CN120337157A - Distribution network fault diagnosis method, device, terminal and medium based on large model - Google Patents
Distribution network fault diagnosis method, device, terminal and medium based on large modelInfo
- Publication number
- CN120337157A CN120337157A CN202510764101.4A CN202510764101A CN120337157A CN 120337157 A CN120337157 A CN 120337157A CN 202510764101 A CN202510764101 A CN 202510764101A CN 120337157 A CN120337157 A CN 120337157A
- Authority
- CN
- China
- Prior art keywords
- distribution network
- modal
- data
- fault diagnosis
- matrix
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/25—Fusion techniques
- G06F18/253—Fusion techniques of extracted features
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
- G01N33/00—Investigating or analysing materials by specific methods not covered by groups G01N1/00 - G01N31/00
- G01N33/0004—Gaseous mixtures, e.g. polluted air
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01R—MEASURING ELECTRIC VARIABLES; MEASURING MAGNETIC VARIABLES
- G01R31/00—Arrangements for testing electric properties; Arrangements for locating electric faults; Arrangements for electrical testing characterised by what is being tested not provided for elsewhere
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01W—METEOROLOGY
- G01W1/00—Meteorology
- G01W1/02—Instruments for indicating weather conditions by measuring two or more variables, e.g. humidity, pressure, temperature, cloud cover or wind speed
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/10—Pre-processing; Data cleansing
- G06F18/15—Statistical pre-processing, e.g. techniques for normalisation or restoring missing data
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/042—Knowledge-based neural networks; Logical representations of neural networks
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/044—Recurrent networks, e.g. Hopfield networks
- G06N3/0442—Recurrent networks, e.g. Hopfield networks characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU]
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Computing Systems (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Molecular Biology (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Chemical & Material Sciences (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Environmental & Geological Engineering (AREA)
- Biochemistry (AREA)
- Probability & Statistics with Applications (AREA)
- Environmental Sciences (AREA)
- Ecology (AREA)
- Biodiversity & Conservation Biology (AREA)
- Atmospheric Sciences (AREA)
- Pathology (AREA)
- Immunology (AREA)
- Analytical Chemistry (AREA)
- Medicinal Chemistry (AREA)
- Food Science & Technology (AREA)
- Combustion & Propulsion (AREA)
- Supply And Distribution Of Alternating Current (AREA)
Abstract
The application discloses a power distribution network fault diagnosis method, device, terminal and medium based on a large model, which relates to the technical field of power distribution networks, the scheme provided by the application is based on the acquired operating data such as topology, electric quantity and operating environment of the power distribution network, the real-time topology dynamic change of the power distribution network is reflected by constructing a dynamic graph attention network model, and the cross-modal attention mechanism is used for fusing the multi-modal data characteristics, so that the interference of the multi-modal data distribution difference on the characteristic extraction is eliminated, the robustness of the model to fault diagnosis operation under the complex fault mode is enhanced, the problems of fault misjudgment and missed judgment caused by insufficient fusion of the static model and the multi-modal data in the traditional method are solved, and the fault diagnosis accuracy of the power distribution network is improved.
Description
Technical Field
The application relates to the technical field of power distribution networks, in particular to a power distribution network fault diagnosis method, device, terminal and medium based on a large model.
Background
Along with the large-scale access of high-proportion renewable energy sources and novel power loads, the operation of a modern power distribution network presents strong uncertainty and multi-time space coupling characteristics, and higher requirements are put forward on the instantaneity and the accuracy of fault diagnosis.
The current power distribution network fault diagnosis technology based on a large model is mostly based on fixed topology and manual rule reasoning, and the mode is effective in a simple scene, but depends on a static model and a preset threshold value, so that the problem that the adaptation is difficult to complex changes such as frequent access of distributed new energy, dynamic reconstruction of network topology and the like is easy to cause fault positioning deviation due to topology mismatch, and therefore fault diagnosis and positioning are easy to cause missed judgment and misjudgment, and the technical problem of low accuracy is caused.
Disclosure of Invention
The application provides a power distribution network fault diagnosis method, device, terminal and medium based on a large model, which are used for solving the technical problem of low accuracy of the existing power distribution network fault diagnosis technology.
In order to solve the technical problems, a first aspect of the present application provides a power distribution network fault diagnosis method based on a large model, including:
acquiring power distribution network operation data, wherein the power distribution network operation data comprises topology data, electrical quantity data and environment data;
constructing a dynamic graph annotation network model according to topology data and electric quantity data in the power distribution network operation data, wherein the dynamic graph annotation network model is used for reflecting the real-time topology dynamic change relation of the power distribution network;
And carrying out feature fusion processing on the power distribution network operation data of different modes through a cross-mode attention mechanism based on a preset multi-mode large model to obtain mode fusion features, so as to obtain a fault diagnosis result of the power distribution network based on the mode fusion features.
Preferably, the constructing a dynamic graph annotation network model according to the operation data of the power distribution network includes:
Constructing a node characteristic matrix according to the power distribution network operation data, wherein the node characteristic matrix is used for reflecting node characteristics of each power distribution network node in different time steps;
Constructing the dynamic adjacency matrix according to the node characteristic matrix, wherein the dynamic adjacency matrix is used for reflecting the electric coupling strength between nodes of different power distribution networks at different times;
And constructing a dynamic graph annotation network model according to the dynamic adjacency matrix.
