Power distribution network fault multi-source heterogeneous feature identification method, system and device and medium

Abstract

The invention discloses a power distribution network fault multi-source heterogeneous feature recognition method, system and device and a medium, and the method comprises the steps: constructing a multi-source heterogeneous data base covering internal electrical quantity and state quantity and external weather and geographical environment, and improving the quality through employing a data management technology; performing time-frequency localization feature extraction on the transient electrical waveform by using wavelet transform, and constructing a Bayesian network model to process logic uncertainty to realize orthogonal complementation of physical waveform and logic state features; mapping the heterogeneous data to a unified space-time coordinate system through a three-level space-time grid mechanism; constructing a multi-time scale fault prediction model system comprising SVR, LSTM, VAE and GCN; and realizing differentiated operation and maintenance decision recommendation through the power distribution network fault multi-mode knowledge graph. According to the method, the fault identification accuracy and the emergency response efficiency of the power distribution network in a complex environment are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of power system automation technology, and in particular to a method, system, device and medium for identifying multi-source heterogeneous characteristics of distribution network faults. Background Technology

[0002] With rapid socio-economic development and continuous urbanization, residents' quality of life is improving, leading to a surge in electricity demand. Users' expectations for power supply capacity, power quality, and electricity services are rising, making power supply reliability a core indicator for evaluating the service level of power companies. As the "last mile" connecting users, the distribution network has numerous lines, wide coverage, long extension distances, and a particularly complex operating environment. These characteristics result in a persistently high failure rate, severely impacting the continuity of power supply. According to reliable statistics from the power industry, up to 80% of power outages are caused by distribution network line faults, making the distribution network a key bottleneck restricting the improvement of power supply reliability.

[0003] In actual operation and maintenance, current fault handling models and technical methods face numerous challenges. First, the ability to identify fault causes is weak—the system can often only locate the faulty section, but it struggles to pinpoint whether the cause is equipment insulation aging, lightning strikes, tree contact with power lines, animal intrusion, vehicle collisions, or construction damage. Existing methods primarily rely on single fault waveform data to infer the cause, but different faults like tree contact with power lines and high-resistance grounding can have very similar waveforms, limiting the accuracy of identification. Furthermore, faults caused by external factors are often closely related to non-electrical information such as weather, time, and geographical location; relying solely on electrical quantities makes a comprehensive judgment difficult. Second, the utilization rate of multi-source heterogeneous data fusion is low. Valuable resources such as SCADA real-time data, PMS equipment ledgers, emergency repair work orders, user electricity consumption records, meteorological information, and GIS geographic data often operate independently, lacking a unified spatiotemporal fusion framework, making it difficult to form a cohesive force. Third, traditional feature extraction methods, such as Fourier transform, cannot capture the transient details of a fault, and shallow models such as logistic regression and support vector machines cannot uncover deep patterns in high-dimensional data. Finally, the decision support is not intelligent enough, and troubleshooting still requires people to go to the site. The time spent on "finding" is far more than on "repairing", so the efficiency is naturally low. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a method, system, device, and medium for identifying multi-source heterogeneous features of distribution network faults to address the multiple shortcomings of existing distribution network fault handling methods, such as inaccurate fault cause identification, insufficient fusion of multi-source heterogeneous data, outdated feature extraction and modeling methods, and low level of intelligent decision support. These shortcomings make it difficult to meet the urgent needs for accurate, efficient, collaborative, and intelligent operation and maintenance under the background of high-reliability power supply.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a method for identifying multi-source heterogeneous features of faults in a distribution network, comprising: Acquire multi-source heterogeneous data; The multi-source heterogeneous data is subjected to quality optimization and sample balancing to generate a high-quality dataset; Based on the fault recording signals in the high-quality dataset, wavelet transform is used to perform multi-scale time-frequency decomposition to extract the electrical salient features in the time-frequency domain. Based on the switch status, equipment ledger and historical protection action records in the high-quality dataset, and combined with the distribution network physical topology and relay protection logic, a Bayesian network model is constructed. By learning conditional probabilities and inferring real-time alarm information, logical state features are generated. A three-level spatiotemporal grid system is established, and the electrical salient features, the logical state features, and the meteorological and geographical environmental information in the high-quality dataset are mapped to a unified grid coordinate system through spatiotemporal coding to form a fused multimodal feature tensor. Based on the fused multimodal feature tensor, a multi-timescale fault prediction model system is constructed to predict short-term sudden fault risks, medium- and long-term trend fault risks, and spatially propagated fault risks, generating multi-dimensional fault prediction results. A multimodal knowledge graph of distribution network faults is constructed. Based on the feature vectors corresponding to the multi-dimensional fault prediction results, the semantic similarity between the feature vectors and historical cases in the knowledge graph is calculated, and differentiated operation and maintenance emergency repair strategies are generated and recommended.

