An overhead low-voltage line fault analysis method and system based on artificial intelligence

Abstract

The application discloses an overhead low-voltage line fault analysis method and system based on artificial intelligence, relates to the technical field of artificial intelligence, and solves the technical problem that original electrical characteristics are often used for analysis, there is a lack of coupling division and importance judgment of the characteristics, the model is difficult to distinguish the root cause of the fault and the accompanying phenomenon, and the accuracy and efficiency of the fault diagnosis result are low; an initial multi-source heterogeneous data is used to construct a dynamic feature matrix and a dynamic knowledge graph, feature decoupling and fault diagnosis are performed on the dynamic feature matrix and the dynamic knowledge graph, and a fault diagnosis result is obtained; a fault prediction result is obtained based on the fault diagnosis result and the dynamic knowledge graph, original mixed features are separated into independent feature components by using a deep feature decoupling network, the importance of each feature component is dynamically evaluated in combination with a topological relationship in the knowledge graph and an attention mechanism, the fault diagnosis model can clearly distinguish the root cause of the fault and the accompanying phenomenon, and the accuracy and efficiency of the fault diagnosis are improved.

Description

Technical Field

[0001] This application belongs to the field of artificial intelligence technology, specifically an artificial intelligence-based method and system for analyzing faults in overhead low-voltage lines. Background Technology

[0002] In modern power systems, overhead low-voltage lines serve as a crucial link connecting the distribution network and end users, undertaking the vital task of the "last mile" of power distribution. Their operational status directly impacts power supply reliability, electricity safety, and user satisfaction. However, due to their long-term exposure to the natural environment, overhead lines are susceptible to multiple factors such as lightning strikes, windstorms, tree obstructions, external damage, and equipment aging, resulting in a relatively high failure rate. Systematic analysis of overhead low-voltage line faults not only helps to accurately identify the causes of faults and improve repair efficiency but also provides a scientific basis for line design optimization, operation and maintenance strategy formulation, and the construction of intelligent distribution systems.

[0003] Existing technologies often directly use mixed original electrical characteristics for fault analysis, lacking the ability to couple and classify features and assess their importance. This makes it difficult for the model to distinguish between the root cause of the fault and accompanying phenomena, resulting in low accuracy and efficiency of fault diagnosis results. Therefore, the fault analysis method for overhead low-voltage lines still needs further improvement. Summary of the Invention

[0004] This application aims to solve at least one of the technical problems existing in the prior art; to this end, this application proposes an artificial intelligence-based method and system for fault analysis of overhead low-voltage lines, which is used to solve the technical problem that the prior art often directly uses mixed original electrical characteristics for analysis, lacks coupling division of characteristics and importance assessment, making it difficult for the model to distinguish the root cause of the fault and the accompanying phenomena, resulting in low accuracy and efficiency of fault diagnosis results.

[0005] To achieve the above objectives, the first aspect of this application provides an artificial intelligence-based fault analysis method for overhead low-voltage lines, comprising: Acquire initial multi-source heterogeneous data of overhead low-voltage lines; the initial multi-source heterogeneous data includes line topology data, line electrical data, and meteorological data; Spatiotemporal alignment is performed on initial multi-source heterogeneous data to obtain spatiotemporally aligned multi-source heterogeneous data; the spatiotemporal alignment operation refers to establishing a unified spatiotemporal benchmark and aligning the timestamps and spatial location information of several initial multi-source heterogeneous data to a unified standard. Constructing dynamic feature matrices and dynamic knowledge graphs based on spatiotemporally aligned multi-source heterogeneous data; Fault diagnosis results are obtained by decoupling features and diagnosing faults based on dynamic knowledge graphs and dynamic feature matrices. Fault prediction results are obtained by performing forward-looking fault prediction based on fault diagnosis results and dynamic knowledge graph.

[0006] This application effectively decomposes the original mixed features into several independent feature components through the above steps; then, by introducing an attention mechanism through the power grid topology relationship and real-time equipment operating status contained in the knowledge graph, it dynamically evaluates the importance weight of each feature component in the current scenario, enabling the fault diagnosis model to clearly distinguish the root cause of the fault and the accompanying phenomena. This not only improves the accuracy and response efficiency of fault diagnosis, but also enhances the interpretability of the diagnosis results, providing more reliable and transparent technical support for operation and maintenance decisions.

[0007] Furthermore, the construction of the dynamic feature matrix includes: Line electrical data and line topology data are extracted from spatiotemporally aligned multi-source heterogeneous data within a sliding time window; the line topology data includes device ID, geographical location, and device connection relationship; the line electrical data refers to the time-series data of electrical parameters of several monitoring nodes, including three-phase voltage, current, and power; the time length of the sliding time window is set based on experience. Calculate several dynamic characteristic indicators corresponding to several device IDs; the dynamic characteristic indicators are used to measure the line operating status and anomaly sensitivity, including voltage unbalance, current unbalance, effective value of zero-sequence current and its rate of change, voltage deviation, voltage fluctuation rate, load rate and harmonic distortion rate; A dynamic feature matrix X(N) is constructed based on the device ID, the duration of the sliding time window, and several dynamic feature indicators corresponding to the device ID. i ST, F ij ); where i represents the number corresponding to the device ID, N i Let F represent the ID of the i-th device; ST represents the length of the sliding time window; j represents the number corresponding to several dynamic feature indicators; and F represents the ID of the i-th device. ij It is represented as the j-th dynamic feature index in the i-th device ID.

[0008] Furthermore, the construction of the dynamic knowledge graph includes: Acquire historical fault data; the historical fault data includes several association rules; Extract spatiotemporally aligned multi-source heterogeneous data and dynamic feature matrices; A static graph skeleton is constructed by defining entities and defining edge relationships; the defined entities are several device IDs defined as entities in the graph; the defined edge relationships are several associations between the entities, including electrical connection relationships, physical location relationships, logical inclusion relationships, and management affiliation relationships; Extract real-time dynamic feature indicators from the dynamic feature matrix, as well as line topology data from spatiotemporally aligned multi-source heterogeneous data; Historical fault data, real-time dynamic feature indicators, and line topology data are injected into a static knowledge graph skeleton to obtain a dynamic knowledge graph G(V, E, A); where V represents a set of entities containing several entities; the entities include static attributes and dynamic attributes; the static attributes refer to features that remain unchanged; the dynamic attributes refer to features that change dynamically; E represents a set of edge relationships containing several edge relationships; the edge relationships include association relationships r and relationship weights GQ; A represents an adjacency matrix, which represents the connection relationships between entities.