Preferably, the method further comprises:
Based on a selected target node, calculating an attention coefficient between the target node and a neighbor node under the current time step by a preset space attention coefficient calculation formula;
Performing feature aggregation on node features and attention coefficients of each neighbor node through an activation function to obtain updated node features of the target node at the next time step;
Summarizing node characteristics from the target node to the current time step, and extracting time-dependent characteristics through an LSTM network to obtain a time characteristic sequence of the target node;
And obtaining the latest node characteristics of the target node according to the weighted sum of the updated node characteristics and the time sequence characteristics.
Preferably, the feature fusion processing is performed on the power distribution network operation data of different modes through a cross-mode attention mechanism, and the obtaining of the mode fusion features includes:
according to the operation data of the power distribution network, performing feature coding processing through a plurality of preset modal encoders to respectively obtain an electric quantity modal matrix, a topology modal matrix and an environment modal matrix;
According to the electric quantity modal matrix, the topology modal matrix and the environment modal matrix, the electric quantity modal matrix is used as a modal alignment reference, and QKV characteristics are obtained through a preset QKV mapping calculation formula;
and according to the QKV characteristics, combining an attention weight calculation formula to obtain attention weights, and obtaining the mode fusion characteristics based on the sum of the attention weights and the electric quantity mode matrix.
Preferably, the QKV mapping calculation formula is specifically:
In the formula, For the electrical quantity modal matrix,For a combined matrix of the topological modal matrix and the environmental modal matrix,、、The weight matrix of the Q feature, the K feature and the V feature respectively.
Preferably, the method further comprises:
Based on a preset multi-mode data sample of the power distribution network, combining a loss function based on structural causal constraint, carrying out model training on the multi-mode large model, determining the convergence degree of the multi-mode large model based on the output of the loss function, and outputting the multi-mode large model when the convergence degree reaches a preset convergence threshold or the training iteration number reaches a preset number threshold.
Preferably, the acquiring operation data of the power distribution network further includes:
and preprocessing the operation data of the power distribution network, wherein the preprocessing comprises data cleaning, normalization and denoising.
Meanwhile, a second aspect of the present application provides a power distribution network fault diagnosis device based on a large model, including:
The operation data acquisition unit is used for acquiring operation data of the power distribution network, wherein the operation data of the power distribution network comprises topology data, electric quantity data and environment data;
The dynamic graph network construction unit is used for constructing a dynamic graph annotation network model according to topology data and electric quantity data in the power distribution network operation data, wherein the dynamic graph annotation network model is used for reflecting the real-time topology dynamic change relation of the power distribution network;
The power distribution network fault diagnosis unit is used for carrying out feature fusion processing on power distribution network operation data of different modes through a cross-mode attention mechanism based on a preset multi-mode large model to obtain mode fusion features, so that a fault diagnosis result of the power distribution network is obtained based on the mode fusion features.
The third aspect of the application provides a power distribution network fault diagnosis terminal based on a large model, which comprises a memory and a processor;
The memory is used for storing program codes, and the program codes are used for realizing the power distribution network fault diagnosis method based on the large model, which is provided by the first aspect of the application;
the processor is configured to read and execute the program code.
A fourth aspect of the present application provides a computer readable storage medium having program code stored therein for being read and executed by a processor to implement a method for fault diagnosis of a distribution network based on a large model as provided in the first aspect of the present application.
From the above technical scheme, the application has the following advantages:
The scheme provided by the application is based on the acquired operation data such as the topology, the electrical quantity and the operation environment of the power distribution network, the real-time topology dynamic change of the power distribution network is reflected by constructing a dynamic graph attention network model, the multi-mode data characteristics are fused by utilizing a cross-mode attention mechanism, the interference of multi-mode data distribution difference on the characteristic extraction is eliminated, the robustness of the model to the fault diagnosis operation under a complex fault mode is enhanced, the problems of fault misjudgment and missed judgment caused by the insufficient fusion of the static model and the multi-mode data in the traditional method are solved, and the fault diagnosis accuracy of the power distribution network is improved.
Drawings
In order to more clearly illustrate the embodiments of the application or the technical solutions of the prior art, the drawings which are used in the description of the embodiments or the prior art will be briefly described, it being obvious that the drawings in the description below are only some embodiments of the application, and that other drawings can be obtained from these drawings without inventive faculty for a person skilled in the art.
Fig. 1 is a schematic flow chart of an embodiment of a power distribution network fault diagnosis method based on a large model.
Fig. 2 is a schematic structural diagram of an embodiment of a power distribution network fault diagnosis device based on a large model.
Fig. 3 is a schematic structural diagram of an embodiment of a power distribution network fault diagnosis terminal based on a large model.
Detailed Description
The embodiment of the application provides a power distribution network fault diagnosis method, device, terminal and medium based on a large model, which are used for solving the technical problem of low accuracy of the existing power distribution network fault diagnosis technology.
In order to make the objects, features and advantages of the present application more comprehensible, the technical solutions in the embodiments of the present application are described in detail below with reference to the accompanying drawings, and it is apparent that the embodiments described below are only some embodiments of the present application, but not all embodiments of the present application. All other embodiments, which can be made by those skilled in the art based on the embodiments of the application without making any inventive effort, are intended to be within the scope of the application.
