger time is reached, and the action starts to be executed. indicates that the switch is disconnected, otherwise , indicating that the switch maintains its original state or is closed.

[0095] Based on the action signals of the switch at each time point, output the switch action sequence.

[0096] Improve the Dijkstra algorithm to solve the optimal path:

[0097] ;

[0098] In the formula, p is a node or line segment in the path. is the set of all possible paths. is the current passing through path p. is the resistance of path p. represents the power loss on path p. is the weight coefficient. is the indicator function, which is 1 when path p is in the fault area and 0 otherwise.

[0099] Output the reconstructed power supply path. .

[0100] The branch box intelligent monitoring and fault isolation system is used to implement the above method, including:

[0101] The multi-physical field data fusion acquisition module is used to perform multi-physical field data fusion acquisition of the branch box to obtain a spatio-temporal feature matrix;

[0102] The dynamic feature vector fusion module is used to perform dynamic feature extraction on the multi-physical field data of the branch box based on the spatio-temporal feature matrix, and output the fused dynamic feature vector of the branch box;

[0103] The fault probability matrix acquisition module is used to obtain the topological graph structure of the branch box, combine the dynamic feature vector of the branch box, and perform spatio-temporal graph convolutional network analysis to obtain the branch box node fault probability matrix;

[0104] The dynamic blocking instruction output module constructs a physically constrained fault propagation model based on the branch box node fault probability matrix, identifies the critical path, obtains the pre-isolation set, analyzes the nodes in the pre-isolation set, outputs the dynamic blocking instruction, and determines whether to take blocking measures for the nodes in the pre-isolation set;

[0105] The isolation instruction analysis module is used to perform quantum decision optimization based on the branch box node fault probability matrix, use the pre-isolation set as the initial solution space constraint, output the fault type and obtain the isolation instruction;

[0106] The power supply path reconstruction module is used to set the branch box topology reconstruction constraint, combine the isolation instruction to perform closed-loop control on each branch box node, and obtain the switch action sequence of each branch box node and the reconstructed power supply path.

[0107] The present invention has the following beneficial effects:

[0108] Through multi-physical field data fusion acquisition and spatio-temporal feature matrix construction, the present invention comprehensively integrates multi-dimensional data such as temperature, current, vibration, and partial discharge, breaks the limitations of traditional single monitoring, realizes the accurate characterization of the thermal-electric-mechanical state of the branch box, and provides rich basis for fault diagnosis. Further, by using wavelet packet decomposition and energy entropy screening to extract transient features, calculate the temperature-vibration coupling index, and fuse the partial discharge peak value, a dynamic feature vector is constructed to accurately capture the subtle changes in the operation state of the equipment, effectively improving the sensitivity and lead time of fault warning.

[0109] In terms of fault diagnosis and propagation analysis, the solution of the present invention innovatively combines the branch box topology structure with dynamic characteristics, utilizes the spatio-temporal graph convolutional network to deeply mine the spatial correlation and temporal variation law between nodes, and outputs a high-precision node fault probability matrix. A fault propagation model including topological propagation and physical reinforcement is constructed, and the graph diffusion equation is combined to predict the fault diffusion path, identify high-risk key nodes and form a pre-isolation set, providing a scientific basis for formulating accurate isolation strategies and minimizing the scope of fault impact.

[0110] The present invention introduces quantum decision optimization technology into the fault isolation process. With the pre-isolation set as the constraint condition, the optimal switch combination is quickly solved through the quantum annealing algorithm, greatly shortening the decision-making time in complex scenarios. Combining closed-loop control with the improved Dijkstra algorithm, a switch action sequence is dynamically generated and the power supply path is reconstructed to ensure the accuracy and timeliness of the isolation operation. At the same time, a real-time feedback mechanism and dynamic fault tolerance logic are incorporated. After the switch action, the status of key nodes is continuously monitored. Once a fault spread is detected, a secondary reconstruction is immediately triggered, and the standby link is linked to form a self-healing ring network to ensure power supply continuity and system stability, significantly improving the intelligent operation and maintenance level of the distribution network. Brief Description of the Drawings

[0111] Figure 1 is a schematic flow chart of the method of the present invention;

[0112] Figure 2 is a schematic module diagram of the system of the present invention. Detailed Embodiments

[0113] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention.