[0009] This application extracts electrical timing data within a sliding time window and calculates several dynamic feature indicators with clear physical meaning to form a device-level feature matrix, thereby achieving a quantitative description of the operating status. Simultaneously, a static graph skeleton is constructed based on the device connection relationship, and real-time feature indicators, historical fault data, and line topology data are injected as attributes to form a dynamic knowledge graph with weighted edges and state-defined points. This transforms the originally mixed raw data into a quantifiable, associative, and traceable structured knowledge system, providing solid data support for subsequent fault diagnosis. This enables the fault diagnosis model to clearly distinguish the root cause and derivative phenomena of the fault based on physical characteristics and topological relationships, thereby improving the accuracy of fault diagnosis.

[0010] Furthermore, the step of obtaining fault diagnosis results through feature decoupling and fault diagnosis based on dynamic knowledge graphs and dynamic feature matrices includes: Extracting dynamic knowledge graphs and dynamic feature matrices; Extract the global context vector from the dynamic knowledge graph; the global context vector refers to the set of entities in the dynamic knowledge graph whose dynamic attributes are meteorological data. The dynamic feature matrix and the global context vector are processed by a deep feature decoupling network to obtain a decoupled feature vector. The deep feature decoupling network is used to filter out independent feature components with clear physical meaning from the mixed dynamic feature matrix, and to focus on important feature components according to the current global context vector to obtain the decoupled feature vector. The decoupled feature vector is a weighted combination of several feature components with clear physical meaning according to their importance. The fault diagnosis results are obtained by using a spatiotemporal graph neural network model to perform fault reasoning on decoupled feature vectors and dynamic knowledge graphs; the fault diagnosis results include fault points and fault types.

[0011] Furthermore, the deep feature decoupling network consists of a feature encoder, several feature decoders, and a context-aware attention fusion module; The feature encoder is used to extract high-level features from the dynamic feature matrix; The aforementioned feature decoders are executed in parallel to extract specific feature components from the high-level features; The context-aware attention fusion module is used to calculate the attention weights between the global context vector and several feature components through a multi-head attention mechanism, and to perform weighted fusion of several feature components and their corresponding attention weights to obtain a decoupled feature vector.

[0012] Furthermore, the spatiotemporal graph neural network model consists of a spatiotemporal alternating convolution module and a result output layer; wherein, the spatiotemporal alternating convolution module consists of a temporal convolution layer and a multi-relation graph convolution layer, and its input data are decoupled feature vectors and dynamic knowledge graphs; the output data are spatiotemporal features corresponding to several entities; The temporal convolutional layer is used to extract temporal features from the decoupled feature vector of the entity to obtain the temporal features corresponding to the entity; wherein, the decoupled feature vector of the entity is data with a time series, which is consistent with the time length of the sliding time window in the dynamic feature matrix; The multi-relation graph convolutional layer is used to extract spatial features from several relationships of an entity to obtain the spatial features corresponding to the entity. The output layer consists of several fully connected layers, which are used to perform fault identification on the spatiotemporal features of several entities to obtain the fault confidence of the entities, as well as the fault type and its corresponding confidence; entities with fault confidence greater than the fault threshold are identified as fault points; and fault types with fault type confidence greater than the type confidence threshold are identified as the fault types corresponding to the fault points.

[0013] Furthermore, the spatial features satisfy: ;in, Represented as the trainable relation importance weights under the relation r; This is represented as the neighbor aggregation feature corresponding to the i-th entity under the association relationship r; Represented as the sigmoid function; This is represented as the feature corresponding to the i-th entity; This is represented as the spatial feature corresponding to the i-th entity; The neighbor aggregation feature satisfies: ;in, Represented as the normalized attention coefficient; and All are represented as trainable parameter matrices; The feature corresponding to the nth entity in the set of neighboring entities corresponding to the i-th entity is represented by the normalized attention coefficients, which satisfy the following: Where m represents the set of neighboring entities of the i-th entity under the association r. Entity number in It is represented as the attention coefficient between the i-th entity and the n-th entity under the association relationship r; The function is a ReLU activation function, used to introduce non-linearity so that the attention coefficients can better capture the relationships between entities; the attention coefficients satisfy: ;in, This represents the relationship weight between the i-th and n-th entities under the association relationship r; T represents the matrix transpose notation. This is represented as the attention vector corresponding to the association r. This is represented as a feature concatenation operation; It is represented as a trainable scalar used to adjust the contribution of relation weights to attention; The features satisfy: Where R represents the set of associations r, and i and n both represent entity numbers. Let represent the set of neighboring entities of the i-th entity under the association r; denoted as the normalized attention coefficient, and l represents the convolutional layer number.

[0014] This application effectively extracts temporal features from the decoupled feature vectors of various power grid entities through temporal convolutional layers. Simultaneously, it utilizes multi-relationship graph convolutional layers to differentiate the heterogeneous relationships within the power grid, including electrical connections, topological adjacencies, and functional dependencies. A relationship-specific attention mechanism is introduced to dynamically evaluate the importance weights of different relationship types and their corresponding neighboring entities under the current operating state. These relationship-aware weights are deeply integrated with entity features to achieve adaptive aggregation of spatial information for multi-relationship networks. This accurately characterizes the propagation behavior and impact range of faults along specific topological paths. Fine-grained fault diagnosis results are output through fully connected layers, enabling refined modeling of the complex spatiotemporal relationships of the power grid. This allows the fault diagnosis model to closely integrate physical connection structures with real-time operating states, clearly distinguishing the fault source and its derivative propagation effects in the multi-relationship network, significantly improving the accuracy, robustness, and interpretability of fault location and classification.