Firstly, the application provides a power distribution network fault diagnosis method based on a large model, which is concretely described as follows:
referring to fig. 1, the power distribution network fault diagnosis method based on the large model provided in this embodiment includes:
step 101, acquiring operation data of a power distribution network;
the power distribution network operation data comprise topology data, electrical quantity data and environment data;
102, constructing a dynamic graph annotation network model according to topology data and electric quantity data in the operation data of the power distribution network;
The dynamic graph notice network model is used for reflecting the real-time topological dynamic change relation of the power distribution network;
Step 103, performing feature fusion processing on the power distribution network operation data of different modes through a cross-mode attention mechanism based on a preset multi-mode large model to obtain mode fusion features so as to obtain a fault diagnosis result of the power distribution network based on the mode fusion features.
It should be noted that the scheme of the application provides that the operation data of the power distribution network comprising topology data, electric quantity data and environment data is obtained, a dynamic graph attention network model is constructed to reflect real-time topology dynamic change, a multi-mode large model is utilized to obtain a mode fusion characteristic by fusing different mode data through a cross-mode attention mechanism, and fault diagnosis is carried out based on the characteristic. The dynamic graph annotation network model describes a graph structure of electrical coupling strength between nodes changing along with time through a dynamic adjacency matrix, and can be specifically realized by adjusting the connection weight between nodes through a real-time updated attention coefficient, so as to capture the influence of topology dynamic reconstruction on a fault propagation path. The cross-modal attention mechanism is to calculate the query and key value characteristics of different modal data through mapping to determine the association weight, and particularly, the multi-head attention layer can be adopted to align potential association of electric quantity, topology and environmental modes for eliminating the multi-source data distribution difference. The modal fusion feature refers to multi-modal joint characterization after weighted aggregation, and specifically, residual connection can be adopted to retain original feature information for enhancing the characterization capability of fault features.
Specifically, the operation data of the power distribution network is preprocessed and then input into a dynamic graph meaning network, the network constructs a dynamic adjacency matrix according to characteristics such as node voltage and current, neighbor node information is aggregated through a spatial attention coefficient, and a node state evolution rule is extracted through time sequence analysis. The multi-mode large model extracts deep features of all modes through an independent encoder, generates cross-mode attention weight by taking an electric quantity mode as a reference, and performs weighted fusion on topology and environmental features. The fused characteristic input classifier outputs fault type and position information.
Compared with the prior art, the method and the device have the advantages that the node connection relation is adjusted in real time through the dynamic graph structure, the method and the device adapt to the dynamic network reconstruction scene of the power distribution network, the node coupling relation change under the dynamic topology can be accurately captured, the interference of the multi-mode data distribution difference on feature extraction is eliminated, the diagnosis robustness under the complex fault mode is enhanced, the misjudgment risk caused by single data dimension and topology mismatch is effectively reduced, meanwhile, the inter-data semantic association is established through the cross-mode attention mechanism, and the feature fusion effectiveness is improved.
Based on the above basic embodiment, further, attention is paid to the network model for the dynamic graph in step 102, and the construction steps are specifically as follows:
constructing a node characteristic matrix according to the operation data of the power distribution network;
constructing a dynamic adjacency matrix according to the node characteristic matrix;
and constructing a dynamic graph annotation network model according to the dynamic adjacency matrix.
It should be noted that, this embodiment further proposes to construct a node feature matrix according to the operation data of the power distribution network, construct a dynamic adjacency matrix according to the node feature matrix, and construct a dynamic graph attention network model according to the dynamic adjacency matrix. The node characteristic matrix is used for reflecting node characteristic data of each power distribution network node at different time steps, and the dynamic adjacency matrix is used for reflecting electric coupling strength among different power distribution network nodes at different times.
The node characteristic matrix is a multidimensional data set formed by electrical parameters such as voltage, current, power and time step information, and can be specifically realized by periodically sampling and characteristic recombination of electrical quantity data acquired in real time by adopting a sliding time window, and is used for representing a dynamic behavior mode of a power distribution network node in a continuous time sequence, and the expression of the node characteristic matrix can be as follows:
wherein, the Is the number of nodes of the power distribution network; is a characteristic dimension such as voltage, current, power, etc.; Is the time step.
The dynamic adjacency matrix is a time-varying weight matrix generated based on the change rate of the electrical quantity between nodes and the topological connection relation, and can be specifically realized by calculating the electrical coupling degree difference between adjacent nodes by adopting a dynamic time warping algorithm, and is used for capturing the real-time fluctuation of the association strength between nodes in the dynamic reconstruction process of the network topology, wherein the matrix expression and the calculation formula of each element in the matrix are as follows:
wherein, the The real-time electric coupling strength between the nodes can be reflected along with the change of time and running state.
Dynamic adjacency matrix calculation formula:
In the formula, Representing nodes,The higher the power flow, the higher the association.
Specifically, after the operation data of the power distribution network are obtained, firstly, slicing the electrical quantity parameters such as voltage, current, power and the like of each node according to a preset time step to form a three-dimensional tensor structure comprising the node number, the characteristic dimension and the time step, and taking the three-dimensional tensor structure as a node characteristic matrix. And then, calculating the electric coupling strength of different time segments according to the electric quantity change trend of adjacent nodes in the node characteristic matrix, and generating a dynamic adjacent matrix by combining the topological connection relation, wherein the matrix element weight is dynamically adjusted along with the electric quantity association degree. And finally, inputting the dynamic adjacency matrix and the node characteristic matrix into a graph attention network frame, realizing dynamic aggregation and updating of node characteristics through a graph convolution layer and an attention mechanism, and constructing a dynamic graph attention network model capable of adapting to the real-time change of topology.