[0114] Embodiment 1: As Figure 1 shown, the intelligent monitoring and fault isolation method for a branch box includes the following steps:

[0115] S1. Perform multi-physical field data fusion acquisition of the branch box to obtain a spatio-temporal feature matrix;

[0116] S2. Based on the spatio-temporal feature matrix, perform dynamic feature extraction on the multi-physical field data of the branch box, and output the fused dynamic feature vector of the branch box;

[0117] S3. Obtain the topological graph structure of the branch box, combine it with the dynamic feature vector of the branch box, and perform spatio-temporal graph convolutional network analysis to obtain the branch box node fault probability matrix;

[0118] S4. Construct a fault propagation model with physical constraints based on the branch box node fault probability matrix, identify the critical path, obtain the pre-isolation set, analyze the nodes in the pre-isolation set, output dynamic blocking instructions, and determine whether to take blocking measures for the nodes in the pre-isolation set;

[0119] S5. Perform quantum decision optimization based on the branch box node fault probability matrix, use the pre-isolation set as the initial solution space constraint, output the fault type, and obtain the isolation instruction;

[0120] S6. Set the branch box topology reconstruction constraint, perform closed-loop control on each branch box node in combination with the isolation instruction, and obtain the switch action sequence of each branch box node and the reconstructed power supply path.

[0121] In S1, the process of obtaining the spatio-temporal feature matrix is as follows:

[0122] Perform multi-physical field data fusion acquisition of the branch box. The acquisition content includes the branch box temperature field at time t , the branch box current waveform , the branch box vibration signal and the partial discharge amount of the branch box ;

[0123] Construct a three-dimensional data cube X:

[0124] ;

[0125] where N is the number of sensor nodes, L is the time window length, is the total number of multi-modal data channels, R is the set of real numbers, denotes dimensional real space, Concatenate represents concatenation, and Stack represents a stacking operation;

[0126] Reduce the dimension through tensor decomposition, and use Tucker decomposition to extract the core features:

[0127] ;

[0128] where G is the core tensor, is the transformation matrix in the node dimension, is the transformation matrix in the time dimension, is the transformation matrix in the sensor dimension; represents transforming the first dimension of the tensor, i.e., the node dimension; represents transforming the second dimension of the tensor, i.e., the time dimension; represents transforming the third dimension of the tensor, i.e., the sensor dimension;

[0129] Obtain a compact spatio-temporal feature matrix:

[0130] ;

[0131] In the formula, is the spatio-temporal feature matrix, is the total number of updated multi-modal data channels, represents a real number space of dimension, represents the mode-1 expansion matrix of the core tensor G.

[0132] Collect multi-physical field data such as the temperature field, current waveform, vibration signal, and partial discharge quantity of the branch box, which can reflect the operating state of the branch box from multiple dimensions, cover multi-faceted information such as heat, electricity, and machinery, avoid the one-sidedness of single-data monitoring, and comprehensively grasp the health status of the branch box.

[0133] Construct a three-dimensional data cube and reduce the dimension through tensor decomposition to obtain a compact spatio-temporal feature matrix. The dimension reduction operation removes redundant information, reduces the data volume, improves the data storage and processing efficiency, makes the subsequent analysis more efficient, and at the same time retains key features without losing important information.

[0134] Use Tucker decomposition to extract core features, accurately refine key information from multi-dimensional complex data, and the obtained spatio-temporal feature matrix can more accurately characterize the operating characteristics of the branch box, providing a high-quality data basis for subsequent fault diagnosis, analysis, etc., and improving the accuracy and reliability of the analysis results.

[0135] In S2, the process of outputting the fused dynamic feature vector of the branch box is as follows:

[0136] Input the spatio-temporal feature matrix and the original high-frequency current signal ;

[0137] Perform transient current analysis:

[0138] For perform 8-layer wavelet packet decomposition to obtain 256 sub-bands:

[0139] ;

[0140] In the formula, is the result of wavelet packet decomposition, is the coefficient of the k-th sub-band after decomposition, is the wavelet packet basis function of the k-th sub-band (the wavelet packet basis function is an extension of the wavelet basis function and is used for wavelet packet transform);

[0141] Calculate the energy entropy of each sub-band :

[0142] ;

[0143] Select the 3 sub-bands with the largest energy entropy and denote them as , , , and their coefficients , , constitute the transient feature :

[0144] ;

[0145] Conduct temperature-vibration coupling analysis and calculate the correlation index of the temperature gradient and vibration energy :

[0146] ;

[0147] In the formula, is the gradient of temperature T in the x direction of space, is the calculation of the vibration signal energy over the time interval , represents the time interval;

[0148] Obtain the partial discharge peak value of the branch box ;

[0149] Output the fused dynamic feature vector of the branch box :

[0150] ;

[0151] Among them, represents the vector transpose.