[0015] Furthermore, the fault prediction result obtained by performing prospective fault prediction based on fault diagnosis results and dynamic knowledge graph includes: Extract fault diagnosis results and dynamic knowledge graphs; the fault diagnosis results include fault points and fault types. Set the fault labels corresponding to several entities with fault points as explicit faults; Set the fault labels of several entities that do not have fault points to "no fault"; The fault type corresponding to the entity with the fault label as explicit fault, the decoupled feature vector corresponding to several historical time points, and meteorological data are integrated into explicit analysis data. The decoupled feature vectors and meteorological data corresponding to several historical time points of entities with fault labels as fault-free are integrated into implicit analysis data. The dynamic knowledge graph and the corresponding analysis data of the entities are input into the multi-task fault prediction model to obtain the fault prediction results of the entities; the analysis data includes explicit analysis data and implicit analysis data; the multi-task fault prediction model is constructed by an artificial intelligence model and is used to predict the fault evolution of entities with known faults and to predict the faults of entities with unknown faults; in this embodiment, the fault prediction results include the fault deterioration time, fault propagation probability and fault deterioration rate of entities with explicit faults, and the fault occurrence prediction time and fault occurrence probability of entities without faults.

[0016] Furthermore, the multi-task fault prediction model consists of an input layer, a feature extraction layer, a multi-task learning layer, and an output layer; The input layer is used to receive dynamic knowledge graphs and analysis data corresponding to entities; The feature extraction layer is used to extract features from a dynamic knowledge graph with analytical data to obtain comprehensive features corresponding to entities; the feature extraction layer is constructed by the spatiotemporal alternating convolution module in the spatiotemporal graph neural network model; The multi-task learning layer includes two task branches: an explicit prediction task branch and a fault-free prediction task branch. By using the fully connected network corresponding to the two task branches, the comprehensive features of the entity are used to perform evolution prediction for known faults and prediction for fault-free conditions. The output layer is used to output the results of the two task branches; The fully connected network corresponding to the explicit prediction task branch includes a time prediction subnetwork, a propagation probability prediction subnetwork, and a deterioration rate prediction subnetwork. The loss function of the fully connected network corresponding to the explicit prediction task branch is a weighted sum of the loss functions of the time prediction subnetwork, the propagation probability prediction subnetwork, and the deterioration rate prediction subnetwork. The fully connected network corresponding to the fault-free prediction task branch includes a fault probability prediction sub-network and a fault time prediction sub-network. The loss function of the fully connected network corresponding to the fault-free prediction task branch is a weighted sum of the loss functions of the fault probability prediction subnetwork and the fault time prediction subnetwork. The loss function of the multi-task fault prediction model is a weighted sum of the loss functions corresponding to the explicit prediction task branch and the fault-free prediction task branch.

[0017] This application classifies entities into two categories based on fault diagnosis results: explicit faults and non-faults. It then integrates their corresponding historical characteristics, fault types, and meteorological data to form differentiated inputs. A shared feature extraction layer extracts comprehensive features from a dynamic knowledge graph and analytical data. Subsequently, two task branches are executed in parallel within a multi-task learning layer: the explicit prediction branch uses three sub-networks—time, propagation probability, and deterioration rate—to perform fine-grained predictions of the evolution trajectory of known faults; the non-fault prediction branch uses a fault probability and fault time sub-network to provide probabilistic early warnings of potential risks. This unifies the simulation of the evolution of existing faults with risk warnings of potential faults, enabling the overhead low-voltage line fault analysis method to not only predict the development path and impact range of known faults but also identify hidden risks in currently normal equipment. This achieves a leap from passive diagnosis to proactive early warning, enhancing the predictability and proactivity of power grid operation and maintenance.

[0018] Another aspect of the present invention provides an artificial intelligence-based fault analysis system for overhead low-voltage lines, comprising: a data acquisition module and a data analysis module; the data acquisition module and the data analysis module are connected together; The data acquisition module acquires initial multi-source heterogeneous data of overhead low-voltage lines through data acquisition equipment; the initial multi-source heterogeneous data includes line topology data, line electrical data, and meteorological data. The data analysis module includes a feature construction unit, a fault diagnosis unit, and a fault prediction unit; The feature construction unit performs a spatiotemporal alignment operation on the initial multi-source heterogeneous data to obtain spatiotemporally aligned multi-source heterogeneous data; and constructs a dynamic feature matrix and a dynamic knowledge graph based on the spatiotemporally aligned multi-source heterogeneous data. The fault diagnosis unit: performs feature decoupling and fault diagnosis based on dynamic knowledge graph and dynamic feature matrix to obtain fault diagnosis results; The fault prediction unit obtains fault prediction results by performing forward-looking fault prediction based on fault diagnosis results and dynamic knowledge graph.

[0019] Compared with the prior art, the beneficial effects of this application are: 1. This application obtains spatiotemporally aligned multi-source heterogeneous data by performing spatiotemporal alignment on initial multi-source heterogeneous data; constructs a dynamic feature matrix and a dynamic knowledge graph based on the spatiotemporally aligned multi-source heterogeneous data; performs feature decoupling and fault diagnosis based on the dynamic knowledge graph and dynamic feature matrix to obtain fault diagnosis results; and performs fault prospective prediction based on the fault diagnosis results and the dynamic knowledge graph to obtain fault prediction results. A deep feature decoupling network is used to separate the original mixed features into independent feature components, and the importance of each feature component is dynamically evaluated through an attention mechanism, combined with the topological relationships and real-time status in the knowledge graph. This enables the fault diagnosis model to clearly distinguish the root cause of the fault and the accompanying phenomena, thereby improving the accuracy and efficiency of fault diagnosis and enhancing the interpretability of the diagnosis results.

[0020] 2. This application utilizes a feature encoder to extract high-level features from a dynamic feature matrix and decomposes it into independent feature components with clear physical meaning through multiple parallel decoders. It introduces a global context vector from a dynamic knowledge graph and calculates the correlation weights between this context and each feature component via a multi-head attention mechanism, thereby dynamically weighting and fusing the features to obtain a decoupled feature vector focusing on the current main risk factors. The decoupled feature vector and the dynamic knowledge graph are then processed through a spatiotemporal graph neural network, utilizing the topological relationships within the graph for information propagation and reasoning, accurately outputting fault points and fault types. This solves the problem of unclear and mixed original features, dynamically focusing on key features based on the scenario, and improving the accuracy and efficiency of fault diagnosis.