[0007] As a preferred embodiment of the multi-source heterogeneous feature identification method for distribution network faults described in this invention, the step of performing quality optimization and sample equalization processing on the multi-source heterogeneous data includes: Outliers in multi-source heterogeneous data are removed. For data that follows a normal distribution, the three sigma principle is used, and for data that does not follow a normal distribution, a distance threshold algorithm based on K-means clustering is used. Based on the data after removing outliers, weekdays and holidays are distinguished, and the missing time point data is filled by the average of multiple sample data of the same type of day and the same time in history. Based on the complete dataset after missing value imputation, an adaptive synthetic sampling algorithm is used to generate synthetic samples according to the density distribution of minority class fault samples in their local neighborhood.

[0008] As a preferred embodiment of the multi-source heterogeneous feature identification method for distribution network faults described in this invention, the step of using wavelet transform for multi-scale time-frequency decomposition includes: Daubechies wavelets were selected as basis functions. Based on the selected Daubechies wavelet basis function, the fault recording signal is subjected to at least three levels of discrete wavelet decomposition to obtain the detail coefficients of each level. Based on the detail coefficients of each layer, the energy proportion, wavelet energy spectrum Shannon entropy, and modulus maxima of the detail coefficients at each scale are calculated as electrical salient features in the time-frequency domain.

[0009] As a preferred embodiment of the multi-source heterogeneous feature identification method for distribution network faults described in this invention, the construction of the Bayesian network model includes: Based on the power grid public information model, the physical connection relationship of the distribution network and the main and backup coordination logic of relay protection are automatically analyzed to generate the topology of Bayesian network. In the aforementioned topology, the fault status of the distribution network components is set as the parent node, and the corresponding protection device action signal and switch change signal are set as child nodes. Based on the conditional probability relationship between the parent node and the child node, during real-time inference, when there is a missing alarm signal, the most likely faulty component is inferred according to the maximum a posteriori probability criterion, and the corresponding confidence index is output.

[0010] As a preferred embodiment of the multi-source heterogeneous feature identification method for distribution network faults described in this invention, the establishment of a three-level spatiotemporal grid system includes: The first-level grid is divided based on the power supply range of the substation; The second-level grid is divided based on the medium-voltage line connection structure; The third-level grid is divided according to administrative divisions, land use, and natural geographical boundaries, and serves as the smallest unit for data fusion.

[0011] The beneficial effects of this preferred technical solution are that by constructing a three-level spatiotemporal grid system, it achieves fine alignment of the physical topology, operating logic and external environmental elements of the distribution network under a unified spatial benchmark, providing structured support for the efficient fusion of multi-source heterogeneous data and the accurate mapping of fault characteristics.

[0012] As a preferred embodiment of the multi-source heterogeneous feature identification method for distribution network faults described in this invention, the construction of a multi-time-scale fault prediction model system includes: Based on the feature slices of the fused multimodal feature tensor at the current moment, a support vector regression model is used with the radial basis function as the kernel function to predict the risk of short-term sudden failures. Based on the time series sequence of the fused multimodal feature tensor in historical periods, a variational autoencoder is used to model the potential distribution of historical fault sequences to generate missing data samples. Long short-term memory network is used to capture long-term time series dependencies, and the outputs of long short-term memory network, autoregressive integral moving average model and triple exponential smoothing model are fused by stacked ensemble algorithm to predict medium- and long-term trend fault risks. Based on the fused multimodal feature tensor and the corresponding three-level spatiotemporal grid topology, a graph convolutional neural network is used, with the third-level grid as nodes and geographical adjacency and electrical connection as edges, to aggregate spatial neighborhood risk features and predict spatially propagating fault risks.

[0013] The beneficial effects of this preferred technical solution are that by constructing a multi-timescale model system that integrates short-term suddenness, medium- and long-term trend and spatial propagation risk prediction, it achieves a full-dimensional characterization of the evolution law of distribution network faults, and significantly improves the timeliness, robustness and spatial accuracy of fault prediction.

[0014] As a preferred embodiment of the multi-source heterogeneous feature identification method for distribution network faults described in this invention, the step of constructing a multi-modal knowledge graph of distribution network faults includes: Using a bidirectional long short-term memory network and a conditional random field model, fault entities, equipment entities, and policy entities are extracted from historical emergency repair work order texts; Based on the extracted entities, semantic relationship triples between entities are constructed according to the predefined ontology, and the translational distance embedding algorithm is used to map the entities and relationships to a unified low-dimensional vector space. Based on the historical case vectors in the low-dimensional vector space, the cosine similarity between the current fault feature vector and the historical case vectors is calculated, and several maintenance and repair strategies with the highest matching degree are recommended.