[0152] Perform 8-layer wavelet packet decomposition on the original high-frequency current signal to obtain 256 sub-bands, and then screen out the 3 sub-bands with the largest energy entropy through energy entropy to construct transient features. Wavelet packet decomposition can expand the signal at different frequency scales, and energy entropy screening focuses on the frequency components that can best reflect transient changes, which can accurately capture the transient anomalies of the current during the operation of the branch box and timely detect transient fault signs such as short circuits and lightning strikes.

[0153] Calculate the correlation index of the temperature gradient and vibration energy, and comprehensively consider the interaction between the two physical quantities of temperature and vibration. During the operation of the branch box, equipment heating and mechanical vibration are often interrelated. This index can effectively reflect the changes in the internal thermal-mechanical state of the equipment. For example, poor contact can cause local overheating and accompanying abnormal vibration, which can assist in judging potential faults in the mechanical structure and heat conduction of the equipment.

[0154] Obtain the partial discharge peak value. Partial discharge is an important sign of faults such as insulation deterioration in the branch box. The peak value can intuitively reflect the discharge intensity. Highlighting this feature can quickly locate the risk of insulation faults and provide a key basis for evaluating the insulation status of the branch box.

[0155] In S3, the process of obtaining the node fault probability matrix of the branch box is as follows:

[0156] Obtain the topological graph structure of the branch box , where V represents the set of nodes in the topological graph structure of the branch box, including the i-th node and the h-th node , and E represents the set of edges in the topological graph structure of the branch box;

[0157] Set the attributes of the i-th node , is the current of the i-th node, is the temperature of the i-th node, is the equivalent resistance of the i-th node;

[0158] The edge weight between the i-th node and the h-th node ;

[0159] In the formula, is the conductor resistance between the i-th node and the h-th node, is the imaginary unit, is the power grid angular frequency, is the line inductance between the i-th node and the h-th node;

[0160] Construct a spatio-temporal graph convolutional layer:

[0161] ;

[0162] In the formula, is the feature matrix of the output l + 1-th layer, is the feature matrix of the l-th layer, is the activation function, A is the angular matrix, D is the normalized adjacency matrix, represents the negative half power of D, is the weight matrix of the l-th layer, obtained through supervised learning training, using historical fault data and normal data as inputs and the node fault probability as the label, and optimized by backpropagation, is the hyperparameter for balancing spatial convolution and temporal convolution, adjusted on the validation set through cross-validation or Bayesian optimization to balance the contributions of spatial and temporal convolutions, is the temporal convolution operation;

[0163] Input the dynamic feature vector of the branch box and the attributes of each node;

[0164] Output node failure probability:

[0165] ;

[0166] Where, is the failure probability of the i-th node, is the final hidden layer feature of the node after spatiotemporal graph convolution processing, is the classifier weight vector, Sigmoid is the activation function;

[0167] Based on the failure probability of each node, output the branch box node failure probability matrix .

[0168] The branch box topology structure and dynamic feature vectors are integrated. The topology structure reflects the connection relationship and layout of the branch boxes, while the dynamic feature vectors include operating status information such as current and temperature. The combination of the two can comprehensively consider the relationship between branch boxes and their own status, allowing for more accurate fault diagnosis. For example, the fault propagation range can be determined based on the topology, and internal anomalies can be identified based on dynamic features.

[0169] The spatiotemporal graph convolution layer effectively integrates spatial (node connectivity) and temporal (operating status changes over time) features through complex operations, such as processing the adjacency matrix. It also introduces nonlinearity using activation functions to better fit complex fault modes, uncover hidden fault characteristics, and improve fault diagnosis accuracy.

[0170] In S4, the process of obtaining the pre-isolated set is:

[0171] Build physically constrained fault propagation models:

[0172] ;

[0173] Where, is the failure probability of the i-th node at time t, express The rate of change with respect to time t, i.e., the speed at which the fault propagates; is the learnable parameter of the topology propagation term, For physical reinforcement, learnable parameters are used. Historical fault propagation data is used in combination with gradient descent optimization to minimize the difference between the predicted propagation path and the actual fault path. is the adjacency matrix weight, is the failure probability of the hth node, To perform the sum operation on all neighbor nodes of the i-th node, is the set of neighbor nodes of the i-th node, for The maximum value of is the topological propagation term, It is a physical enhancement item;

[0174] Solve the propagation range within the future time interval based on the graph diffusion equation :

[0175] ;

[0176] In the formula, is the failure propagation probability vector of the node at time , is the failure propagation probability vector of the node at time t, is the matrix exponential operation, is the Laplacian matrix, and e is the natural constant;

[0177] Perform critical path identification:

[0178] Define the propagation risk index:

[0179] ;

[0180] In the formula, is the propagation risk index of the i-th node, is the failure probability of the g-th node , represents the sum of the failure propagation probabilities of all nodes on the propagation path of the i-th node, and g represents the node index on the propagation path of the i-th node;

[0181] Screen the propagation risk index higher than the risk index threshold stored in the database The corresponding nodes form a pre-isolation set .