[0021] 3. This application extracts temporal features from the decoupled feature vectors of entities through temporal convolutional layers, and uses multi-relationship graph convolutional layers to differentiate the various relationships in the power grid. By introducing a relationship-specific attention mechanism, it dynamically calculates the importance weights of different relationship types and different neighboring entities, and integrates relationship weights and entity features to aggregate spatial information, accurately capturing the propagation path of faults under specific topological relationships. The diagnostic results are output through fully connected layers, realizing a fine characterization of the complex spatiotemporal relationships of the power grid. This enables the fault diagnosis model to clearly distinguish the fault source and its propagation effect in the multi-relationship network based on physical connections and real-time status, improving the accuracy and reliability of fault location and classification. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1This is a flowchart of an artificial intelligence-based fault analysis method for overhead low-voltage lines according to this application. Figure 2 This is a schematic diagram illustrating the principle of an artificial intelligence-based fault analysis system for overhead low-voltage lines according to this application. Detailed Implementation

[0024] The technical solutions of this application will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0025] Please see Figure 1 The first aspect of this application provides an artificial intelligence-based method for fault analysis of overhead low-voltage lines, including: Acquire initial multi-source heterogeneous data of overhead low-voltage lines; the initial multi-source heterogeneous data includes line topology data, line electrical data, and meteorological data; Spatiotemporal alignment is performed on initial multi-source heterogeneous data to obtain spatiotemporally aligned multi-source heterogeneous data. Spatiotemporal alignment refers to establishing a unified spatiotemporal benchmark and aligning the timestamps and spatial location information of several initial multi-source heterogeneous data to a unified standard. Constructing dynamic feature matrices and dynamic knowledge graphs based on spatiotemporally aligned multi-source heterogeneous data; Fault diagnosis results are obtained by decoupling features and diagnosing faults based on dynamic knowledge graphs and dynamic feature matrices. Fault prediction results are obtained by performing forward-looking fault prediction based on fault diagnosis results and dynamic knowledge graph.

[0026] The construction of the dynamic feature matrix in this embodiment includes: Line electrical data and line topology data are extracted from spatiotemporally aligned multi-source heterogeneous data within a sliding time window. Line topology data includes device ID, geographical location, and device connection relationship. Line electrical data refers to the time series data of electrical parameters of several monitoring nodes, including three-phase voltage, current, and power. The time length of the sliding time window is set based on experience. In this embodiment, the time length of the sliding time window is set to 1 hour. Calculate several dynamic characteristic indicators corresponding to several device IDs; dynamic characteristic indicators are used to measure the line operating status and anomaly sensitivity, including voltage unbalance, current unbalance, effective value of zero-sequence current and its rate of change, voltage deviation, voltage fluctuation rate, load rate and harmonic distortion rate; A dynamic feature matrix X(N) is constructed based on the device ID, the duration of the sliding time window, and several dynamic feature indicators corresponding to the device ID. i ST, Fij ); where i represents the number corresponding to the device ID, N i Let F represent the ID of the i-th device; ST represents the length of the sliding time window; j represents the number corresponding to several dynamic feature indicators; and F represents the ID of the i-th device. ij It is represented as the j-th dynamic feature index in the i-th device ID.

[0027] The construction of the dynamic knowledge graph in this embodiment includes: Obtain historical fault data; historical fault data includes several association rules; in this embodiment, the association rules are fault mode-device type-environmental conditions association rules; Extract spatiotemporally aligned multi-source heterogeneous data and dynamic feature matrices; A static graph skeleton is constructed by defining entities and edge relationships. Defining entities involves defining several device IDs as entities in the graph. Defining edge relationships involves defining several associations between entities, including electrical connection relationships, physical location relationships, logical inclusion relationships, and management attribution relationships. In this embodiment, entities include transformers, line segments, poles, circuit breakers, and fuses, etc. Electrical connection relationships include supplying power to, connecting to, and belonging to, etc. Physical location relationships include being located at and adjacent to, etc. Logical inclusion relationships include containing and belonging to, etc. Management attribution relationships include governing and being responsible for, etc. Extract real-time dynamic feature indicators from the dynamic feature matrix, as well as line topology data from spatiotemporally aligned multi-source heterogeneous data; Historical fault data, real-time dynamic feature indicators, and line topology data are injected into a static knowledge graph skeleton to obtain a dynamic knowledge graph G(V, E, A). Here, V represents a set of entities, containing several entities; entities include static and dynamic attributes; static attributes refer to features that remain unchanged, including entity ID and location, etc.; dynamic attributes refer to features that change dynamically, including several dynamic feature indicators and meteorological data, etc.; E represents a set of edge relationships, containing several edge relationships; edge relationships include association relationship r and relationship weight GQ; A represents an adjacency matrix, representing the connection relationships between entities. In this embodiment, the relationship weight varies depending on the association relationship. When two entities are adjacent towers, wind speed is used as the relationship weight; when two entities are transformers and power supply units of a line segment, load rate is used as the relationship weight. In this embodiment, the relationship weight GQ refers to the set of relationship weights gq between several entities.

[0028] This embodiment extracts electrical timing data within a sliding time window, calculates several dynamic feature indicators with clear physical meaning, and constructs a device-level feature matrix to achieve a refined and quantitative characterization of the operating status of power distribution equipment. Simultaneously, a static graph skeleton is constructed based on the electrical connection relationships between devices, and real-time feature indicators, historical fault records, and line topology information are injected as attributes of nodes and edges, forming a dynamic knowledge graph where nodes contain states and edges carry weights. This effectively transforms the original, mixed, and unstructured data into a quantifiable, associative, and traceable structured knowledge system, providing a high-quality, semantically rich data foundation for subsequent fault diagnosis. The fault diagnosis model can fully integrate physical features and topological relationships, accurately distinguishing the root cause of the fault from its derivative phenomena, significantly improving the accuracy and reliability of the diagnosis.