[0015] Secondly, the present invention provides a system for identifying multi-source heterogeneous characteristics of faults in a distribution network, comprising: A multi-source heterogeneous data access module is used to acquire multi-source heterogeneous data; The data quality enhancement and balancing module is used to perform quality optimization and sample balancing on the multi-source heterogeneous data to generate a high-quality dataset. The electrical transient feature extraction module is used to extract time-frequency domain electrical salient features by performing multi-scale time-frequency decomposition using wavelet transform based on the fault recording signal in the high-quality dataset. The logic state reasoning modeling module is used to construct a Bayesian network model based on the switch status, equipment ledger and historical protection action records in the high-quality dataset, combined with the distribution network physical topology and relay protection logic. By learning conditional probabilities and reasoning real-time alarm information, logic state features are generated. The multimodal spatiotemporal fusion coding module is used to establish a three-level spatiotemporal grid system and map the electrical salient features, the logical state features, and the meteorological and geographical environmental information in the high-quality dataset to a unified grid coordinate system through spatiotemporal coding to form a fused multimodal feature tensor. The multi-scale fault risk prediction module is used to construct a multi-time-scale fault prediction model system based on the fused multimodal feature tensor, and to predict short-term sudden fault risks, medium- and long-term trend fault risks and spatially propagated fault risks, respectively, and generate multi-dimensional fault prediction results. The knowledge-driven emergency repair strategy recommendation module is used to construct a multimodal knowledge graph of distribution network faults. Based on the feature vectors corresponding to the multi-dimensional fault prediction results, it calculates the semantic similarity between the feature vectors and historical cases in the knowledge graph, and generates and recommends differentiated operation and maintenance emergency repair strategies.

[0016] Thirdly, the present invention provides an electronic device, comprising: Memory, used to store programs; A processor is configured to execute the computer-executable instructions, which, when executed by the processor, implement the steps of the method for identifying multi-source heterogeneous features of power distribution network faults.

[0017] Fourthly, the present invention provides a computer-readable storage medium, comprising: when the program is executed by a processor, the steps of implementing the method for identifying multi-source heterogeneous features of power distribution network faults.

[0018] The beneficial effects of this invention are as follows: By constructing a three-level spatiotemporal grid system, this invention integrates the substation power supply range, medium-voltage line connection structure, administrative divisions, land use, and natural geographical boundaries at each level. Using the third-level grid as the smallest unit of data fusion, it achieves accurate alignment and structured organization of multi-source heterogeneous data (electrical, logical, meteorological, and geographical) under a unified spatiotemporal coordinate system. By fusing electrical salient features extracted from wavelet time-frequency decomposition with logical state features generated based on the Common Information Model of the Power Grid (CIM) and Bayesian network inference, and combining adaptive synthetic sampling (ADASYN) sample equalization and multimodal spatiotemporal coding, it achieves effective characterization of high-quality, high-dimensional fault features and enhanced noise robustness. By constructing a multi-timescale fault prediction model system covering short-term suddenness, medium-to-long-term trends, and spatial propagation risks, and linking it with the multimodal knowledge graph of distribution network faults for semantic similarity matching, it achieves closed-loop intelligent decision-making from "risk prediction" to "differentiated emergency repair strategy recommendation," significantly improving the accuracy, coverage, and operation and maintenance response efficiency of distribution network fault early warning. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein: Figure 1 This is a basic flowchart of a method for identifying multi-source heterogeneous features of power distribution network faults, provided in one embodiment of the present invention. Detailed Implementation

[0020] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0021] Example 1, referring to Figure 1 As an embodiment of the present invention, a method for identifying multi-source heterogeneous features of distribution network faults is provided, comprising: S100: Acquire multi-source heterogeneous data; S200: Perform quality optimization and sample balancing on the multi-source heterogeneous data to generate a high-quality dataset; S300: Based on the fault recording signal in the high-quality dataset, wavelet transform is used to perform multi-scale time-frequency decomposition to extract the electrical salient features in the time-frequency domain. S400: Based on the switch status, equipment ledger and historical protection action records in the high-quality dataset, and combined with the distribution network physical topology and relay protection logic, a Bayesian network model is constructed. By learning conditional probabilities and inferring real-time alarm information, logical state features are generated. S500: Establish a three-level spatiotemporal grid system, and map the electrical salient features, the logical state features, and the meteorological and geographical environmental information in the high-quality dataset to a unified grid coordinate system through spatiotemporal coding to form a fused multimodal feature tensor; S600: Based on the fused multimodal feature tensor, a multi-timescale fault prediction model system is constructed to predict short-term sudden fault risks, medium- and long-term trend fault risks, and spatially propagated fault risks, respectively, and generate multi-dimensional fault prediction results. S700: Construct a multimodal knowledge graph of distribution network faults, calculate the semantic similarity between the feature vectors corresponding to the multi-dimensional fault prediction results and historical cases in the knowledge graph based on the feature vectors, and generate and recommend differentiated operation and maintenance emergency repair strategies.

[0022] It should be noted that existing power distribution network fault handling methods face a series of challenges during operation, including: data sources being scattered and highly heterogeneous; a lack of unified spatiotemporal benchmarks for multi-source information such as waveform recording, switch position changes, meteorology, and geography, making effective integration difficult; traditional fault diagnosis models often rely on single electrical quantities or rule bases, resulting in poor adaptability to alarm-deficient scenarios such as protection failure to operate or communication interruptions, and limited logical reasoning capabilities; furthermore, existing prediction methods typically only focus on short-term or static risks, lacking collaborative modeling of fault suddenness, equipment aging trends, and fault spatial propagation, leading to a single prediction dimension and insufficient timeliness; finally, repair strategies often rely on manual experience or simple template matching, failing to extract structured knowledge from historical work orders and achieve semantic-level intelligent recommendations, making it difficult to support differentiated and precise operation and maintenance decisions.