[0182] In S4, the process of determining whether to take blocking measures for the nodes in the pre-isolation set is as follows:

[0183] Set the dynamic blocking instruction :

[0184] ;

[0185] In the formula, 1 means that blocking measures need to be taken for this node, and 0 means that no blocking is required;

[0186] Output the dynamic blocking instruction .

[0187] The physically constrained fault propagation model combines the topological propagation term and the physical strengthening term. The topological propagation term considers the influence of the connection relationship between branch boxes on fault propagation based on the adjacency matrix weights and the fault probabilities of adjacent nodes; the physical strengthening term combines the temperature-vibration correlation index and the partial discharge peak value to strengthen the fault propagation analysis from the physical property level. The combination of the two can more accurately describe the fault propagation mechanism, optimize the parameters using historical data, and make the model better fit the actual fault propagation situation.

[0188] Based on the solution of the graph diffusion equation for the future fault propagation range, considering the branch box topological structure and the time-varying fault propagation probability, the diffusion boundary of the fault within a certain time can be predicted in advance, which can gain processing time for the operation and maintenance personnel and make preparations in advance.

[0189] Define the propagation risk index, which integrates the node fault probability and the node fault propagation probability on the propagation path, can comprehensively evaluate the risk degree of the node in fault propagation, screen out the key path nodes with high risk, and help the operation and maintenance personnel focus on the key objects of attention.

[0190] Determine the pre-isolation set according to the propagation risk index and the preset threshold, output the dynamic blocking instruction, and judge whether to take blocking measures for the nodes in the pre-isolation set. This instruction clarifies the nodes that need to be blocked, avoids blind operations, improves the pertinence and effectiveness of fault isolation, reduces the fault influence range, and ensures the stable operation of the power system.

[0191] In S5, the process of outputting the fault type and obtaining the isolation instruction is as follows:

[0192] Based on the node fault probability of the branch box, perform quantum decision optimization, and first perform fault classification:

[0193] ;

[0194] In the formula, is the quantum potential function value corresponding to the y-th type of fault (used to describe the potential energy characteristics of the quantum state), u is the upper limit of the summation term, representing the number of characteristic dimensions participating in the calculation, is the r-th weight parameter related to the y-th type of fault, which is optimized through quantum neural network training with the fault classification accuracy as the objective function, is the dynamic feature vector The r-th eigenvalue in, is the central value of the y-th type of fault in the r-th characteristic dimension. Cluster the dynamic feature vector Fdynamic of each type of fault (such as K-means), and take the cluster center as , represents the quantum phase similarity function (a function used in quantum computing and quantum information processing);

[0195] Then calculate the classification probability:

[0196] ;

[0197] where y is the index of the fault category, Y is the total number of fault categories, represents the sample belonging to the probability of the y-th type of fault, is the temperature parameter;

[0198] Output the fault type of the branch box node;

[0199] Isolation strategy optimization, quantum annealing to solve the optimal switch combination, and use the pre-isolation set as the initial solution space constraint:

[0200] ;

[0201] where is the optimal switch combination, argmin represents finding the parameter value that minimizes the objective function, is the set of all possible switch combination vectors, n is the total number of switches, is the fault classification loss weight coefficient, is the switch operation loss weight coefficient, is the fault classification loss function, is the switch operation loss function;

[0202] Output the optimal switch combination as the isolation instruction.

[0203] The process of using the pre-isolation set as the initial solution space constraint is:

[0204] Modify the loss function to:

[0205] ;

[0206] where is the modified loss function, is the original loss function, which includes the fault classification loss function and the switch operation loss function, is the weight coefficient, is the i-th node in the node set representing the branch box topology structure corresponding switch state, is the pre-isolation set.

[0207] Fault classification is carried out based on quantum potential function calculation, and the weight parameters are trained by a quantum neural network, with the fault classification accuracy as the objective function for optimization. At the same time, the central values of various faults in the feature dimension are determined through clustering, which can more accurately capture the differences of different fault types in multiple feature dimensions, improve the accuracy of fault classification, and accurately judge whether the branch box has faults such as overload, insulation breakage, or poor contact.