[0029] In this embodiment, the fault diagnosis results obtained by feature decoupling and fault diagnosis based on dynamic knowledge graphs and dynamic feature matrices include: Extracting dynamic knowledge graphs and dynamic feature matrices; Extract the global context vector from the dynamic knowledge graph; the global context vector refers to the set of entities in the dynamic knowledge graph whose dynamic attributes are meteorological data. The dynamic feature matrix and global context vector are passed through a deep feature decoupling network to obtain a decoupled feature vector. The deep feature decoupling network is used to filter out independent feature components with clear physical meaning from the mixed dynamic feature matrix, and to focus on important feature components according to the current global context vector to obtain the decoupled feature vector. The decoupled feature vector is a weighted combination of several feature components with clear physical meaning according to their importance. The fault diagnosis results are obtained by using a spatiotemporal graph neural network model to perform fault reasoning on decoupled feature vectors and dynamic knowledge graphs; the fault diagnosis results include fault points and fault types.

[0030] The deep feature decoupling network in this embodiment consists of a feature encoder, several feature decoders, and a context-aware attention fusion module; The feature encoder is used to extract high-level features from the dynamic feature matrix; in this embodiment, the feature encoder is a shared feature encoder, which adopts the Transformer encoder. Several feature decoders are executed in parallel to extract specific feature components from high-level features. In this embodiment, the feature decoders include electrical decoders and load decoders, etc. The electrical decoder reconstructs and outputs feature components that include, but are not limited to, pure electrical feature indicators such as voltage imbalance and current imbalance, and the load decoder reconstructs and outputs feature components that include, but are not limited to, pure load feature indicators such as load rate. The context-aware attention fusion module is used to calculate the attention weights between the global context vector and several feature components through a multi-head attention mechanism, and to perform weighted fusion of several feature components and their corresponding attention weights to obtain a decoupled feature vector.

[0031] This embodiment utilizes a feature encoder to extract high-order semantic features from a dynamic feature matrix and decomposes them into several independent feature components with clear physical meanings through multiple parallel decoders. Subsequently, the global context vector contained in the dynamic knowledge graph is introduced, and the correlation weight between the context and each feature component is calculated using a multi-head attention mechanism to achieve dynamic weighting and adaptive fusion of the feature components, thereby generating a decoupled feature vector focusing on the current dominant risk factor. This decoupled feature vector and the dynamic knowledge graph are jointly input into a spatiotemporal graph neural network, and cross-node information propagation and causal reasoning are carried out based on the topological structure in the graph. Finally, the fault location is accurately located and the fault type is identified, overcoming the problems of mixed original features and semantic ambiguity. It can dynamically focus on key features according to the operating scenario, significantly improving the accuracy, efficiency and robustness of fault diagnosis.

[0032] The spatiotemporal graph neural network model in this embodiment consists of a spatiotemporal alternating convolution module and a result output layer. The spatiotemporal alternating convolution module comprises a temporal convolution layer and a multi-relation graph convolution layer. Its input data are decoupled feature vectors and a dynamic knowledge graph; the output data are spatiotemporal features corresponding to several entities. In this embodiment, the temporal convolution layer processes the time series of decoupled feature vectors corresponding to several entities in the dynamic knowledge graph. Subsequently, the dynamic knowledge graph is spatially processed by the multi-relation graph convolution layer to obtain spatiotemporal features. The temporal convolutional layer is used to extract temporal features from the decoupled feature vector of an entity to obtain the temporal features corresponding to the entity; wherein, the decoupled feature vector of the entity is data with a time series, which is consistent with the time length of the sliding time window in the dynamic feature matrix; in this embodiment, the temporal convolutional layer uses the TCN method to extract temporal features. Multi-relation graph convolutional layers are used to extract spatial features from several relationships between entities to obtain the spatial features corresponding to the entities; The output layer consists of several fully connected layers, used to perform fault identification on the spatiotemporal features of several entities to obtain the fault confidence of the entities, as well as the fault type and its corresponding confidence. Entities with a fault confidence greater than a fault threshold are identified as fault points; fault types with a fault type confidence greater than a type confidence threshold are identified as the fault types corresponding to the fault points. In this embodiment, the sigmoid activation function is used for the fully connected layer corresponding to the fault confidence of fault location, and the softmax activation function is used for the fully connected layer corresponding to the fault type of fault classification.

[0033] The spatial characteristics in this embodiment satisfy: ;in, This is represented as the trainable relation importance weight under the relation r, which can be calculated through a global attention mechanism in this embodiment; This is represented as the neighbor aggregation feature corresponding to the i-th entity under the association relationship r; Represented as the sigmoid function; This is represented as the feature corresponding to the i-th entity; This is represented as the spatial feature corresponding to the i-th entity; Neighbor aggregation features satisfy: ;in, Represented as the normalized attention coefficient; and All are represented as trainable parameter matrices; Let the feature be the feature corresponding to the nth entity in the set of neighboring entities corresponding to the i-th entity; the normalized attention coefficients satisfy: Where m represents the set of neighboring entities of the i-th entity under the association r. Entity number in It is represented as the attention coefficient between the i-th entity and the n-th entity under the association relationship r; The function is a ReLU activation function, used to introduce non-linearity so that the attention coefficients can better capture the relationships between entities; the attention coefficients satisfy: ;in, This represents the relationship weight between the i-th and n-th entities under the association relationship r; T represents the matrix transpose notation. This is represented as the attention vector corresponding to the association r. This is represented as a feature concatenation operation; It is represented as a trainable scalar used to adjust the contribution of relation weights to attention; feature satisfy: Where R represents the set of associations r, and i and n both represent entity numbers. Let represent the set of neighboring entities of the i-th entity under the association r; denoted as the normalized attention coefficient, and l represents the convolutional layer number.