Compared with the prior art, the method and the device have the advantages that the electric coupling strength among the nodes is updated in real time through the dynamic adjacency matrix, so that the model can be used for capturing the dynamic change characteristics of the topology in a self-adaptive mode, and fault misjudgment caused by topology mismatch is avoided.
It can be understood that, in order to further improve the node feature characterization capability of the present solution and reduce the risk of misjudgment caused by the change of the network structure, the update logic example of the node feature matrix provided in this embodiment is specifically as follows:
Based on the selected target node, calculating the attention coefficient between the target node and the neighbor node under the current time step by a preset space attention coefficient calculation formula;
Performing feature aggregation on node features and attention coefficients of each neighbor node through an activation function to obtain updated node features of the target node at the next time step;
Summarizing node characteristics from a target node to a current time step, and extracting time-dependent characteristics through an LSTM network to obtain a time characteristic sequence of the target node;
And obtaining the latest node characteristics of the target node according to the weighted sum of the updated node characteristics and the time sequence characteristics.
It is noted that it is further proposed that the attention coefficient between the target node and the neighboring nodes is calculated according to a preset space attention coefficient calculation formula based on the selected target node, the node characteristics and the attention coefficients of each neighboring node are subjected to characteristic aggregation through an activation function to obtain updated node characteristics of the target node at the next time step, the node characteristics of the target node up to the current time step are summarized, time-dependent characteristic extraction is performed through an LSTM network to obtain a time characteristic sequence of the target node, and the latest node characteristics of the target node are obtained according to the weighted sum of the updated node characteristics and the time sequence characteristics.
The spatial attention coefficient refers to a degree of association between a dynamic calculation target node and a neighboring node, specifically, for a node i, j, an attention coefficient at a time t may be specifically implemented by calculating cosine similarity or dot product similarity after linear transformation, and is used for capturing coupling strength between nodes under dynamic topology change, where the embodiment refers to the spatial attention coefficientThe calculation formula of (2) is as follows:
In the formula, Represents the hidden state of node i at time t,Representing the hidden state of node j at time t,In order for the matrix of parameters to be learnable,For vector concatenation operation LeakyReLU (LEAKY RECTIFIED LINEAR Unit, leakage correction linear Unit) is an improved activation function, mainly used for solving the defect that the gradient of the traditional ReLU (RECTIFIED LINEAR Unit) is zero in the negative interval.
The activation function refers to a function of nonlinear conversion, such as a ReLU or Sigmoid function, and can be specifically used to perform nonlinear mapping on the aggregated features, so as to enhance the model expression capability. Feature aggregation refers to weighted superposition of features of neighbor nodes according to attention coefficients, and can be realized by adopting a weighted summation or splicing mode, and is used for fusing spatial association information among multiple nodes.
The LSTM network refers to a long-term and short-term memory cyclic neural network, and can be specifically realized through a hidden state transmission and time gating mechanism and is used for extracting the dependency relationship of node characteristics in the time dimension. The time sequence features refer to feature change trend of the target node in the historical time step, and particularly, the time sequence segments can be intercepted through a sliding window for encoding so as to reflect the dynamic evolution rule of the node state.
Specifically, for a selected target node, the strength of electrical coupling between the node and its neighboring nodes is first quantified using a spatial attention coefficient calculation based on the node feature matrix for the current time step. For example, abrupt current changes or voltage fluctuations of neighboring nodes can dynamically affect the feature update of the target node through the attention coefficient. Then, multiplying the characteristics of the neighbor nodes with the corresponding attention coefficients, accumulating, and performing nonlinear transformation through an activation function to generate preliminary updated characteristics of the target node at the next moment, wherein the expression is as follows:
represents the hidden state of node i at time t +1, For the neighbor set of node i, byThe definition of the term "a" or "an" is,In order for the matrix of parameters to be learnable,To activate the function.
Meanwhile, the characteristic sequences of the target node at the current time step and the historical time step are input into the LSTM network, and the memory unit is utilized to capture the fault propagation mode on the time sequence, wherein the expression is as follows:
Is a sequence of cell states, which is defined by the sequence of cell states, Is a sequence of hidden states.
Finally, the updated features after spatial aggregation and the time features extracted by the LSTM are subjected to weighted fusion, for example, higher weight is given to the time features to reflect causal chronology of fault propagation, so that the latest multidimensional feature representation of the target node is generated, and the expression is as follows:
are learnable parameters for balancing the importance of spatiotemporal features.
According to the scheme, through a dynamic spatial attention mechanism, the association weights among the nodes are adjusted in real time, and meanwhile, the LSTM network mining depth time sequence mode is combined, so that the node characteristics can be fused with spatial dynamic association and time causal evolution information at the same time, model parameters are dynamically adjusted according to the fault severity, and the attention to a key area is promoted:
wherein, the Representing the fault confidence of node i,In order to adjust the coefficient of the coefficient,In order to adjust the associated weights before they are adjusted,And (5) the adjusted association weight.
Through the technical scheme, the characterization capability of the node characteristics can be improved, so that the fault diagnosis model can capture the abnormal propagation path under the dynamic topology more accurately, the misjudgment risk caused by the change of the network structure is reduced, the extraction capability of the multi-time-scale fault characteristics is enhanced, and the reliability of the diagnosis result is remarkably improved.