[0023] Therefore, addressing the shortcomings of existing power distribution network fault handling methods, such as inaccurate fault cause identification, insufficient fusion of multi-source heterogeneous data, outdated feature extraction and modeling methods, and low level of intelligent decision support, which make it difficult to meet the urgent needs for accurate, efficient, collaborative, and intelligent operation and maintenance under the background of high-reliability power supply, this paper proposes a closed-loop intelligent operation and maintenance system for power distribution network faults. This system, through steps S100-S700, integrates multi-source heterogeneous data governance, multi-modal feature engineering, three-level spatiotemporal grid alignment, multi-timescale prediction, and knowledge graph-driven decision-making. This system achieves accurate perception, full-dimensional prediction, and intelligent strategy generation of power distribution network faults, significantly improving the accuracy, robustness, and refinement of fault early warning and emergency repair response.

[0024] Example 2, this is an embodiment of the present invention, which provides a method for identifying multi-source heterogeneous features of distribution network faults based on the previous embodiment, including: In this embodiment of the application, the multi-source heterogeneous data in step S100 includes internal data and external data. The internal data includes fault recording signals, switch status, and equipment ledgers, while the external data includes meteorological information and geographic environment information. In the internal data, the sampling frequency of the fault recording signals is 4kHz or 10kHz, and the equipment ledgers include key fields {Device_ID, Type, Install_Date, Capacity, Location_GPS}. In the external data, the meteorological information is obtained by connecting to the meteorological bureau API, and the spatial resolution is 1km×1km gridded data, which includes elements such as rainfall (mm / h), lightning density, and wind speed (m / s).

[0025] In this embodiment of the application, step S200, which involves quality optimization and sample balancing of multi-source heterogeneous data, includes: Outliers in multi-source heterogeneous data are removed. For data that follows a normal distribution, the three sigma principle is used, and for data that does not follow a normal distribution, a distance threshold algorithm based on K-means clustering is used. Based on the data after removing outliers, weekdays and holidays are distinguished, and the missing time point data is filled by the average of multiple sample data of the same type of day and the same time in history. Based on the complete dataset after missing value imputation, the Adaptive Synthetic Sampling (ADASYN) algorithm is used to generate synthetic samples according to the density distribution of minority class fault samples in their local neighborhoods.

[0026] In this embodiment of the application, outlier removal in step S200 is achieved by using the three sigma principle to remove outliers from data that follows a normal distribution, and by calculating the distance from the sample to the cluster center based on K-means clustering for non-normally distributed data, and identifying and removing outliers based on a preset distance threshold.

[0027] In an optional implementation, the outlier removal in step S200 can also determine the interquartile range for each feature of the multi-source heterogeneous data, identify low or high value data that deviate significantly from the range as outliers, and remove them from the dataset.

[0028] In an optional implementation, outlier removal in step S200 can also be carried out by constructing an isolated forest model based on the multidimensional features in the high-quality dataset, generating isolation trees through random segmentation, calculating anomaly scores based on the average path length of the samples, and identifying data points with anomaly scores higher than a preset threshold as outliers for removal.

[0029] In this embodiment, during the missing value imputation stage, considering the significant daily / weekly periodicity of the load data, the "conditional periodic mean method" is adopted: if the missing time is a working day, the average value of the data at the same time on all working days within the past N=4 weeks is used for imputation. The specific calculation formula is as follows: in, For missing moments The fill value, Indicates the previous number Historical observations for the same hour on the same day of the week. =4.

[0030] In this embodiment of the application, step S300 uses wavelet transform for multi-scale time-frequency decomposition, including: Daubechies wavelets were selected as basis functions. Based on the selected Daubechies wavelet basis function, the fault recording signal is subjected to at least three levels of discrete wavelet decomposition to obtain the detail coefficients of each level. Based on the detail coefficients at each level, the energy proportion, wavelet energy spectrum Shannon entropy, and modulus maxima of the detail coefficients at each scale are calculated as time-frequency domain electrical salient features; among which, the first... The energy of the layer wavelet detail coefficients is defined as: Normalized energy percentage ,in The formula for calculating the Shannon entropy of wavelet energy spectrum is: in, For the first Layer energy percentage.

[0031] In this embodiment of the application, the wavelet basis function selection in step S300 is to use the Daubechies 4 (db4) wavelet as the mother wavelet, and take advantage of its good compact support and orthogonality to perform multi-scale discrete wavelet decomposition on the fault recording signal, so as to effectively extract transient fault features in different frequency bands.