[0208] The quantum phase similarity function comprehensively considers the relationship between multiple eigenvalues of the dynamic feature vector and the fault central value, explores the potential correlation between features, rather than looking at each feature in isolation, making the classification more in line with the actual fault feature distribution and enhancing the reliability of the classification results.

[0209] The quantum annealing algorithm is used to solve the optimal switch combination. Utilizing the characteristics of quantum computing to quickly search in the complex solution space, compared with traditional algorithms, it can more efficiently find the switch combination that minimizes the comprehensive loss of fault classification and switch operation loss, shortening the decision-making time and improving the fault isolation efficiency.

[0210] Taking the pre-isolation set as the initial solution space constraint, focusing on high-risk nodes, and at the same time modifying the loss function to force the high-risk nodes to be isolated first, making the isolation strategy more targeted, giving priority to dealing with key nodes, reducing the risk of fault propagation, and ensuring the stable operation of the power system.

[0211] In S6, the process of obtaining the switch action sequence of each branch box node and the reconstructed power supply path is as follows:

[0212] Set the branch box topology reconstruction constraints, including the allowable range of line rated current and voltage;

[0213] Input the isolation instruction, and generate a switch control signal according to the isolation instruction:

[0214] ;

[0215] In the formula, is the action signal of switch d at time t, is the isolation instruction of switch d, is the time point to trigger the switch action;

[0216] When , that is, when the isolation instruction requires the switch to be disconnected and , it reaches the trigger time and starts to execute the action, indicates that the switch is disconnected, otherwise , indicating that the switch maintains its original state or is closed;

[0217] Based on the action signals of the switch at each time point, output the switch action sequence;

[0218] Improve the Dijkstra algorithm to solve the optimal path:

[0219] ;

[0220] Wherein, p is a node or a line segment in the path, is the set of all possible paths, is the current passing through path p, is the resistance of path p, represents the power loss on path p, is the weight coefficient, is the indicator function, whose value is 1 when path p is in the fault area and 0 otherwise;

[0221] Output the reconstructed power supply path .

[0222] Generate a switch control signal according to the isolation instruction, clearly stipulate the trigger condition of the switch action. When the isolation instruction requirement is met and the trigger time is reached, accurately control the switch to disconnect or maintain the state, ensure that the switch operation is precisely matched with the fault isolation requirement, and avoid misoperation.

[0223] Output the switch action sequence, make the operations of multiple switches proceed in a reasonable order, ensure the orderly process of fault isolation, and reduce the risk of secondary faults caused by chaotic operations.

[0224] Obtain the constraints for branch box topology reconstruction, such as the rated current of the line and the allowable voltage range. When reconstructing the power supply path, ensure that the new path meets the operating requirements of electrical equipment, avoid problems such as overcurrent and abnormal voltage in the line caused by reconstruction, and ensure the safe and stable operation of the power system.

[0225] Use the improved Dijkstra algorithm to solve the optimal path, comprehensively consider factors such as path power loss and whether it is in the fault area, which can not only reduce energy consumption but also avoid the fault area, improve the power supply reliability and efficiency, and achieve reasonable resource allocation.

[0226] When calculating each formula, the dimensionality of the parameters involved can be removed to simplify the calculation.

[0227] Embodiment 2:

[0228] Based on the power grid state parameters with real-time feedback, specifically including the load rate, voltage deviation rate, and temperature rise rate, dynamically adjust the optimization weights of the power supply path. Calculate the dynamic change trends of the load rate, voltage deviation rate, and temperature rise rate through a sliding time window, express the dynamic change trends through the slopes of the load rate, voltage deviation rate, and temperature rise rate, obtain the slope-weight factor mapping set of the load rate, voltage deviation rate, and temperature rise rate stored in the database, and determine the weight factors of the matching load rate, voltage deviation rate, and temperature rise rate. Normalize each index to the interval [0, 1], and design a weighted scoring function: Comprehensive score = weight factor of load rate × (1 - load rate) + weight factor of voltage deviation rate × (1 - voltage deviation rate) + weight factor of temperature rise rate × temperature rise rate.

[0229] By constructing a comprehensive evaluation function of load balance degree, voltage stability coefficient, and line margin, automatically balance power supply security and economy during reconstruction, and preferentially select the path with the optimal electrical parameters and redundant fault tolerance capabilities. Calculate the standard deviation of the load rates of each line in the area. The smaller the standard deviation, the more balanced the load distribution. The optimization goal is to minimize the standard deviation to avoid local overload. Monitor the deviation ratio of the node voltage to the rated value, and use the inverse exponential function (e^(-k×deviation)) to quantify the stability. The larger the deviation, the lower the coefficient. Combine the rated current of the line, real-time current, and temperature rise prediction model to calculate the remaining safety margin ((rated current - real-time current) / rated current). The higher the margin, the greater the scoring contribution.

[0230] At the same time, design a dynamic fault tolerance logic. Continuously monitor the health status of key nodes after the switch action is executed. If it is detected that the isolation area spreads or new fault symptoms appear, immediately trigger a secondary reconstruction and link the standby links of adjacent branch boxes to form a grid-shaped self-healing network. After the pre-isolation set is executed, continuously track parameters such as the temperature, vibration, and discharge amount of adjacent nodes. If any parameter exceeds the limit and grows within the set time, it is determined as fault spread, trigger a secondary fault tolerance response, establish periodic communication with adjacent branch boxes, exchange link status in real time, select the available link with the highest score from the standby link pool according to the comprehensive score, preferentially enable the path with low load, stable voltage, and large margin, and complete the timing control of disconnecting the old path and closing the new path through the switch logic table preset in the database to ensure power supply continuity. When the risk cannot be eliminated by local link switching, send a collaborative reconstruction request to the upper-level node, trigger a multi-branch box joint path calculation, form a grid-shaped power supply loop network that bypasses the fault area, and achieve multi-point mutual backup.

[0231] Example 3: As Figure 2 shown, the intelligent monitoring and fault isolation system for branch boxes includes:

[0232] A multi-physical field data fusion acquisition module, which is used to perform multi-physical field data fusion acquisition of the branch box to obtain a spatio-temporal feature matrix;

[0233] A dynamic feature vector fusion module, which is used to perform dynamic feature extraction on the multi-physical field data of the branch box based on the spatio-temporal feature matrix and output the fused dynamic feature vector of the branch box;

[0234] A fault probability matrix acquisition module, which is used to obtain the topological graph structure of the branch box, combine it with the dynamic feature vector of the branch box, and perform spatio-temporal graph convolution network analysis to obtain the branch box node fault probability matrix;

[0235] A dynamic blocking instruction output module, which constructs a fault propagation model with physical constraints based on the branch box node fault probability matrix, identifies the critical path, obtains the pre-isolation set, analyzes the nodes in the pre-isolation set, outputs dynamic blocking instructions, and determines whether to take blocking measures for the nodes in the pre-isolation set;

[0236] An isolation instruction analysis module, which is used to perform quantum decision optimization based on the branch box node fault probability matrix, use the pre-isolation set as the initial solution space constraint, output the fault type and obtain the isolation instruction;

[0237] A power supply path reconstruction module, which is used to set the branch box topology reconstruction constraint, combine the isolation instruction to perform closed-loop control on each branch box node, and obtain the switch action sequence of each branch box node and the reconstructed power supply path.

Claims

1. Branch box intelligent monitoring and fault isolation method, characterized in that It includes the following steps: S1. Perform multi-physical field data fusion acquisition of the branch box to obtain a spatio-temporal feature matrix; S2. Based on the spatio-temporal feature matrix, extract dynamic features from the multi-physical field data of the branch box, and output the fused dynamic feature vector of the branch box; S3. Obtain the topological graph structure of the branch box, combine it with the dynamic feature vector of the branch box, and perform spatio-temporal graph convolutional network analysis to obtain the node fault probability matrix of the branch box; S4. Construct a physically constrained fault propagation model based on the node fault probability matrix of the branch box, identify the critical path, obtain the pre-isolation set, analyze the nodes in the pre-isolation set, output dynamic blocking instructions, and determine whether to take blocking measures for the nodes in the pre-isolation set; S5. Perform quantum decision optimization based on the node fault probability matrix of the branch box, use the pre-isolation set as the initial solution space constraint, output the fault type and obtain the isolation instruction; S6. Set the topological reconstruction constraint of the branch box, and perform closed-loop control on each branch box node in combination with the isolation instruction to obtain the switch action sequence of each branch box node and the reconstructed power supply path; In the above S1, the process of obtaining the spatio-temporal feature matrix is as follows: Perform multi-physical field data fusion acquisition for the branch box, and the acquisition content includes the temperature field of the branch box at time t , the current waveform of the branch box , the vibration signal of the branch box and the partial discharge amount of the branch box ; Construct a three-dimensional data cube X: ； where N is the number of sensor nodes, L is the length of the time window, is the total number of multi-modal data channels, R is the set of real numbers, denotes dimensional real space, Concatenate denotes concatenation, and Stack denotes the stacking operation; Reduce the dimension through tensor decomposition, and use Tucker decomposition to extract the core features: ； where G is the core tensor, is the transformation matrix on the node dimension, is the transformation matrix on the time dimension, is the transformation matrix on the sensor dimension; represents the transformation on the first dimension of the tensor, i.e., the node dimension; represents the transformation on the second dimension of the tensor, i.e., the time dimension; represents the transformation on the third dimension of the tensor, i.e., the sensor dimension; Obtain a compact spatio-temporal feature matrix: ； In the formula, is the spatio-temporal feature matrix, is the total number of updated multi-modal data channels, denotes a real number space of dimension represents the mode-1 unfolding matrix of the core tensor G.