Further, the embodiment also provides that feature fusion processing is carried out on the power distribution network operation data of different modes through a cross-mode attention mechanism, and the obtained mode fusion features comprise the steps of carrying out feature encoding processing through a plurality of preset mode encoders according to the power distribution network operation data to respectively obtain an electric quantity mode matrix, a topology mode matrix and an environment mode matrix, obtaining QKV features through a preset QKV mapping calculation formula according to the electric quantity mode matrix, the topology mode matrix and the environment mode matrix serving as mode alignment references, obtaining attention weights according to QKV features and combining attention weight calculation formulas, and obtaining the mode fusion features based on the sum of the attention weights and the electric quantity mode matrix.
The modal encoder refers to a neural network module designed for different data modes, and can be specifically realized by adopting a convolutional neural network or a full-connection layer, and is used for mapping heterogeneous data such as original electric quantity, topological structure, environmental parameters and the like to a unified feature space. The mode encoder related to the embodiment mainly comprises an electric quantity encoder and an input(Time-series electric quantity), output electric quantity modal matrix。
Topology encoder, inputs DGAT (DYNAMIC GRAPH Attention Network ) generated dynamic graph (adjacency matrix)) Output topology modal matrix。
Wherein, the And the geographic data of each node of the power distribution network.
Environmental encoder for inputting environmental image characteristicsMeteorological characteristicsOutput environment modal matrix。
QKV (Query-Key-Value) mapping calculation formula refers to a mathematical expression of Query, key and Value feature vectors generated by linear transformation by taking an electric quantity modal matrix as a reference, and specifically, input data can be projected by using a trainable weight matrix to realize cross-modal feature alignment. The attention weight calculation formula is that through calculating the similarity between the query characteristics and the key characteristics, the association strength coefficients among different modes are generated, and a dot product attention mechanism can be specifically selected, so that the contribution degree of each mode to fault diagnosis is dynamically distributed.
The map calculation formula QKV mentioned in this embodiment specifically includes:
In the formula, In the form of an electrical quantity modal matrix,Is a combined matrix of a topological modal matrix and an environmental modal matrix,、、The weight matrix of the Q feature, the K feature and the V feature respectively.
Specifically, the electrical quantity modal matrix comprises time sequence characteristics of real-time measurement data such as voltage and current, the topological modal matrix reflects dynamic connection relation among network nodes, and the environmental modal matrix integrates external influence factors such as temperature and humidity. And performing QKV mapping by taking the electric quantity mode as a reference, and generating the attention weight reflecting the cross-mode correlation by calculating the matching degree of the key characteristics corresponding to the topology and the environment mode and the reference query characteristics. This weight is used to adjust the fusion ratio of the different modality features, for example in thunderstorm weather, the humidity features of the environmental modality may be given a higher weight. And finally, superposing the weighted multi-modal characteristics and the reference electrical quantity characteristics to form a fusion characteristic vector containing cross-modal associated information, and providing more complete input information for a downstream fault classifier. According to the scheme, a cross-modal attention mechanism is introduced, a dynamic association model between modalities is built in a feature space, so that external interference such as changes of the air condition, topological reconstruction events and the like can be quantitatively evaluated and integrated into a diagnosis decision, the problem of feature conflict caused by static fusion rules is effectively solved, through the technical scheme, efficient collaboration of multi-modal data in a feature layer is realized, and the problem of fault misjudgment caused by single data dimension in a traditional method is solved. For example, in a distributed power supply frequent switching scene, normal fluctuation and real fault signals caused by topology change can be accurately distinguished by dynamically adjusting the association weight of the topology mode and the electrical quantity mode, the false alarm probability caused by network reconstruction is reduced, and the diagnosis robustness under complex working conditions is improved.
Further, the solution of this embodiment may further include the following steps:
Based on a preset multi-mode data sample of the power distribution network, combining a loss function based on structural causal constraint, carrying out model training on the multi-mode large model to determine the convergence degree of the multi-mode large model based on the output of the loss function, and outputting the multi-mode large model when the convergence degree reaches a preset convergence threshold or the training iteration number reaches a preset number threshold.
The loss function of the causal constraint of the structure refers to causal rationality of constraint model output by introducing priori knowledge of power grid fault propagation logic, and the causal rationality can be realized by combining causal inference theory and regularized loss terms, for example, causal graph structure constraint terms are added in the loss function, so that model parameter updating directions are limited. The multi-mode data sample refers to a power distribution network operation data set containing multi-dimensional information such as electric quantity, topology, environment and the like, and can be specifically constructed by combining historical fault records, real-time monitoring data and meteorological environment data, for example, voltage, current and power data are aligned with temperature, humidity and wind speed data according to time stamps. The convergence threshold refers to an error threshold for judging whether to terminate iteration in the model training process, and specifically may use the failure diagnosis accuracy rate or the loss value change rate on the verification set as a judgment basis, for example, when the accuracy rate of the verification set fluctuates by less than 1% in three continuous iterations.