[0032] In an optional implementation, the wavelet basis function selection in step S300 can also use Morlet wavelet to perform continuous wavelet transform on the fault recording signal, generate a high-resolution time-frequency spectrum on the time-frequency plane, and extract time-frequency domain electrical salient features such as energy concentration areas and instantaneous frequency change points from it.

[0033] In an optional implementation, the wavelet basis function selection in step S300 can also use Symlets wavelets, which have approximate symmetry and tight support characteristics, to perform multi-scale discrete wavelet decomposition on the fault recording signal, so as to effectively extract the transient electrical features of each frequency band while preserving the accurate location of the fault initiation time.

[0034] In this embodiment, the Daubechies 4 (db4) wavelet, which has good compact support and orthogonality, is selected as the mother wavelet. A 5-level discrete wavelet decomposition (DWT) is performed on the zero-sequence current signal with a sampling rate of 10kHz. The signal is decomposed into: approximation coefficients. (0-156.25Hz, predominantly fundamental component) and detail coefficients (2.5k-5kHz) (1.25k-2.5kHz), ..., (156.25-312.5Hz). Among them The layer corresponds to the extremely high frequency band, which is usually caused by lightning surges; - The layer corresponds to the mid-frequency band and typically contains the characteristics of random electric arcs generated when trees collide with power lines.

[0035] In this embodiment of the application, step S400, which involves constructing a Bayesian network model, includes: Based on the power grid public information model, the physical connection relationship of the distribution network and the main and backup coordination logic of relay protection are automatically analyzed to generate the topology of Bayesian network. In the aforementioned topology, the fault status of the distribution network components is set as the parent node, and the corresponding protection device action signal and switch change signal are set as child nodes. Based on the conditional probability relationship between the parent node and the child node, during real-time inference, when there is a missing alarm signal, the most likely faulty component is inferred according to the maximum a posteriori probability criterion, and the corresponding confidence index is output.

[0036] Specifically, with line faults Circuit breaker tripping For example, when the protection signal is missing but observation is made... When the line fault occurs, the posterior probability is calculated using the following formula: in, Learned from historical data, Let the prior probability of a line fault be denominator. The posterior probability is obtained by marginalizing all fault assumptions; when this posterior probability exceeds a preset confidence threshold (e.g., 85%), the line is determined to be faulty.

[0037] In this embodiment, the topology is a three-layer directed graph: the parent node layer is the fault source (such as line fault, transformer fault), the middle node layer is the main / backup protection device, and the child node layer is the circuit breaker tripping action element. The connection relationship strictly follows the physical causal chain of "fault → protection → switch". In the embodiments of this application, the conditional probability table is learned from 5 years of historical fault records through maximum likelihood estimation. For example, the protection failure rate P(main protection = 0 | line fault = 1) = 0.05, which supports fault-tolerant inference through edge computing when the protection signal is lost.

[0038] In this embodiment of the application, step S500, which establishes a three-level spatiotemporal grid system, includes: The first-level grid is divided based on the power supply range of the substation; The second-level grid is divided based on the medium-voltage line connection structure; The third-level grid is divided according to administrative divisions, land use, and natural geographical boundaries, and serves as the smallest unit for data fusion.

[0039] In this embodiment, the third-level grid adopts a standard 500m×500m grid and is fine-tuned using GIS geofencing technology combined with natural boundaries. Each grid is assigned a unique GeoHash code for unified mapping of static attributes (such as average altitude and vegetation coverage) and dynamic environmental factors (such as maximum wind speed and cumulative rainfall). Through Pearson correlation analysis, it was found that the influence of environmental factors differs significantly among different grid types. For example, the correlation coefficient between wind speed and failure rate in the "mountainous area-high vegetation" grid reaches 0.82, which verifies the necessity of refined grid division.

[0040] In this embodiment of the application, step S600, which involves constructing a multi-timescale fault prediction model system, includes: Based on the feature slices of the fused multimodal feature tensor at the current moment, a support vector regression (SVR) model is used with the radial basis function as the kernel function to predict the risk of short-term sudden failures. Based on the time series of the fused multimodal feature tensor in historical periods, a variational autoencoder (VAE) is used to model the potential distribution of historical fault sequences to generate missing data samples. Long short-term memory network (LSTM) is used to capture long-term time series dependencies. The outputs of the LSTM network, the autoregressive integral moving average model and the triple exponential smoothing model are combined with the stacked ensemble algorithm to predict medium- and long-term trend fault risks. Based on the fused multimodal feature tensor and the corresponding three-level spatiotemporal grid topology, a graph convolutional neural network is used, with the third-level grid as nodes and geographical adjacency and electrical connection as edges, to aggregate spatial neighborhood risk features and predict spatially propagating fault risks.

[0041] In this embodiment of the application, in step S600, the short-term failure risk prediction model extracts the fused multimodal feature tensor slices at the current moment, inputs them into the support vector regression (SVR) model using radial basis function (RBF) as the kernel function, and outputs the sudden failure risk value in the short term after hyperparameter optimization.