[0034] In this embodiment, the fault prediction result obtained by performing forward-looking fault prediction based on fault diagnosis results and dynamic knowledge graph includes: Extract fault diagnosis results and dynamic knowledge graphs; fault diagnosis results include fault points and fault types. Set the fault labels corresponding to several entities with fault points as explicit faults; Set the fault labels of several entities that do not have fault points to "no fault"; The fault type corresponding to the entity with the fault label as explicit fault, the decoupled feature vector corresponding to several historical time points, and meteorological data are integrated into explicit analysis data. The decoupled feature vectors and meteorological data corresponding to several historical time points of entities with fault labels as fault-free are integrated into implicit analysis data. The dynamic knowledge graph and the corresponding analysis data of the entities are input into the multi-task fault prediction model to obtain the fault prediction results of the entities; the analysis data includes explicit analysis data and implicit analysis data; the multi-task fault prediction model is constructed through an artificial intelligence model and is used to predict the fault evolution of entities with known faults and to predict the faults of entities with unknown faults.

[0035] The multi-task fault prediction model in this embodiment consists of an input layer, a feature extraction layer, a multi-task learning layer, and an output layer. The input layer is used to receive dynamic knowledge graphs and corresponding entity analysis data; The feature extraction layer is used to extract features from the dynamic knowledge graph with analysis data to obtain the comprehensive features corresponding to the entities; the feature extraction layer is constructed by the spatiotemporal alternating convolution module in the spatiotemporal graph neural network model; in this embodiment, the comprehensive features and the spatiotemporal features output by the spatiotemporal alternating convolution module are consistent in nature, and the comprehensive features are data with spatiotemporal feature properties; The multi-task learning layer includes two task branches: an explicit prediction task branch and a fault-free prediction task branch. By using the fully connected network corresponding to the two task branches, the evolution prediction of known faults and the prediction of fault-free status are performed on the comprehensive features of the entity. The output layer is used to output the results of the two task branches; The fully connected network corresponding to the explicit prediction task branch includes a time prediction subnetwork, a propagation probability prediction subnetwork, and a deterioration rate prediction subnetwork; The loss function of the fully connected network corresponding to the explicit prediction task branch is a weighted sum of the loss functions of the time prediction subnetwork, the propagation probability prediction subnetwork, and the deterioration rate prediction subnetwork. In this embodiment, the loss functions of the time prediction subnetwork, the propagation probability prediction subnetwork, and the deterioration rate prediction subnetwork are respectively Huber loss, binary cross-entropy loss, and mean squared error loss. The fully connected network corresponding to the fault-free prediction task branch includes a fault probability prediction subnetwork and a fault time prediction subnetwork. The loss function of the fully connected network corresponding to the fault-free prediction task branch is a weighted sum of the loss functions of the fault probability prediction subnetwork and the fault time prediction subnetwork. In this embodiment, the loss functions of the fault probability prediction subnetwork and the fault time prediction subnetwork are cross-entropy loss and negative log-likelihood loss, respectively. The loss function of the multi-task fault prediction model is a weighted sum of the loss functions corresponding to the explicit prediction task branch and the fault-free prediction task branch.

[0036] Based on the fault diagnosis results, this embodiment divides power grid entities into two categories: explicit faults and no faults. It integrates corresponding historical operating characteristics, fault types, meteorological environment, and other multi-source information to construct differentiated input representations. Through a shared feature extraction layer, it fuses and extracts comprehensive features with spatiotemporal semantics from dynamic knowledge graphs and multi-dimensional analysis data. A multi-task learning architecture is designed using two parallel task branches: the explicit fault prediction branch, by fusing three sub-networks—time evolution, fault propagation probability, and deterioration rate—performs fine-grained modeling and prediction of the subsequent development trajectory, impact range, and severity of identified faults; the no-fault risk early warning branch, through two sub-networks—fault occurrence probability and potential fault time—performs probabilistic assessment and early warning of latent risks for currently normal equipment. This achieves an organic unity between dynamic evolution simulation of existing faults and proactive risk identification of potential faults, breaking through the limitations of traditional fault analysis which is confined to post-event diagnosis. It promotes the transformation of overhead low-voltage line fault analysis from passive response to proactive prevention, significantly enhancing the predictability, proactivity, and intelligence level of power grid operation and maintenance.

[0037] Please see Figure 2 Another embodiment of this application provides an artificial intelligence-based overhead low-voltage line fault analysis system, including: a data acquisition module and a data analysis module; the data acquisition module and the data analysis module are electrically and / or communicatively connected; Data acquisition module: Acquires initial multi-source heterogeneous data of overhead low-voltage lines through data acquisition equipment; the initial multi-source heterogeneous data includes line topology data, line electrical data, and meteorological data; the data acquisition equipment includes several sensors, etc. The data analysis module includes a feature construction unit, a fault diagnosis unit, and a fault prediction unit; Feature construction unit: Spatiotemporal alignment is performed on the initial multi-source heterogeneous data to obtain spatiotemporally aligned multi-source heterogeneous data; dynamic feature matrix and dynamic knowledge graph are constructed based on the spatiotemporally aligned multi-source heterogeneous data; Fault diagnosis unit: Based on dynamic knowledge graph and dynamic feature matrix, feature decoupling and fault diagnosis are performed to obtain fault diagnosis results; Fault prediction unit: Based on fault diagnosis results and dynamic knowledge graph, it performs forward-looking fault prediction to obtain fault prediction results.

[0038] Some of the data in the above formula are calculated by removing dimensions and taking their numerical values. The formula is the closest to the real situation obtained by software simulation of a large amount of collected data. The preset parameters and preset thresholds in the formula are set by those skilled in the art according to the actual situation or obtained through simulation of a large amount of data.

[0039] The working principle of this application is as follows: Initial multi-source heterogeneous data of overhead low-voltage lines is acquired; spatiotemporal alignment is performed on the initial multi-source heterogeneous data to obtain spatiotemporally aligned multi-source heterogeneous data; a dynamic feature matrix and a dynamic knowledge graph are constructed based on the spatiotemporally aligned multi-source heterogeneous data; feature decoupling and fault diagnosis are performed based on the dynamic knowledge graph and dynamic feature matrix to obtain fault diagnosis results; and fault prediction results are obtained based on the fault diagnosis results and the dynamic knowledge graph. A deep feature decoupling network is used to separate the original mixed features into independent feature components. Combined with the topological relationships and real-time status in the knowledge graph, the importance of each feature component is dynamically evaluated through an attention mechanism. This enables the fault diagnosis model to clearly distinguish the root cause of the fault from accompanying phenomena, thereby improving the accuracy and efficiency of fault diagnosis. Simultaneously, it enhances the interpretability of the diagnosis results, avoiding the problem that existing technologies often directly use mixed original electrical features for analysis, lacking feature coupling and importance assessment, making it difficult for the model to distinguish the root cause of the fault from accompanying phenomena, resulting in low accuracy and efficiency of fault diagnosis results.