Specifically, the training process firstly extracts batch data from the multi-mode data sample and inputs the batch data into the multi-mode large model, and after the model outputs a fault diagnosis result, the difference between a prediction result and a real label is calculated by combining a loss function of the causal constraint of the structure. The loss function consists of a data fitting term and a causal constraint term, wherein the data fitting term is used for measuring a prediction error, and the causal constraint term is used for constructing a causal graph according to a power grid fault propagation path and constraining characteristic weight distribution of a model. For example, the causal constraint term can adopt a graph regularization method based on an adjacency matrix, so that the fault association strength among nodes is ensured to accord with the actual physical rule. In the training process, the output of the loss function is used for back propagation to update model parameters, and when the performance index of the model on the verification set reaches a preset threshold or the iteration number reaches an upper limit, the training is terminated and the optimized multi-mode large model is output.
According to the scheme, the prior knowledge such as the topological connection relation of the power grid, the fault propagation path and the like is encoded to the loss function by introducing the structural causal constraint, so that the model training process is not only concerned with the prediction precision, but also forcedly learns the characteristic association conforming to the actual causal logic, and the rationality and the robustness of the diagnosis result are improved.
Further, after step 101, before step 102, the method may further include:
the data is preprocessed, and the preprocessing comprises data cleaning, normalization and denoising.
The data cleaning is to correct or remove abnormal values or missing values in the operation data of the power distribution network, and specifically, the method can be realized by filling missing data segments by adopting an interpolation method or identifying and filtering abnormal fluctuation data based on a sliding window statistical threshold value, and the step can eliminate the interference of sensor acquisition errors on the input of a subsequent model.
The normalization is to convert electrical measurement data of different dimensions into a unified numerical range, and specifically, the normalization can be realized by mapping voltage and current characteristics into a distribution space with a mean value of 0 and a variance of 1 by adopting a Z-score normalization method, and the step eliminates the magnitude difference among multi-mode data to improve the model convergence efficiency. The denoising is to restrain high-frequency interference signals in the operation data of the power distribution network, and specifically can be realized by removing high-frequency noise components after carrying out multi-scale decomposition on current waveforms by wavelet transformation, and the step is helpful for keeping effective frequency band information of fault characteristics.
Specifically, the operation data of the power distribution network is easily influenced by factors such as sensor drift, communication interference and the like in the acquisition process, so that noise, dimension difference and incompleteness of the data exist. The method comprises the steps of cleaning original data, repairing missing values caused by communication interruption, eliminating abnormal mutation points caused by equipment faults, converting electric measurement data with different dimensions such as voltage, power and the like into standardized feature vectors through normalization processing, facilitating collaborative training of a follow-up dynamic graph annotation network and a multi-mode large model, carrying out time-frequency decomposition on current waveforms through wavelet denoising, filtering high-frequency noise interference, retaining low-frequency components of fault features, and accordingly improving the identification capability of the model on weak fault signals.
Through the technical scheme, the method and the device can effectively inhibit the problem of model overfitting caused by training data noise or missing, enhance the generalization capability of the multi-mode large model in a complex dynamic scene, enable the fault diagnosis result to not only accord with the data statistics rule, but also meet the physical operation constraint of a power grid, and remarkably reduce the fault misjudgment and missed judgment probability in a high-permeability environment of new energy.
Furthermore, knowledge distillation and light deployment can be adopted for the multi-mode large model, so that fault diagnosis is accelerated correspondingly, and the running stability of the power distribution network is improved. The loss function is set as:
In the formula, The hidden layer characteristic representation is respectively represented by a teacher model and a student model, lambda is a weight parameter, and KL divergence is used for measuring the output probability distribution P of the teacher model and the student model.
Through the distillation design, the student model can achieve order-of-magnitude efficiency improvement under the condition of extremely low precision loss, and key technical support is provided for the method.
The embodiment of the power distribution network fault diagnosis method based on the large model provided by the application is described in detail below.
Referring to fig. 2, a power distribution network fault diagnosis device based on a large model provided in this embodiment includes:
An operation data obtaining unit 201, configured to obtain operation data of a power distribution network, where the operation data of the power distribution network includes topology data, electrical quantity data, and environmental data;
A dynamic graph network construction unit 202, configured to construct a dynamic graph annotation network model according to topology data and electrical quantity data in the operation data of the power distribution network, where the dynamic graph annotation network model is used to reflect a real-time topology dynamic change relationship of the power distribution network;
The power distribution network fault diagnosis unit 203 is configured to perform feature fusion processing on power distribution network operation data of different modes through a cross-mode attention mechanism based on a preset multi-mode large model to obtain mode fusion features, so as to obtain a fault diagnosis result of the power distribution network based on the mode fusion features.
As shown in fig. 3, the embodiment also provides a power distribution network fault diagnosis terminal based on a large model, which comprises a memory 33 and a processor 31, wherein the memory 33 and the processor 31 can be connected through a communication bus 34;
the memory 33 is configured to store program codes for implementing a power distribution network fault diagnosis method based on a large model as provided in the above embodiment;
The processor 31 is for reading and executing the program code.
The memory is a hardware device with a data storage function, and can be specifically realized by a solid state disk or a flash memory chip, and is used for storing program codes comprising dynamic graph meaning network construction logic, a multi-mode feature fusion algorithm and fault diagnosis rules. The processor is a computing unit with a data processing function, and can be realized by a multi-core central processing chip or a graphic processing chip, and is used for analyzing program codes and executing real-time topology dynamic modeling, cross-mode data alignment and fault reasoning operation. The program code refers to a software module containing a computer instruction sequence, and can be specifically written and realized by Python or C++ language, and is used for converting the operation data of the power distribution network into the structural characteristics of the dynamic graph, and realizing the cooperative analysis of electric quantity, topology and environmental data by a cross-modal attention mechanism.