2. The intelligent monitoring and fault isolation method for branch boxes according to claim 1, characterized in that: In the above S2, the process of outputting the fused dynamic feature vector of the branch box is as follows: Input spatio-temporal feature matrix and the original high-frequency current signal ; Perform transient current analysis: Pair Perform 8-layer wavelet packet decomposition to obtain 256 sub-bands: ； In the formula, is the result of wavelet packet decomposition, is the coefficient of the k-th subband after decomposition, is the wavelet packet basis function of the k-th subband; Calculate the energy entropy of each subband : ； Select the 3 sub-bands with the largest energy entropy and denote them as , , . Their coefficients , , constitute the transient feature : ； Perform temperature-vibration coupling analysis and calculate the correlation index between the temperature gradient and the vibration energy : ； In the formula, is the gradient of temperature T in the x direction of space, is the calculation of the energy of the vibration signal in the time interval , represents the time interval; Obtain the partial discharge peak value of the branch box ; Output the dynamically characteristic vectors of the fused branch boxes : ； Among them, represents vector transpose.

3. The intelligent monitoring and fault isolation method for the branch box according to claim 2, characterized in that: In the above S3, the process of obtaining the node fault probability matrix of the branch box is as follows: Obtain the topological graph structure of the branch box , where V represents the set of nodes in the topological graph structure of the branch box, including the i-th node and the h-th node , and E represents the set of edges in the topological graph structure of the branch box; Set the attributes of the i-th node , is the current of the i-th node, is the temperature of the i-th node, is the equivalent resistance of the i-th node; The edge weight between the i-th node and the h-th node ; wherein, is the conductor resistance between the i-th node and the h-th node, is the imaginary unit, is the grid angular frequency, is the line inductance between the i-th node and the h-th node; Construct a spatio-temporal graph convolutional layer: ； wherein, is the feature matrix of the (l + 1)-th layer of the output, is the feature matrix of the l-th layer, is the activation function, A is the angular matrix, D is the normalized adjacency matrix, represents the negative half power of D, is the weight matrix of the l-th layer, is the hyperparameter for balancing spatial convolution and temporal convolution, is the temporal convolution operation; Input the dynamic feature vector of the branch box and the attributes of each node; Output the node fault probability: ； wherein, is the failure probability of the i-th node, is the final hidden layer feature of the node after spatio-temporal graph convolution processing, is the classifier weight vector, and Sigmoid is the activation function; Output the fault probability matrix of the branch box nodes based on the fault probabilities of each node .

4. The intelligent monitoring and fault isolation method for branch boxes according to claim 3, characterized in that: In the above S4, the process of obtaining the pre-isolation set is as follows: Construct a physically constrained fault propagation model: ； Wherein, is the failure probability of the i-th node at time t, represents the rate of change with respect to time t, that is, the speed of fault propagation; is the learnable parameter of the topological propagation term, is the learnable parameter of the physical reinforcement term, is the weight of the adjacency matrix, is the failure probability of the h-th node, is the summation operation over all neighbor nodes of the i-th node, is the set of neighbor nodes of the i-th node, is the maximum value of, is the topological propagation term, is the physical reinforcement term; Solving for the propagation range within a future time interval based on the graph diffusion equation inside: ； Wherein, is the fault propagation probability vector of the node at time , is the fault propagation probability vector of the node at time t, is the matrix exponential operation, is the Laplacian matrix, and e is the natural constant; Perform critical path identification: Define the propagation risk index: ； In the formula, is the propagation risk index of the i-th node, is the failure probability of the g-th node , represents the summation of the failure propagation probabilities of all nodes on the propagation path of the i-th node, and g represents the node index on the propagation path of the i-th node; Screening transmission risk index Higher than the risk index threshold stored in the database The corresponding nodes form a pre-isolation set .