[0042] In an optional implementation, the short-term failure risk prediction model in step S600 can also take the fused multimodal feature tensor slice at the current moment as input, use gradient boosting tree algorithms such as XGBoost or LightGBM to automatically learn the nonlinear relationship between features and failure risk, and output continuous predicted values ​​of short-term sudden failure risk by integrating multiple decision trees.

[0043] In an optional implementation, in step S600, the short-term failure risk prediction model can also input the current fused multimodal feature tensor slice into a multilayer perceptron consisting of 2 to 3 hidden layers, perform end-to-end training through ReLU activation function and Dropout regularization, and finally output continuous prediction values ​​of short-term sudden failure risk.

[0044] In this embodiment, the radial basis function (RBF) is selected as the kernel function to handle nonlinear relationships, and its mathematical expression is as follows: in, For the input feature vector, Denotes the Euclidean norm. The kernel function width parameter controls the influence range of a single training sample. In this embodiment, the Support Vector Regression (SVR) model optimizes the penalty coefficient C and kernel parameter γ using a grid search, and its objective function is: in, For the weight vector, For bias terms, , As slack variables, For regularization parameters, This represents the number of samples.

[0045] In this embodiment of the application, LSTM controls the information flow through a gating mechanism, and its core computation includes: in, , , These are the forget gate, input gate, and output gate, respectively. In cellular state, In hidden state, It is the Sigmoid activation function. This represents element-wise product.

[0046] In this embodiment of the application, in the Stacking ensemble framework, the outputs of LSTM, Autoregressive Integral Moving Average (ARIMA) model and Holt-Winters triple exponential smoothing model are used as meta-features and input into XGBoost for final fusion, thereby further improving prediction stability. In this embodiment, GCN uses a normalized adjacency matrix with self-loops for feature propagation, and its first... layer to the first The update formula for the layer is: in, To add a self-loop adjacency matrix, for The degree matrix, For the first Layer node feature matrix For learnable weight matrix, It is a non-linear activation function.

[0047] In this embodiment of the application, step S700, which involves constructing a multimodal knowledge graph of distribution network faults, includes: Using a bidirectional long short-term memory network and a conditional random field model, fault entities, equipment entities, and policy entities are extracted from historical emergency repair work order texts; Based on the extracted entities, semantic relationship triples between entities are constructed according to the predefined ontology, and the translational distance embedding algorithm is used to map the entities and relationships to a unified low-dimensional vector space. Based on the historical case vectors in the low-dimensional vector space, the cosine similarity between the current fault feature vector and the historical case vectors is calculated, and several maintenance and repair strategies with the highest matching degree are recommended.

[0048] In this embodiment of the application, the knowledge graph entity relationship embedding in step S700 uses the translational distance embedding algorithm (TransE) to map the entities such as faults, equipment and strategies extracted from historical emergency repair work orders and their semantic relationships to a unified low-dimensional vector space, so that the relationship triples satisfy the translational property of "head entity vector + relationship vector ≈ tail entity vector", thereby supporting intelligent strategy recommendation based on vector similarity.

[0049] In an optional implementation, the knowledge graph entity relation embedding in step S700 can also model each relation as a rotation transformation from the head entity to the tail entity in the complex vector space. By learning the complex embedding representation of entities and relations, the triples can satisfy "tail entity vector ≈ head entity vector ∘ relation rotation", thereby more accurately characterizing the symmetric, antisymmetric and inverse relations in the distribution network fault knowledge graph, and supporting high-precision semantic similarity calculation and strategy recommendation.

[0050] In an optional implementation, the knowledge graph entity relationship embedding in step S700 can also use entities of the power distribution network fault knowledge graph as nodes and semantic relationships as edges, and use GraphSAGE or Graph Attention Network (GAT) to aggregate neighbor node information on the graph structure to adaptively generate entity embedding vectors containing local topology and semantic context, and then realize intelligent recommendation of historical emergency repair strategies through vector similarity matching.

[0051] In this embodiment of the application, the extracted entity types are further subdivided into {Fault, Cause, Equipment, Action, Tool}, and semantic relationships are established in the graph database; In this embodiment, the recommendation results are combined with GIS route planning to provide auxiliary decision-making information such as estimated arrival time.

[0052] In this embodiment, the system adopts a microservice architecture: the data access layer accesses the SCADA stream in real time through Kafka and synchronizes the PMS ledger through ETL; the algorithm service layer integrates PyTorch, PyWavelets, pgmpy and Scikit-learn libraries based on Python 3.8; the application interaction layer uses Vue.js and WebGL to implement 3D grid heatmap visualization, supporting 5-minute risk updates and grid profile radar chart display.

[0053] Example 3 is an embodiment of the present invention. This embodiment differs from the first embodiment in that it provides a multi-source heterogeneous feature identification system for distribution network faults.

[0054] It should be noted that the technical solution of the distribution network fault multi-source heterogeneous feature identification system is based on the same concept as the above-mentioned distribution network fault multi-source heterogeneous feature identification method. For details not described in detail in the technical solution of the distribution network fault multi-source heterogeneous feature identification system in this embodiment, please refer to the description of the above-mentioned distribution network fault multi-source heterogeneous feature identification method.

[0055] This embodiment provides a multi-source heterogeneous feature identification system for distribution network faults, comprising: A multi-source heterogeneous data access module is used to acquire multi-source heterogeneous data; The data quality enhancement and balancing module is used to perform quality optimization and sample balancing on the multi-source heterogeneous data to generate a high-quality dataset. The electrical transient feature extraction module is used to extract time-frequency domain electrical salient features by performing multi-scale time-frequency decomposition using wavelet transform based on the fault recording signal in the high-quality dataset. The logic state reasoning modeling module is used to construct a Bayesian network model based on the switch status, equipment ledger and historical protection action records in the high-quality dataset, combined with the distribution network physical topology and relay protection logic. By learning conditional probabilities and reasoning real-time alarm information, logic state features are generated. The multimodal spatiotemporal fusion coding module is used to establish a three-level spatiotemporal grid system and map the electrical salient features, the logical state features, and the meteorological and geographical environmental information in the high-quality dataset to a unified grid coordinate system through spatiotemporal coding to form a fused multimodal feature tensor. The multi-scale fault risk prediction module is used to construct a multi-time-scale fault prediction model system based on the fused multimodal feature tensor, and to predict short-term sudden fault risks, medium- and long-term trend fault risks and spatially propagated fault risks, respectively, and generate multi-dimensional fault prediction results. The knowledge-driven emergency repair strategy recommendation module is used to construct a multimodal knowledge graph of distribution network faults. Based on the feature vectors corresponding to the multi-dimensional fault prediction results, it calculates the semantic similarity between the feature vectors and historical cases in the knowledge graph, and generates and recommends differentiated operation and maintenance emergency repair strategies.

[0056] This embodiment also provides an electronic device applicable to a method for identifying multi-source heterogeneous characteristics of distribution network faults, including: The system includes a memory and a processor. The memory stores computer-executable instructions, and the processor executes these instructions to implement a method for identifying multi-source heterogeneous features of power distribution network faults, as proposed in the above embodiments.

[0057] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements a method for identifying multi-source heterogeneous features of distribution network faults as proposed in the above embodiments.

[0058] The storage medium proposed in this embodiment belongs to the same inventive concept as the method for identifying multi-source heterogeneous features of distribution network faults proposed in the above embodiments. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0059] Based on the above description of the implementation methods, those skilled in the art can clearly understand that the present invention can be implemented using software and necessary general-purpose hardware, and of course, it can also be implemented using hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk, or optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of the various embodiments of the present invention.

[0060] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for identifying multi-source heterogeneous features of distribution network faults, characterized in that, include: Acquire multi-source heterogeneous data; The multi-source heterogeneous data is subjected to quality optimization and sample balancing to generate a high-quality dataset; Based on the fault recording signals in the high-quality dataset, wavelet transform is used to perform multi-scale time-frequency decomposition to extract the electrical salient features in the time-frequency domain. Based on the switch status, equipment ledger and historical protection action records in the high-quality dataset, and combined with the distribution network physical topology and relay protection logic, a Bayesian network model is constructed. By learning conditional probabilities and inferring real-time alarm information, logical state features are generated. A three-level spatiotemporal grid system is established, and the electrical salient features, the logical state features, and the meteorological and geographical environmental information in the high-quality dataset are mapped to a unified grid coordinate system through spatiotemporal coding to form a fused multimodal feature tensor. Based on the fused multimodal feature tensor, a multi-timescale fault prediction model system is constructed to predict short-term sudden fault risks, medium- and long-term trend fault risks, and spatially propagated fault risks, generating multi-dimensional fault prediction results. A multimodal knowledge graph of distribution network faults is constructed. Based on the feature vectors corresponding to the multi-dimensional fault prediction results, the semantic similarity between the feature vectors and historical cases in the knowledge graph is calculated, and differentiated operation and maintenance emergency repair strategies are generated and recommended.

2. The method for identifying multi-source heterogeneous features of distribution network faults as described in claim 1, characterized in that: The process of quality optimization and sample balancing for multi-source heterogeneous data includes: Outliers in multi-source heterogeneous data are removed. For data that follows a normal distribution, the three sigma principle is used, and for data that does not follow a normal distribution, a distance threshold algorithm based on K-means clustering is used. Based on the data after removing outliers, weekdays and holidays are distinguished, and the missing time point data is filled by the average of multiple sample data of the same type of day and the same time in history. Based on the complete dataset after missing value imputation, an adaptive synthetic sampling algorithm is used to generate synthetic samples according to the density distribution of minority class fault samples in their local neighborhood.

3. The method for identifying multi-source heterogeneous features of distribution network faults as described in claim 1 or 2, characterized in that: The multi-scale time-frequency decomposition using wavelet transform includes: Daubechies wavelets were selected as basis functions. Based on the selected Daubechies wavelet basis function, the fault recording signal is subjected to at least three levels of discrete wavelet decomposition to obtain the detail coefficients of each level. Based on the detail coefficients of each layer, the energy proportion, wavelet energy spectrum Shannon entropy, and modulus maxima of the detail coefficients at each scale are calculated as electrical salient features in the time-frequency domain.

4. The method for identifying multi-source heterogeneous features of distribution network faults as described in claim 3, characterized in that: The construction of the Bayesian network model includes: Based on the power grid public information model, the physical connection relationship of the distribution network and the main and backup coordination logic of relay protection are automatically analyzed to generate the topology of Bayesian network. In the aforementioned topology, the fault status of the distribution network components is set as the parent node, and the corresponding protection device action signal and switch change signal are set as child nodes. Based on the conditional probability relationship between the parent node and the child node, during real-time inference, when there is a missing alarm signal, the most likely faulty component is inferred according to the maximum a posteriori probability criterion, and the corresponding confidence index is output.

5. The method for identifying multi-source heterogeneous features of distribution network faults as described in claim 4, characterized in that: The establishment of the three-level spatiotemporal grid system includes: The first-level grid is divided based on the power supply range of the substation; The second-level grid is divided based on the medium-voltage line connection structure; The third-level grid is divided according to administrative divisions, land use, and natural geographical boundaries, and serves as the smallest unit for data fusion.

6. The method for identifying multi-source heterogeneous features of distribution network faults as described in claim 5, characterized in that: The construction of the multi-timescale fault prediction model system includes: Based on the feature slices of the fused multimodal feature tensor at the current moment, a support vector regression model is used with the radial basis function as the kernel function to predict the risk of short-term sudden failures. Based on the time series sequence of the fused multimodal feature tensor in historical periods, a variational autoencoder is used to model the potential distribution of historical fault sequences to generate missing data samples. Long short-term memory network is used to capture long-term time series dependencies, and the outputs of long short-term memory network, autoregressive integral moving average model and triple exponential smoothing model are fused by stacked ensemble algorithm to predict medium- and long-term trend fault risks. Based on the fused multimodal feature tensor and the corresponding three-level spatiotemporal grid topology, a graph convolutional neural network is used, with the third-level grid as nodes and geographical adjacency and electrical connection as edges, to aggregate spatial neighborhood risk features and predict spatially propagating fault risks.

7. The method for identifying multi-source heterogeneous features of distribution network faults as described in claim 6, characterized in that: The construction of the multimodal knowledge graph of distribution network faults includes: Using a bidirectional long short-term memory network and a conditional random field model, fault entities, equipment entities, and policy entities are extracted from historical emergency repair work order texts; Based on the extracted entities, semantic relationship triples between entities are constructed according to the predefined ontology, and the translational distance embedding algorithm is used to map the entities and relationships to a unified low-dimensional vector space. Based on the historical case vectors in the low-dimensional vector space, the cosine similarity between the current fault feature vector and the historical case vectors is calculated, and several maintenance and repair strategies with the highest matching degree are recommended.

8. A multi-source heterogeneous feature identification system for distribution network faults, using the method described in any one of claims 1-7, characterized in that, include: A multi-source heterogeneous data access module is used to acquire multi-source heterogeneous data; The data quality enhancement and balancing module is used to perform quality optimization and sample balancing on the multi-source heterogeneous data to generate a high-quality dataset. The electrical transient feature extraction module is used to extract time-frequency domain electrical salient features by performing multi-scale time-frequency decomposition using wavelet transform based on the fault recording signal in the high-quality dataset. The logic state reasoning modeling module is used to construct a Bayesian network model based on the switch status, equipment ledger and historical protection action records in the high-quality dataset, combined with the distribution network physical topology and relay protection logic. By learning conditional probabilities and reasoning real-time alarm information, logic state features are generated. The multimodal spatiotemporal fusion coding module is used to establish a three-level spatiotemporal grid system and map the electrical salient features, the logical state features, and the meteorological and geographical environmental information in the high-quality dataset to a unified grid coordinate system through spatiotemporal coding to form a fused multimodal feature tensor. The multi-scale fault risk prediction module is used to construct a multi-time-scale fault prediction model system based on the fused multimodal feature tensor, and to predict short-term sudden fault risks, medium- and long-term trend fault risks and spatially propagated fault risks, respectively, and generate multi-dimensional fault prediction results. The knowledge-driven emergency repair strategy recommendation module is used to construct a multimodal knowledge graph of distribution network faults. Based on the feature vectors corresponding to the multi-dimensional fault prediction results, it calculates the semantic similarity between the feature vectors and historical cases in the knowledge graph, and generates and recommends differentiated operation and maintenance emergency repair strategies.

9. An electronic device, characterized in that, include: Memory, used to store programs; A processor for loading the program to perform the steps of the method as claimed in any one of claims 1-7.

10. A computer-readable storage medium storing a program, characterized in that, When the program is executed by a processor, it implements the steps of the method as described in any one of claims 1-7.

Images