[0040] The above embodiments are only used to illustrate the technical methods of this application and are not intended to limit it. Although this application 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 methods of this application without departing from the spirit and scope of the technical methods of this application.

Claims

1. A fault analysis method for overhead low-voltage lines based on artificial intelligence, characterized in that, include: Acquire initial multi-source heterogeneous data of overhead low-voltage lines; The initial multi-source heterogeneous data includes line topology data, line electrical data, and meteorological data; Spatiotemporal alignment is performed on the initial multi-source heterogeneous data to obtain spatiotemporally aligned multi-source heterogeneous data. The spatiotemporal alignment operation refers to establishing a unified spatiotemporal benchmark and aligning the timestamps and spatial location information of several initial multi-source heterogeneous data to a unified standard. Constructing dynamic feature matrices and dynamic knowledge graphs based on spatiotemporally aligned multi-source heterogeneous data; Fault diagnosis results are obtained by decoupling features and diagnosing faults based on dynamic knowledge graphs and dynamic feature matrices. Fault prediction results are obtained by performing forward-looking fault prediction based on fault diagnosis results and dynamic knowledge graph. The fault diagnosis results obtained by feature decoupling and fault diagnosis based on dynamic knowledge graphs and dynamic feature matrices include: Extracting dynamic knowledge graphs and dynamic feature matrices; Extract the global context vector from the dynamic knowledge graph; the global context vector refers to the set of entities in the dynamic knowledge graph whose dynamic attributes are meteorological data. The dynamic feature matrix and the global context vector are passed through a deep feature decoupling network to obtain a decoupled feature vector. The deep feature decoupling network is used to filter out independent feature components with clear physical meaning from the mixed dynamic feature matrix, and to focus on important feature components based on the current global context vector to obtain the decoupled feature vector. The fault diagnosis result is obtained by using a spatiotemporal graph neural network model to perform fault reasoning on decoupled feature vectors and dynamic knowledge graphs; the fault diagnosis result includes the fault point and fault type. The deep feature decoupling network consists of a feature encoder, several feature decoders, and a context-aware attention fusion module. The feature encoder is used to extract high-level features from the dynamic feature matrix; The aforementioned feature decoders are executed in parallel to extract specific feature components from the high-level features; The context-aware attention fusion module is used to calculate the attention weights between the global context vector and several feature components through a multi-head attention mechanism, and to perform weighted fusion of several feature components and their corresponding attention weights to obtain a decoupled feature vector.

2. The method for fault analysis of overhead low-voltage lines based on artificial intelligence according to claim 1, characterized in that, The construction of the dynamic feature matrix includes: Extract line electrical data and line topology data from spatiotemporally aligned multi-source heterogeneous data within a sliding time window; the line topology data includes device ID, geographical location, and device connection relationship; the line electrical data refers to the time-series data of electrical parameters of several monitoring nodes, including three-phase voltage, current, and power; Calculate several dynamic characteristic indicators corresponding to several device IDs; the dynamic characteristic indicators are used to measure the line operating status and anomaly sensitivity, including voltage unbalance, current unbalance, effective value of zero-sequence current and its rate of change, voltage deviation, voltage fluctuation rate, load rate and harmonic distortion rate; A dynamic feature matrix X(N) is constructed based on the device ID, the duration of the sliding time window, and several dynamic feature indicators corresponding to the device ID. i ST, F ij Where i represents the device ID, and N is the device number. i Let F represent the ID of the i-th device; ST represents the length of the sliding time window; j represents the number corresponding to several dynamic feature indicators; and F represents the ID of the i-th device. ij It is represented as the j-th dynamic feature index in the i-th device ID.

3. The method for fault analysis of overhead low-voltage lines based on artificial intelligence according to claim 1, characterized in that, The construction of the dynamic knowledge graph includes: Acquire historical fault data; the historical fault data includes several association rules; Extract spatiotemporally aligned multi-source heterogeneous data and dynamic feature matrices; A static graph skeleton is constructed by defining entities and defining edge relationships; the defined entities are several device IDs defined as entities in the graph; the defined edge relationships are several associations between the entities, including electrical connection relationships, physical location relationships, logical inclusion relationships, and management affiliation relationships; Extract real-time dynamic feature indicators from the dynamic feature matrix, as well as line topology data from spatiotemporally aligned multi-source heterogeneous data; Historical fault data, real-time dynamic feature indicators, and line topology data are injected into a static graph skeleton to obtain a dynamic knowledge graph G(V, E, A); where V represents a set of entities containing several entities; the entities include static attributes and dynamic attributes; E represents a set of edge relationships containing several edge relationships; the edge relationships include association relationship r and relationship weight GQ; and A represents an adjacency matrix, which represents the connection relationships between entities.

4. The method for fault analysis of overhead low-voltage lines based on artificial intelligence according to claim 1, characterized in that, The spatiotemporal graph neural network model consists of a spatiotemporal alternating convolution module and a result output layer; wherein, the spatiotemporal alternating convolution module consists of a temporal convolution layer and a multi-relation graph convolution layer; The temporal convolutional layer is used to extract temporal features from the decoupled feature vector of the entity to obtain the temporal features corresponding to the entity; The multi-relation graph convolutional layer is used to extract spatial features from several relationships of an entity to obtain the spatial features corresponding to the entity. The output layer consists of several fully connected layers, which are used to perform fault identification on the spatiotemporal features of several entities to obtain the fault confidence of the entities, as well as the fault type and its corresponding confidence; entities with fault confidence greater than the fault threshold are identified as fault points; and fault types with fault type confidence greater than the type confidence threshold are identified as the fault types corresponding to the fault points.

5. The method for fault analysis of overhead low-voltage lines based on artificial intelligence according to claim 4, characterized in that, The spatial features satisfy: ;in, Represented as the trainable relation importance weights under the relation r; This is represented as the neighbor aggregation feature corresponding to the i-th entity under the association relationship r; Represented as the sigmoid function; This is represented as the feature corresponding to the i-th entity; This is represented as the spatial feature corresponding to the i-th entity; The neighbor aggregation feature satisfies: ;in, Represented as the normalized attention coefficient; and All are represented as trainable parameter matrices; The feature corresponding to the nth entity in the set of neighboring entities corresponding to the i-th entity is represented by the normalized attention coefficients, which satisfy the following: Where m represents the set of neighboring entities of the i-th entity under the association r. Entity number in It is represented as the attention coefficient between the i-th entity and the n-th entity under the association relationship r; The function is a ReLU activation function; the attention coefficients satisfy: ;in, This represents the relationship weight between the i-th and n-th entities under the association relationship r; T represents the matrix transpose notation. This is represented as the attention vector corresponding to the association r. This is represented as a feature concatenation operation; Represented as a trainable scalar; The features satisfy: Where R represents the set of associations r, and i and n both represent entity numbers. Let represent the set of neighboring entities of the i-th entity under the association r; denoted as the normalized attention coefficient, and l represents the convolutional layer number.

6. The method for fault analysis of overhead low-voltage lines based on artificial intelligence according to claim 1, characterized in that, The fault prediction results obtained by performing forward-looking fault prediction based on fault diagnosis results and dynamic knowledge graphs include: Extract fault diagnosis results and dynamic knowledge graphs; the fault diagnosis results include fault points and fault types. Set the fault labels corresponding to several entities with fault points as explicit faults; Set the fault labels of several entities that do not have fault points to "no fault"; The fault type corresponding to the entity with the fault label as explicit fault, the decoupled feature vector corresponding to several historical time points, and meteorological data are integrated into explicit analysis data. The decoupled feature vectors and meteorological data corresponding to several historical time points of entities with fault labels as fault-free are integrated into implicit analysis data. The dynamic knowledge graph and the corresponding analysis data of the entities are input into the multi-task fault prediction model to obtain the fault prediction results of the entities; the analysis data includes explicit analysis data and implicit analysis data; the multi-task fault prediction model is constructed by an artificial intelligence model and is used to predict the fault evolution of entities with known faults and to predict the faults of entities with unknown faults.

7. The method for fault analysis of overhead low-voltage lines based on artificial intelligence according to claim 6, characterized in that, The multi-task fault prediction model consists of an input layer, a feature extraction layer, a multi-task learning layer, and an output layer. The input layer is used to receive dynamic knowledge graphs and analysis data corresponding to entities; The feature extraction layer is used to extract features from a dynamic knowledge graph with analytical data to obtain comprehensive features corresponding to entities; the feature extraction layer is constructed by the spatiotemporal alternating convolution module in the spatiotemporal graph neural network model; The multi-task learning layer includes two task branches: an explicit prediction task branch and a fault-free prediction task branch. By using the fully connected network corresponding to the two task branches, the comprehensive features of the entity are used to perform evolution prediction for known faults and prediction for fault-free conditions. The output layer is used to output the results of the two task branches; The fully connected network corresponding to the explicit prediction task branch includes a time prediction subnetwork, a propagation probability prediction subnetwork, and a deterioration rate prediction subnetwork. The fully connected network corresponding to the fault-free prediction task branch includes a fault probability prediction sub-network and a fault time prediction sub-network. The loss function of the multi-task fault prediction model is a weighted sum of the loss functions corresponding to the explicit prediction task branch and the fault-free prediction task branch.

8. An artificial intelligence-based fault analysis system for overhead low-voltage lines, characterized in that, include: Interconnected data acquisition and data analysis modules; The data acquisition module acquires initial multi-source heterogeneous data of overhead low-voltage lines through data acquisition equipment. The initial multi-source heterogeneous data includes line topology data, line electrical data, and meteorological data; The data analysis module includes a feature construction unit, a fault diagnosis unit, and a fault prediction unit; The feature construction unit performs a spatiotemporal alignment operation on the initial multi-source heterogeneous data to obtain spatiotemporally aligned multi-source heterogeneous data. Constructing dynamic feature matrices and dynamic knowledge graphs based on spatiotemporally aligned multi-source heterogeneous data; The fault diagnosis unit: performs feature decoupling and fault diagnosis based on dynamic knowledge graphs and dynamic feature matrices to obtain fault diagnosis results; the process of performing feature decoupling and fault diagnosis based on dynamic knowledge graphs and dynamic feature matrices to obtain fault diagnosis results includes: Extracting dynamic knowledge graphs and dynamic feature matrices; Extract the global context vector from the dynamic knowledge graph; the global context vector refers to the set of entities in the dynamic knowledge graph whose dynamic attributes are meteorological data. The dynamic feature matrix and the global context vector are passed through a deep feature decoupling network to obtain a decoupled feature vector. The deep feature decoupling network is used to filter out independent feature components with clear physical meaning from the mixed dynamic feature matrix, and to focus on important feature components based on the current global context vector to obtain the decoupled feature vector. The fault diagnosis result is obtained by using a spatiotemporal graph neural network model to perform fault reasoning on decoupled feature vectors and dynamic knowledge graphs; the fault diagnosis result includes the fault point and fault type. The deep feature decoupling network consists of a feature encoder, several feature decoders, and a context-aware attention fusion module. The feature encoder is used to extract high-level features from the dynamic feature matrix; The aforementioned feature decoders are executed in parallel to extract specific feature components from the high-level features; The context-aware attention fusion module is used to calculate the attention weights between the global context vector and several feature components through a multi-head attention mechanism, and to perform weighted fusion of several feature components and their corresponding attention weights to obtain a decoupled feature vector. The fault prediction unit obtains fault prediction results by performing forward-looking fault prediction based on fault diagnosis results and dynamic knowledge graph.

Images