5. The intelligent monitoring and fault isolation method for branch boxes according to claim 4, characterized in that: In the above S4, the process of determining whether to take blocking measures for the nodes in the pre-isolation set is as follows: Set dynamic blocking instruction : ； In the formula, 1 indicates that blocking measures need to be taken for this node, and 0 indicates that no blocking is required; Output dynamic blocking instruction .

6. The intelligent monitoring and fault isolation method for branch boxes according to claim 1, characterized in that: In the above S5, the process of outputting the fault type and obtaining the isolation instruction is as follows: Perform quantum decision optimization based on the node fault probability of the branch box, and first perform fault classification: ； In the formula, is the quantum potential function value corresponding to the y-th type of fault, u is the upper limit of the number of terms for summation, representing the number of characteristic dimensions participating in the calculation, is the r-th weight parameter related to the y-th type of fault, is the dynamic feature vector is the r-th eigenvalue in, is the central value of the y-th type of fault on the r-th characteristic dimension, represents the quantum phase similarity function; Then calculate the classification probability: ； where y is the index of the fault category and Y is the total number of fault categories, denotes the sample and the probability that it belongs to the y-th type of fault, and θ is the temperature parameter; Output the fault type of the branch box node; Optimize the isolation strategy, use quantum annealing to solve the optimal switch combination, and use the pre-isolation set as the initial solution space constraint: ； Wherein, is the optimal switch combination, and argmin represents finding the parameter value that minimizes the objective function, is the set of all possible switch combination vectors, and n is the total number of switches, is the fault classification loss weight coefficient, is the switch operation loss weight coefficient, is the fault classification loss function, is the switch operation loss function; Output the optimal switch combination as an isolation instruction.

7. The intelligent monitoring and fault isolation method for branch boxes according to claim 6, characterized in that: The process of using the pre-isolation set as the initial solution space constraint is as follows: Modify the loss function to: ； In the formula, is the modified loss function, is the original loss function, which includes the fault classification loss function and the switch operation loss function, is the weight coefficient, is the \(i\)-th node in the node set representing the branch box topology structure corresponding switch state, is the pre-isolation set.

8. The intelligent monitoring and fault isolation method for branch boxes according to claim 1, characterized in that: In the above S6, the process of obtaining the switch action sequence of each branch box node and the reconstructed power supply path is as follows: Set the topological reconstruction constraint of the branch box, including the rated current of the line and the allowable voltage range; Input the isolation instruction, and generate a switch control signal according to the isolation instruction: ； wherein, is the action signal of switch d at time t, is the isolation instruction of switch d, is the time point for triggering the switch action; When it is required that the switch is disconnected in the isolation instruction and when it reaches the trigger time, the action starts to be executed, indicating that the switch is disconnected, otherwise it indicates that the switch maintains its original state or is closed; Based on the action signals of the switch at each time point, output the switch action sequence; Improve the Dijkstra algorithm to solve the optimal path: ； Wherein, p is a node or line segment in the path, is the set of all possible paths, is the current passing through path p, is the resistance of path p, represents the power loss on path p, is the weight coefficient, is the indicator function, which has a value of 1 when path p is in the fault area and 0 otherwise; Output the reconstructed power supply path .

9. The intelligent monitoring and fault isolation system for branch boxes is used to implement the method described in any one of claims 1-8, and is characterized in that It includes: A multi-physical field data fusion acquisition module, which is used to perform multi-physical field data fusion acquisition of the branch box to obtain a spatio-temporal feature matrix; A dynamic feature vector fusion module, which is used to extract dynamic features from the multi-physical field data of the branch box based on the spatio-temporal feature matrix, and output the fused dynamic feature vector of the branch box; A fault probability matrix acquisition module, which is used to obtain the topological graph structure of the branch box, combine the dynamic feature vectors of the branch box, perform spatio-temporal graph convolution network analysis, and obtain the branch box node fault probability matrix; A dynamic blocking instruction output module, which constructs a fault propagation model with physical constraints based on the branch box node fault probability matrix, identifies the critical path, obtains the pre-isolation set, analyzes the nodes in the pre-isolation set, outputs dynamic blocking instructions, and determines whether to take blocking measures for the nodes in the pre-isolation set; An isolation instruction analysis module, which is used to perform quantum decision optimization based on the branch box node fault probability matrix, use the pre-isolation set as the initial solution space constraint, output the fault type and obtain the isolation instruction; A power supply path reconstruction module, which is used to set the branch box topology reconstruction constraint, combine the isolation instruction to perform closed-loop control on each branch box node, and obtain the switch action sequence of each branch box node and the reconstructed power supply path.

Images