Specifically, when the terminal operates, after the program code preset in the memory is loaded by the processor, the real-time topology data, the electrical measurement information and the meteorological environment monitoring data are firstly acquired from the external data acquisition system, and the data cleaning and normalization operation is executed. And then, the processor constructs a dynamic graph annotation network model according to the topological connection relation and the electrical quantity time sequence characteristic, captures the dynamic coupling relation between nodes through a spatial attention mechanism, and extracts a time dimension characteristic evolution rule by combining with an LSTM network. In the multi-mode fusion stage, the processor inputs coding features of different modes into a cross-mode attention layer, aligns topology and environmental features by taking electric quantity data as a reference, and calculates the inter-mode association weight through QKV mapping so as to generate a fused global feature vector. Finally, the processor performs fault type classification and positioning calculation based on the fusion characteristics, and outputs a diagnosis result to the man-machine interaction interface or the control execution system.
The present application also provides a computer readable storage medium, in which a program code is stored, where the program code is used to be read and executed by a processor, so as to implement a power distribution network fault diagnosis method based on a large model provided in the foregoing embodiment.
It will be clearly understood by those skilled in the art that, for convenience and brevity of description, specific working procedures of the terminal, apparatus and unit described above may refer to corresponding procedures in the foregoing method embodiments, which are not repeated herein.
In the several embodiments provided in the present application, it should be understood that the disclosed terminal, apparatus and method may be implemented in other manners. For example, the apparatus embodiments described above are merely illustrative, e.g., the division of the units is merely a logical function division, and there may be additional divisions when actually implemented, e.g., multiple units or components may be combined or integrated into another system, or some features may be omitted or not performed. Alternatively, the coupling or direct coupling or communication connection shown or discussed with each other may be an indirect coupling or communication connection via some interfaces, devices or units, which may be in electrical, mechanical or other form.
The terms "first," "second," "third," "fourth," and the like in the description of the application and in the above figures, if any, are used for distinguishing between similar objects and not necessarily for describing a particular sequential or chronological order. It is to be understood that the data so used may be interchanged where appropriate such that the embodiments of the application described herein may be implemented, for example, in sequences other than those illustrated or otherwise described herein. Furthermore, the terms "comprises," "comprising," and "having," and any variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, system, article, or apparatus that comprises a list of steps or elements is not necessarily limited to those steps or elements expressly listed but may include other steps or elements not expressly listed or inherent to such process, method, article, or apparatus.
It should be understood that in the present application, "at least one (item)" means one or more, and "a plurality" means two or more. "and/or" is used to describe an association relationship of an associated object, and indicates that three relationships may exist, for example, "a and/or B" may indicate that only a exists, only B exists, and three cases of a and B exist simultaneously, where a and B may be singular or plural. The character "/" generally indicates that the context-dependent object is an "or" relationship. "at least one of" or the like means any combination of these items, including any combination of single item(s) or plural items(s). For example, at least one of a, b or c may represent a, b, c, "a and b", "a and c", "b and c", or "a and b and c", wherein a, b, c may be single or plural.
The units described as separate units may or may not be physically separate, and units shown as units may or may not be physical units, may be located in one place, or may be distributed on a plurality of network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.
In addition, each functional unit in the embodiments of the present invention may be integrated in one processing unit, or each unit may exist alone physically, or two or more units may be integrated in one unit. The integrated units may be implemented in hardware or in software functional units.
The integrated units, if implemented in the form of software functional units and sold or used as stand-alone products, may be stored in a computer readable storage medium. Based on such understanding, the technical solution of the present invention may be embodied essentially or in part or all of the technical solution or in part in the form of a software product stored in a storage medium, including instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method according to the embodiments of the present invention. The storage medium includes a U disk, a removable hard disk, a Read-Only Memory (ROM), a random access Memory (RAM, random Access Memory), a magnetic disk, an optical disk, or other various media capable of storing program codes.
While the application has been described in detail with reference to the foregoing embodiments, it will be understood by those skilled in the art that the foregoing embodiments may be modified or equivalents may be substituted for some of the features thereof, and that the modifications or substitutions do not depart from the spirit and scope of the embodiments of the application.
Claims (10)
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202510764101.4A CN120337157B (en) | 2025-06-10 | 2025-06-10 | Distribution network fault diagnosis method, device, terminal and medium based on large model |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202510764101.4A CN120337157B (en) | 2025-06-10 | 2025-06-10 | Distribution network fault diagnosis method, device, terminal and medium based on large model |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN120337157A true CN120337157A (en) | 2025-07-18 |
| CN120337157B CN120337157B (en) | 2025-09-30 |
Family
ID=96361076
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202510764101.4A Active CN120337157B (en) | 2025-06-10 | 2025-06-10 | Distribution network fault diagnosis method, device, terminal and medium based on large model |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN120337157B (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN121352035A (en) * | 2025-12-17 | 2026-01-16 | 国网上海市电力公司 | A Smart Fault Diagnosis Device and Method for Distribution Networks Based on AI Large Model |
| CN121432026A (en) * | 2025-12-30 | 2026-01-30 | 北京昊创瑞通电气设备股份有限公司 | Distribution transformer analysis methods, devices, equipment, and media based on intelligent fusion terminals |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20190286506A1 (en) * | 2018-03-13 | 2019-09-19 | Nec Laboratories America, Inc. | Topology-inspired neural network autoencoding for electronic system fault detection |
| CN117609824A (en) * | 2023-11-09 | 2024-02-27 | 武汉华源电力设计院有限公司 | Active power distribution network topology identification and fault diagnosis analysis method, device and equipment |
| WO2024187506A1 (en) * | 2023-03-16 | 2024-09-19 | 广东电网有限责任公司广州供电局 | Fault locating method and apparatus applied to power distribution network, device, and medium |
| CN119619734A (en) * | 2025-02-11 | 2025-03-14 | 国网山西省电力公司运城供电公司 | A distribution network fault location method based on FSGATv2 and topology similarity |
| CN120011563A (en) * | 2025-01-09 | 2025-05-16 | 中国船舶集团有限公司第七一九研究所 | Social network user opinion classification method, device, computing device and storage medium |
-
2025
- 2025-06-10 CN CN202510764101.4A patent/CN120337157B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20190286506A1 (en) * | 2018-03-13 | 2019-09-19 | Nec Laboratories America, Inc. | Topology-inspired neural network autoencoding for electronic system fault detection |
| WO2024187506A1 (en) * | 2023-03-16 | 2024-09-19 | 广东电网有限责任公司广州供电局 | Fault locating method and apparatus applied to power distribution network, device, and medium |
| CN117609824A (en) * | 2023-11-09 | 2024-02-27 | 武汉华源电力设计院有限公司 | Active power distribution network topology identification and fault diagnosis analysis method, device and equipment |
| CN120011563A (en) * | 2025-01-09 | 2025-05-16 | 中国船舶集团有限公司第七一九研究所 | Social network user opinion classification method, device, computing device and storage medium |
| CN119619734A (en) * | 2025-02-11 | 2025-03-14 | 国网山西省电力公司运城供电公司 | A distribution network fault location method based on FSGATv2 and topology similarity |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN121352035A (en) * | 2025-12-17 | 2026-01-16 | 国网上海市电力公司 | A Smart Fault Diagnosis Device and Method for Distribution Networks Based on AI Large Model |
| CN121432026A (en) * | 2025-12-30 | 2026-01-30 | 北京昊创瑞通电气设备股份有限公司 | Distribution transformer analysis methods, devices, equipment, and media based on intelligent fusion terminals |
| CN121432026B (en) * | 2025-12-30 | 2026-03-20 | 北京昊创瑞通电气设备股份有限公司 | Distribution transformer analysis methods, devices, equipment, and media based on intelligent fusion terminals |
Also Published As
| Publication number | Publication date |
|---|---|
| CN120337157B (en) | 2025-09-30 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN119580022B (en) | Method and system for analyzing defects in wafer manufacturing based on big data | |
| CN120337157B (en) | Distribution network fault diagnosis method, device, terminal and medium based on large model | |
| CN117154680A (en) | A wind power power prediction method based on non-stationary Transformer model | |
| CN120579457B (en) | A Multimodal Electromechanical Transient Modeling Method for Wind Turbines Based on Artificial Intelligence | |
| CN119622648A (en) | A method for fault detection of aviation inverter based on information fusion algorithm | |
| CN117110748A (en) | A method for detecting abnormal operating status of substation main equipment based on converged terminals | |
| CN118094112B (en) | Irrigation data assimilation method and device, electronic equipment and storage medium | |
| CN118861869A (en) | A cable fault prediction method and system introducing uncertain factors | |
| CN120724182A (en) | Enterprise data clustering processing method and system based on NLP and machine learning | |
| CN120146291A (en) | A method, system and medium for predicting photovoltaic power generation capacity based on spatiotemporal graph neural network | |
| CN119513503A (en) | A data conversion method, device, medium and equipment based on machine learning | |
| CN120893626A (en) | A method, system, computer equipment, and dielectric for short-term photovoltaic power prediction | |
| CN119622561A (en) | Distributed data anomaly diagnosis method for virtual power plants based on deep neural network | |
| CN120542984B (en) | Intelligent wind power prediction method and system based on actual measurement power and wind speed cooperative assimilation | |
| CN116680624A (en) | Classification method, system and storage medium for metadata of power system | |
| CN120123773B (en) | A method and system for cross-domain data fusion and mining based on multimodal power systems | |
| CN121210929A (en) | A multimodal and reinforcement learning-based intelligent evaluation method and system | |
| CN120974946A (en) | A method and system for optimizing and training geotechnical engineering models based on a multi-source knowledge base | |
| CN119670982B (en) | Method for improving the accuracy of photovoltaic power generation capacity prediction | |
| CN118550757B (en) | Missing data-oriented micro-service system root cause positioning method, medium and device | |
| CN120509661A (en) | Social management command scheduling system and method based on multi-source data fusion and intelligent control | |
| CN120410346A (en) | Power system digital twin model generation method, device, computer equipment, readable storage medium and program product | |
| CN119719853A (en) | Power consumption anomaly detection method and system based on Beidou synchronous multi-source data and LSTM | |
| CN119809038A (en) | A node voltage prediction method and device for a new energy distribution network | |
| CN116881663B (en) | A method for quality testing of low-voltage current transformers based on digital twins |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |