Power line health state evaluation and prediction method and system based on big data

Abstract

The invention provides an electric power line health state assessment prediction method and system based on big data, and relates to the technical field of electric power line health assessment, and the method comprises the steps: preprocessing historical operation data, generating standardized data, extracting fault features through employing a space-time heterogeneous graph network, constructing a basic hierarchical knowledge base for automatic marking, and carrying out the prediction of the health state of an electric power line. Finally generating a training sample set; a directed acyclic graph is modeled, a main transmission line is identified, risk sections are divided, a knowledge base is optimized, an inspection scheme is adaptively generated based on an inspection strategy model of reinforcement learning, the health state of a power line is evaluated, and a health state report is generated; according to the invention, the fault prediction accuracy and inspection efficiency of the power line are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of power line health assessment, and particularly to a method and system for assessing and predicting the health status of power lines based on big data. Background Art

[0002] With the continuous expansion of the scale of the power system, the health status of power lines directly affects the safe and stable operation of the power system. At present, the assessment of the health status of power lines mainly relies on regular inspections and real-time monitoring data. By collecting multi-source heterogeneous data such as the operating parameters, environmental parameters, and equipment status of power lines, the assessment and prediction of the line health status are realized. The development of big data technology provides a new technical means for the assessment of the health status of power lines, which can mine fault characteristics and operation rules from a large amount of historical data and provide data support for the health status assessment.

[0003] However, there are still deficiencies in the existing technologies. Traditional health status assessment methods often process each monitoring data independently, ignoring the spatio-temporal correlation between data, resulting in insufficient feature extraction; most of the existing inspection strategies adopt a fixed-cycle inspection mode, and fail to dynamically adjust the inspection frequency according to the importance of the line and the fault risk, causing waste of inspection resources; there is a lack of in-depth analysis of the topological structure characteristics of power lines, unable to accurately identify key transmission paths, and it is difficult to achieve differentiated management.

[0004] In summary, the present invention aims to solve the above technical problems, and proposes a method for assessing and predicting the health status of power lines based on big data. By extracting the correlation features of multi-source data through a spatio-temporal heterogeneous graph network, establishing a hierarchical knowledge base to effectively organize knowledge, using a reinforcement learning method to optimize the inspection strategy, and combining the topological features of power lines to achieve differentiated inspection and assessment, thereby improving the accuracy of power line health status assessment and the inspection efficiency. Summary of the Invention

[0005] An embodiment of the present invention provides a method and system for assessing and predicting the health status of power lines based on big data, which can solve the problems in the existing technologies.

[0006] In the first aspect of the embodiment of the present invention, A method for assessing and predicting the health status of power lines based on big data is provided, including: Obtaining the historical operation data of the power line, preprocessing the historical operation data according to the preset data quality rules, and generating standardized operation data; Extracting fault features from the standardized operation data based on a spatio-temporal heterogeneous graph network, establishing a mapping relationship between the features and the operation status, constructing a basic hierarchical knowledge base, automatically annotating the standardized operation data, and generating a training sample set; Model the power line system as a directed acyclic graph, calculate the transmission path eigenvalue, identify the backbone transmission lines, divide the risk sections, and integrate them into the basic hierarchical knowledge base to obtain an optimized hierarchical knowledge base; According to the optimized hierarchical knowledge base and the backbone transmission lines, establish an inspection strategy model based on the reinforcement learning method, and adaptively generate an inspection execution plan; execute the inspection execution plan and evaluate the health status of the power line to obtain the health status evaluation result; Generate a power line health status report based on the health status evaluation result.

[0007] In an alternative embodiment, Extract fault features from the standardized operation data based on the spatio-temporal heterogeneous graph network, establish the mapping relationship between the features and the operation status, and construct the basic hierarchical knowledge base including: The standardized operation data includes equipment data, environmental data, and historical fault data; Based on the standardized operation data, construct device status nodes, environmental feature nodes, and historical fault nodes respectively; Based on the physical connection relationship between device status nodes, establish the topological association edges between devices; based on the influence relationship between device status nodes and environmental feature nodes, establish the device-environment association edges; based on the corresponding relationship between device status nodes and historical fault nodes, establish the device-fault association edges, and jointly form a heterogeneous information network; For each node in the heterogeneous information network, perform type-aware weighted aggregation extraction of fault features based on the multi-head attention mechanism to obtain the fault feature vector, and perform feature aggregation through differentiable pooling guided by node importance to obtain the subgraph feature representation; Use the causal convolution kernel to perform sliding convolution operations on the subgraph feature representation, generate the time series feature vector through time series-based information processing, calculate the contrast loss between normal operation state samples and the boundary loss between abnormal operation state samples, and obtain the comprehensive anomaly score of the fault feature; Store the time series feature vector, the comprehensive anomaly score, and the corresponding fault cases in the instance layer, and establish the mapping relationship between the fault features and the operation status; aggregate the fault cases with the Euclidean distance between the time series feature vectors corresponding to the fault cases less than the first preset threshold into the fault patterns and store them in the pattern layer; store the evolution relationship between the fault patterns with the time series correlation degree greater than the second preset threshold in the rule layer; the instance layer, the pattern layer, and the rule layer constitute the basic hierarchical knowledge base.

[0008] In an alternative embodiment, For each node in the heterogeneous information network, weighted aggregation based on the multi-head attention mechanism is performed to extract fault features for type perception, obtaining a fault feature vector. Feature aggregation is carried out through differentiable pooling guided by node importance, and the subgraph feature representation is obtained, including: According to different types of neighbor nodes of each node in the heterogeneous information network, a multi-head attention weight matrix is constructed, and the multi-head attention weight matrix is used to calculate the dynamic association strength between the node and the different types of neighbor nodes; Multiply the original feature vector of the node by the transformation matrix corresponding to the type of the node to obtain a type-aware feature vector, and perform a weighted aggregation operation on the type-aware feature vector based on the dynamic association strength to generate a fault feature vector; Concatenate the fault feature vector with the global feature vector of the heterogeneous information network, calculate the node importance score through a preset multi-layer perceptron, sample the nodes based on the node importance score, and construct a local network structure; For the fault feature vectors in the local network structure, feature aggregation is performed using differentiable pooling operations, and the aggregated features are updated through a gated recurrent unit to obtain the subgraph feature representation.

[0009] In an alternative embodiment, Model the power line system as a directed acyclic graph, calculate the transmission path eigenvalue, identify the main transmission lines, divide the risk sections, and integrate them into the basic hierarchical knowledge base to obtain an optimized hierarchical knowledge base, including: Construct a directed acyclic graph of the power line system, where the directed acyclic graph includes a set of operation feature nodes and a set of power transmission edges. Obtain the operation feature values of each node in the training sample set, and set the power transmission efficiency corresponding to each edge as the edge eigenvalue; Calculate the sum of the connected edge eigenvalues of each operation feature node in the directed acyclic graph to obtain the node importance value, determine the power transmission path based on the node importance value, and calculate the product of the edge eigenvalues of each power transmission path to obtain the path eigenvalue; Obtain the line load threshold of each operation feature node, identify the main transmission lines that meet the load threshold constraints through the maximum flow algorithm with iterative capacity adjustment, and calculate the corresponding edge cut set load ratio; Obtain the failure probability and the set of downstream impact nodes of each operation feature node, calculate the node risk value, correct the node risk value based on the edge cut set load ratio, and divide the risk sections through spectral clustering algorithm; Write the backbone transmission line into the skeleton layer of the basic hierarchical knowledge base, divide the protection levels based on the risk sections, record the path characteristic values and node risk values, establish a hierarchical association relationship including the load ratio of the edge cut set, and update the priority of the control strategy of the basic hierarchical knowledge base based on the hierarchical association relationship to obtain an optimized hierarchical knowledge base.

[0010] In an alternative embodiment, Obtain the line load threshold of each operating characteristic node, identify the backbone transmission line that meets the load threshold constraint through the maximum flow algorithm of iterative capacity adjustment, and calculate the corresponding load ratio of the edge cut set, including: Obtain the line load threshold of each operating characteristic node in the power system network, construct an initial capacity matrix, under the constraint that the actual load is less than or equal to the line load threshold, obtain the initial maximum flow value through the maximum flow algorithm, determine the corresponding minimum cut set based on the initial maximum flow value, and obtain the initial backbone transmission line; Apply a perturbation coefficient to the initial backbone transmission line, correct the initial capacity matrix to obtain a corrected capacity matrix, calculate the maximum flow value after perturbation, and calculate the ratio with the initial maximum flow value to obtain a network fitness index; Based on the difference between the network fitness index and the preset fitness threshold, update the corrected capacity matrix to obtain an iterative capacity matrix, calculate the iterative maximum flow value corresponding to the iterative capacity matrix, repeat the iteration until the change rate of the iterative maximum flow values in two adjacent iterations is less than the preset change rate threshold, obtain the final capacity matrix, screen the lines that meet the line load threshold from the final capacity matrix as the optimized backbone transmission lines, and calculate the corresponding load ratio of the edge cut set.

[0011] In an alternative embodiment, According to the optimized hierarchical knowledge base and the backbone transmission line, establish an inspection strategy model based on the reinforcement learning method, and adaptively generate an inspection execution plan, including: Based on the fault characteristics in the optimized hierarchical knowledge base, the mapping relationship with the operating state, and the transmission path characteristic values of the backbone transmission line, construct a regional characteristic function, and determine the boundary of the dynamically associated sub-region according to the regional characteristic function; Construct a local state evaluator for the dynamically associated sub-region, multiply the risk factor score of the local state evaluator by the weight coefficient to obtain the regional risk degree, and combine the regional risk degree with the inspection time interval and the load level to calculate the regional inspection urgency; Construct a feature similarity matrix of the dynamically associated sub-region, generate a regional knowledge transfer function according to the feature similarity matrix, and fuse the regional knowledge transfer function with the historical inspection experience to obtain the regional shared experience; Construct a state vector for the regional inspection urgency, regional risk level, and regional shared experience construction, construct an action vector for the inspection priority parameter, and establish a reward function based on the state vector and the action vector; Use the reinforcement learning method to optimize the reward function to obtain the policy model parameters, and generate an inspection execution plan according to the policy model parameters.

[0012] In an alternative embodiment, Construct a feature similarity matrix for dynamically associated sub-regions, generate a regional knowledge transfer function according to the feature similarity matrix, and fuse the regional knowledge transfer function with historical inspection experience to obtain regional shared experience, including: Construct the dynamically associated sub-regions into a graph network model. Each of the dynamically associated sub-regions in the graph network model corresponds to a regional node. The regional node contains fault features and operating status data. Based on the physical connection relationship between the dynamically associated sub-regions, construct edge connections, and extract the topological association features between the regional nodes through a graph convolutional network to obtain a feature similarity matrix; Calculate the similarity of adjacent regional nodes to obtain a gating coefficient, and construct a regional knowledge transfer function according to the gating coefficient and the feature similarity matrix. The regional knowledge transfer function represents the knowledge transfer relationship between regional nodes; Obtain positive sample pairs by constructing the features of the same dynamically associated sub-region within different time windows, obtain negative sample pairs by constructing the features of different dynamically associated sub-regions, and train a feature extraction model through a contrast loss function to obtain historical inspection experience; Fuse the regional knowledge transfer function with the historical inspection experience to obtain regional shared experience.

[0013] In the second aspect of the embodiments of the present invention, Provide a power line health status assessment and prediction system based on big data, including: The first unit is used to obtain the historical operation data of the power line, preprocess the historical operation data according to the preset data quality rules, and generate standardized operation data; The second unit is used to extract fault features from the standardized operation data based on a spatio-temporal heterogeneous graph network, establish a mapping relationship between features and operating status, construct a basic hierarchical knowledge base, automatically annotate the standardized operation data, and generate a training sample set; The third unit is used to model the power line system as a directed acyclic graph, calculate the transmission path eigenvalue, identify the main transmission line, divide the risk section, and integrate it into the basic hierarchical knowledge base to obtain an optimized hierarchical knowledge base; A fourth unit is configured to establish an inspection strategy model based on a reinforcement learning method according to an optimized hierarchical knowledge base and a backbone transmission line, and adaptively generate an inspection execution plan; execute the inspection execution plan and evaluate the health status of the power line to obtain a health status evaluation result. A fifth unit is configured to generate a power line health status report based on the health status evaluation result.

[0014] In a third aspect of the embodiments of the present invention, a kind of electronic device is provided, including: a processor; a memory for storing instructions executable by the processor; wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0015] In a fourth aspect of the embodiments of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0016] In the embodiments of the present invention, through the preprocessing and standardization of the historical operation data of the power line, the quality and consistency of the data can be improved, thereby providing a reliable data basis for subsequent fault feature extraction and status evaluation, and being able to more accurately reflect the actual operation status of the power line; extracting fault features based on the spatio-temporal heterogeneous graph network and establishing the mapping relationship between the features and the operation status can effectively identify potential fault risks, improve the accuracy of fault detection, provide a scientific basis for the maintenance and management of the power line, and reduce the probability of faults; the adaptive inspection strategy model generated by combining the reinforcement learning method can realize the dynamic evaluation and real-time monitoring of the health status of the power line, optimize the allocation of inspection resources, improve the inspection efficiency, and ensure the safe and stable operation of the power line. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 is a schematic flow chart of the power line health status evaluation and prediction method based on big data according to the embodiments of the present invention; Figure 2 is a performance analysis chart of the early warning time; Figure 3 is a comparison chart of the network topology and load distribution before and after the optimization of the backbone transmission line of the power system; Figure 4 is a visualization chart of the knowledge tree-like migration flow and experience accumulation. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Apparently, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0019] The technical solutions of the present invention will be described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments.

[0020] Figure 1 The following is a schematic flowchart of a method for evaluating and predicting the health status of a power line based on big data in an embodiment of the present invention. As Figure 1 shown, the method includes: Obtain the historical operation data of the power line, preprocess the historical operation data according to preset data quality rules, and generate standardized operation data; Extract fault features from the standardized operation data based on a spatio-temporal heterogeneous graph network, establish a mapping relationship between the features and the operation status, construct a basic hierarchical knowledge base, automatically annotate the standardized operation data, and generate a training sample set; Model the power line system as a directed acyclic graph, calculate the transmission path eigenvalue, identify the main transmission line, divide the risk section, and integrate it into the basic hierarchical knowledge base to obtain an optimized hierarchical knowledge base; Based on the optimized hierarchical knowledge base and the main transmission line, establish an inspection strategy model based on the reinforcement learning method, and adaptively generate an inspection execution plan; execute the inspection execution plan and evaluate the health status of the power line to obtain a health status evaluation result; Generate a power line health status report based on the health status evaluation result.

[0021] In an alternative embodiment, extracting fault features from the standardized operation data based on a spatio-temporal heterogeneous graph network, establishing a mapping relationship between the features and the operation status, and constructing a basic hierarchical knowledge base includes: The standardized operation data includes equipment data, environmental data, and historical fault data; Based on the standardized operation data, respectively construct equipment status nodes, environmental feature nodes, and historical fault nodes; Based on the physical connection relationship between the equipment status nodes, establish an inter-device topological association edge. Based on the influence relationship between the equipment status nodes and the environmental feature nodes, establish an equipment-environment association edge. Based on the correspondence between the equipment status nodes and the historical fault nodes, establish an equipment-fault association edge, and jointly form a heterogeneous information network; For each node in the heterogeneous information network, based on the multi-head attention mechanism, perform type-aware weighted aggregation to extract fault features, obtain a fault feature vector, and perform feature aggregation through differentiable pooling guided by node importance to obtain a subgraph feature representation; Use causal convolution kernels to perform sliding convolution operations on the subgraph feature representation, generate a time-series feature vector through time-series-based information processing, calculate the contrast loss between normal operating state samples and the boundary loss between abnormal operating state samples, and obtain a comprehensive anomaly score for the fault features; Store the time-series feature vector, the comprehensive anomaly score, and the corresponding fault cases in the instance layer to establish a mapping relationship between fault features and operating states; Aggregate fault cases with the Euclidean distance between the time-series feature vectors corresponding to the fault cases less than the first preset threshold into a fault mode and store it in the mode layer; Store the evolution relationship between fault modes with the time-series correlation degree corresponding to the fault mode greater than the second preset threshold in the rule layer; The instance layer, the mode layer, and the rule layer constitute a basic hierarchical knowledge base.

[0022] In a specific implementation, obtain standardized operation data, where the operation data includes device data, environmental data, and historical fault data. The device data may include parameters such as the temperature, pressure, and operation time of the device, the environmental data may include external environmental conditions such as humidity, temperature, and air pressure, and the historical fault data includes the types of faults that have occurred in the device in the past and their occurrence times.

[0023] The device status node represents the current status of the device, the environmental feature node represents the features of the environment, and the historical fault node represents the past fault records of the device. By analyzing the physical connection relationships between device status nodes, establish topological association edges between devices. For example, if two devices are connected by a cable, establish a topological association edge between them. Based on the influence relationship between the device status node and the environmental feature node, establish a device-environment association edge. If the operating state of the device is significantly affected by the environmental temperature, establish an association edge between the device status node and the environmental feature node. Based on the correspondence between the device status node and the historical fault node, establish a device-fault association edge. If a device has experienced a fault in a specific environment, establish an association edge between its status node and the corresponding fault node. Construct a heterogeneous information network containing device status nodes, environmental feature nodes, and historical fault nodes.

[0024] In the constructed heterogeneous information network, a multi-head attention mechanism is adopted for type-aware weighted aggregation to extract fault features. For each node, according to its type (device, environment, or fault), the features of adjacent nodes are weighted and aggregated through different attention heads to obtain the fault feature vector of the node; a differentiable pooling method guided by node importance is used to aggregate the features to obtain the feature representation of the subgraph. By analyzing the importance of nodes in the network, the features of important nodes are selectively aggregated, thereby improving the effectiveness of the feature representation.

[0025] The causal convolution kernel is used to perform sliding convolution operations on the subgraph feature representation. By sliding the convolution kernel over the time series, time series features are extracted to generate a time series feature vector. By calculating the contrast loss between normal operating state samples and the boundary loss between abnormal operating state samples, a comprehensive anomaly score of the fault features is obtained. The smaller the difference between the feature vectors of normal samples and the larger the difference between abnormal samples, the higher the comprehensive anomaly score, which reflects the severity of the fault.

[0026] The time series feature vector, the comprehensive anomaly score, and the corresponding fault cases are stored in the instance layer to establish the mapping relationship between fault features and operating states. The instance layer is used to store the detailed information of each fault case, including its time series feature vector and comprehensive anomaly score. This information will be used for subsequent fault mode recognition and evolution relationship analysis.

[0027] In the instance layer, the Euclidean distance between the time series feature vectors corresponding to the fault cases is calculated. If the distance is less than the first preset threshold, these fault cases are aggregated into a fault mode and stored in the mode layer. The fault mode represents a group of fault cases with similar features and manifestations, which is convenient for subsequent analysis and processing.

[0028] In the mode layer, the time series correlation degree between the fault modes is analyzed. If the time series correlation degree between some fault modes is greater than the second preset threshold, the evolution relationship between these fault modes is recorded and stored in the rule layer. The rule layer is used to store the evolution relationship between fault modes to help analyze the potential development trend and mutual influence of faults.

[0029] Through the above steps, a basic hierarchical knowledge base is constructed, including the instance layer, the mode layer, and the rule layer. The basic hierarchical knowledge base can provide an important basis for the prediction and diagnosis of equipment faults, help maintenance personnel take timely measures, reduce the equipment failure rate, and improve the reliability and safety of the equipment.

[0030] Exemplarily, assume that during the operation of a certain device, the temperature data is 70°C, the pressure is 1.5 MPa, and the ambient humidity is 60%. In the historical failure data, the device has had a failure once under the same temperature and humidity. Through the above method, in the constructed heterogeneous information network, an association edge is established between the device status node and the environmental feature node, forming a complete fault feature extraction system. Based on the operating status of the device and the historical failure records, the generated comprehensive anomaly score is 0.85, indicating that the device has a relatively high failure risk in the current state.

[0031] In an alternative embodiment, for each node in the heterogeneous information network, a type-aware weighted aggregation is performed based on the multi-head attention mechanism to extract fault features, obtaining a fault feature vector. Feature aggregation is performed through differentiable pooling guided by node importance, and the subgraph feature representation includes: According to the different types of neighbor nodes of each node in the heterogeneous information network, a multi-head attention weight matrix is constructed, and the multi-head attention weight matrix is used to calculate the dynamic association strength between the node and the different types of neighbor nodes; Multiply the original feature vector of the node by the transformation matrix corresponding to the type of the node to obtain a type-aware feature vector, and perform a weighted aggregation operation on the type-aware feature vector based on the dynamic association strength to generate a fault feature vector; Concatenate the fault feature vector with the global feature vector of the heterogeneous information network, and calculate the node importance score through a preset multi-layer perceptron. Based on the node importance score, sample the nodes to construct a local network structure; For the fault feature vector in the local network structure, perform a feature aggregation operation using a differentiable pooling operation, and update the aggregated features through a gated recurrent unit to obtain a subgraph feature representation.

[0032] In a specific embodiment, a node type mapping table is established, and each type of node in the power line is assigned a unique type identifier: device nodes such as transmission towers and transformers are recorded as type 1, environmental sensor nodes such as temperature, humidity, and wind speed are recorded as type 2, and historical failure record nodes such as insulator detachment and wire breakage are recorded as type 3. For each node in the network, an adjacency relationship table is established based on the topological structure, recording all its directly connected neighbor nodes and types.

[0033] For different types of node pairs, calculate the parameter matrices of 8 attention heads. Each attention head independently captures one aspect of the correlation features between nodes. During attention calculation, for the original feature vectors of the central node and each type of neighbor node, query vectors, key vectors, and value vectors are generated respectively through the corresponding type of linear transformation matrix. The query vector and the key vector are subjected to a scaled dot product operation to obtain the original attention scores. The original scores are normalized using the softmax function to obtain the attention weights reflecting the correlation strength between node pairs. After each attention head independently calculates a set of attention weights, the results of the 8 heads are concatenated and linearly transformed to obtain the final multi-head attention weight matrix.

[0034] Define a dedicated feature transformation matrix for each type of node, and the matrix dimension is adapted to the original feature dimension of the node. Taking the device node as an example, collect multi-dimensional original features including running duration, load rate, maintenance records, vibration data, etc. to form a feature vector. Multiply the feature vector by the transformation matrix exclusive to this type of node to obtain the initial feature representation incorporating type information.

[0035] For each node, process its neighbor nodes of different types separately. According to the obtained multi-head attention weights, weight the type-aware features of the neighbor nodes. Multiply the feature vectors of the neighbor nodes by the corresponding attention weights, and then sum the weighted features of all neighbors to obtain the aggregated representation of the neighborhood information. Concatenate the type-aware features of the node itself with the neighborhood aggregated features, and process them through the non-linear activation function LeakyReLU to extract the fault features and generate the fault feature vector. The features obtained in this way retain both the node's own information and the influence of different types of neighbor nodes.

[0036] Conduct global feature statistics on the entire heterogeneous information network, calculate topological features such as the average node degree, clustering coefficient, eigenvalue distribution of the network; statistically analyze the mean, variance, and distribution characteristics of device operation parameters; analyze the spatio-temporal distribution law of fault occurrences. Concatenate these global statistical features with the fault feature vector of each node to obtain the comprehensive feature vector.

[0037] Construct a three-layer perceptron network for importance evaluation. The dimension of the input layer matches the comprehensive feature vector. The first hidden layer uses 256 neurons, and the second hidden layer uses 128 neurons. Both hidden layers use the ReLU activation function to achieve non-linear transformation, and at the same time, the Dropout mechanism is applied to prevent overfitting. The output layer uses one neuron with the sigmoid activation function to map the result to the 0-1 interval as the node importance score.

[0038] Sort the nodes based on the importance scores and select the nodes with scores in the top 30% as key nodes. For each key node, collect its second-order neighbors (neighbors of neighbors) to construct a local network structure for subsequent feature aggregation. This importance-based sampling method ensures that subsequent analysis focuses on the key regions in the network.

[0039] For each local network structure, construct an affinity matrix between nodes. Calculate the cosine similarity of the feature vectors between node pairs and use a Gaussian kernel function to map the similarity to the 0-1 interval. Based on the affinity matrix, perform soft assignment on the nodes, and assign larger weights to the features of nodes with high similarity for aggregation to generate the feature representations of several cluster centers.

[0040] Use a gated recurrent unit (GRU) to process the time series feature sequence. Input the features of 30 consecutive time steps into the GRU in sequence. At each time step, the reset gate determines how much historical information to forget based on the current input and the hidden state at the previous moment; the update gate determines the degree of incorporation of the current input information; and the final output is the feature representation that integrates historical information and the current state. To improve the effect of time series modeling, adopt a bidirectional GRU structure and consider both forward and backward time series dependencies.

[0041] Weightedly combine the feature representations at different time steps through an attention mechanism to obtain a subgraph representation that reflects the dynamic evolution characteristics of the local network. This hierarchical feature aggregation and time series modeling method can effectively capture the multi-scale spatio-temporal correlation patterns in the power line.

[0042] Exemplarily, the analysis process of the node of the No. 1 tower on a certain 500 kV transmission line: 1. Attention calculation: This tower is connected to the No. 2 tower (equipment node), temperature sensor (environmental node), and the record of the shedding of a primary insulator (fault node). Calculate three groups of attention weights respectively: for the No. 2 tower, considering the similarity of their operating parameters and physical distance, the initial score of 0.85 is obtained through dot product operation, and after normalization, it is 0.8; for the temperature sensor, the score of 0.65 is calculated based on the influence law of temperature on equipment performance, and after normalization, it is 0.6; for the fault record, the score of 0.45 is calculated according to the correlation of the equipment state at the time of the fault, and after normalization, it is 0.4; 2. Feature processing: The original features of the tower include numerical values such as [operating duration: 5 years, load rate: 85%, number of maintenance times: 3 times], etc. After being mapped by a transformation matrix and weighted by neighbor features, a fault feature vector is obtained, which reflects the comprehensive information of equipment state, environmental impact, and historical faults.

[0043] 3. Importance assessment: Combine the fault characteristics of this tower with the overall statistical characteristics of the line (average load rate of 75%, average failure rate of 0.8 times per year, etc.). Through calculation by a multi-layer perceptron, an importance score of 0.85 is obtained, indicating that this tower is a key monitoring point.

[0044] 4. Feature aggregation: Combine this tower with adjacent important towers to form a local structure. Process the status sequence of the last 30 days through a gated recurrent unit to generate a feature representation reflecting the evolution of the health status of this section.

[0045] The theoretical basis of the solution in this embodiment stems from recent research results in the field of graph neural networks, especially the cross-application of heterogeneous graph neural networks and temporal graph modeling. Currently, traditional CNN temporal methods, unidirectional GRU methods, and differentiable pooling methods are mainly used for power system fault prediction. The traditional CNN temporal method slices power equipment data according to time windows and converts it into a two-dimensional matrix form for processing. Although it can capture local temporal patterns, it is difficult to model long-distance dependencies and cannot effectively express the topological associations between devices, resulting in a limited warning time. For example, for insulator shedding faults, it can only give a warning 21.5 hours in advance. The unidirectional GRU method directly processes parameter sequences and can capture temporal dependencies to a certain extent, but ignores the spatial topological structure in the power network. Although the prediction lead time has been improved, it is still limited. For example, for equipment aging faults, it can only give a warning 39.1 hours in advance. The differentiable pooling method extracts network features through graph convolution and pooling operations, and can better capture the topological relationships between devices, but often treats different types of nodes as homogeneous nodes and lacks sufficient modeling of temporal evolution. For example, for grounding faults, it can only give a warning 27.8 hours in advance.

[0046] To address the above limitations, the solution in this embodiment proposes a comprehensive improvement method based on differentiable pooling of heterogeneous information networks and bidirectional GRU temporal modeling. First, different from the traditional method of simplifying the power system into a homogeneous graph, this solution clearly distinguishes different types of nodes such as transmission equipment, environmental sensors, and historical fault records, and constructs a dedicated attention calculation mechanism for each type of node pair. This design is based on the understanding that the power system is essentially a complex system composed of multiple heterogeneous entities, and there are significant differences in the interaction patterns between different types of nodes. Second, this solution designs dedicated feature transformation matrices for different types of nodes and performs differential aggregation of neighborhood information based on dynamically calculated attention weights, considering that different types of power system components have different working characteristics and fault modes, and their features need to be mapped to a feature space suitable for heterogeneous network learning through different transformations.

[0047] This solution also combines node features and global network characteristics, evaluates the importance of nodes through a multi-layer perceptron, and samples based on this to construct a local network. This design stems from the understanding that not all nodes in the power system are equally important for fault prediction. Focusing on key nodes and their local structures can reduce noise interference and improve computational efficiency. Most innovatively, it organically combines the differentiable pooling operation in the spatial domain with the bidirectional GRU model in the temporal domain to construct a spatio-temporal fusion feature representation. This is because power system faults are usually complex events with spatio-temporal correlations, having both propagation characteristics in spatial topology and evolution laws in time series. The bidirectional GRU can consider both historical trends and possible future developments, while differentiable pooling effectively aggregates the structural information of the local network.

[0048] This significant extension of the early warning time brings remarkable practical value to power system maintenance, enabling the operation and maintenance team to deploy resources in advance and plan the best repair strategies, thereby reducing power outage time, lowering repair costs, and improving power grid reliability. Especially for severe faults such as insulator detachment and wire breakage, this solution can provide early warnings about two days in advance, fully meeting the time requirements for on-site emergency repairs. This solution performs evenly across different types of faults, demonstrating its strong modeling and generalization capabilities for diverse fault patterns in the power system, and providing a powerful decision-making support tool for the preventive maintenance of the power system.

[0049] such as Figure 2As shown, it demonstrates the early warning time performance of different network structure feature aggregation and time series modeling methods in power line fault prediction. The horizontal axis in the figure represents different types of power line faults, and the vertical axis represents the average time (in hours) that the system can predict the fault in advance. The numbers above the data points indicate the early warning time that the corresponding method can achieve for this fault type. It can be clearly seen from the figure that the technical solution (using the method combining differentiable pooling and bidirectional GRU) performs best for all fault types. Especially for the insulator detachment fault, the technical solution can predict the fault occurrence 47.3 hours (about 2 days) in advance, which is nearly 26 hours earlier than the traditional CNN time series method (21.5 hours) and about 15 hours earlier than the method using only differentiable pooling (32.6 hours). For the conductor breakage fault, the early warning time of the technical solution is 42.8 hours, which is 9.3 hours higher than the second-best method (using only differentiable pooling). In the lightning strike fault prediction, although the early warning time of each method is generally short (which is related to the suddenness of lightning strike events), the technical solution can still give an early warning 11.6 hours in advance, which is better than other methods. For the three types of progressive faults, namely equipment aging, grounding fault, and connector damage, the technical solution realizes the early warning times of 58.7 hours, 38.2 hours, and 45.6 hours respectively, always maintaining a significant advantage. Notably, for the equipment aging type, the early warning time of the technical solution reaches an astonishing 58.7 hours (more than 2 days), which is nearly 30 hours earlier than the traditional method, and this has great practical value for arranging maintenance personnel and resources. Generally speaking, by combining the advantages of differentiable pooling (effectively aggregating local network structure information) and bidirectional GRU (capturing time series dependencies), the technical solution significantly improves the early warning time of fault prediction, being on average 24.3 hours earlier than the traditional method, providing sufficient response time for power line maintenance.

[0050] In this embodiment, the multi-head attention mechanism is used to dynamically weight different types of neighbor nodes, enabling the model to fully exploit the multi-dimensional correlation features among various types of nodes, improving the fine-grainedness and accuracy of feature representation; by multiplying the original features with the exclusive transformation matrix, type-aware feature representations are generated, thus better retaining the specific information of various types of nodes and effectively enhancing the complementarity of node self-information and neighborhood information; combining global statistical features with local key node sampling not only reflects the overall network topology characteristics but also focuses on local key regions, enabling the model to have a global perspective when dealing with large-scale networks and accurately capture local anomalies or important patterns; through importance-based sampling and subsequent soft assignment of the affinity matrix and differentiable pooling operations, efficient aggregation of neighborhood features is achieved, reducing redundant information and improving data utilization efficiency; using bidirectional GRU and the attention mechanism to model continuous time-series features can capture the multi-scale spatio-temporal evolution laws in power lines, providing a more reliable basis for fault prediction and maintenance decision-making.

[0051] In an alternative embodiment, the power line system is modeled as a directed acyclic graph, the transmission path eigenvalue is calculated, the backbone transmission line is identified, the risk section is divided, and integrated into the basic hierarchical knowledge base to obtain an optimized hierarchical knowledge base, including: Construct a directed acyclic graph of the power line system, the directed acyclic graph includes a set of operation feature nodes and a set of power transmission edges, obtain the operation feature value of each node in the training sample set, and set the power transmission efficiency corresponding to each edge as the edge feature value; Calculate the sum of the connected edge feature values of each operation feature node in the directed acyclic graph to obtain the node importance value, determine the power transmission path based on the node importance value, and calculate the product of the edge feature values of each power transmission path to obtain the path eigenvalue; Obtain the line load threshold of each operation feature node, identify the backbone transmission line that meets the load threshold constraint through the maximum flow algorithm with iterative capacity adjustment, and calculate the corresponding edge cut set load ratio; Obtain the failure probability and the set of downstream affected nodes of each operation feature node, calculate the node risk value, correct the node risk value based on the edge cut set load ratio, and divide the risk section through the spectral clustering algorithm; Write the backbone transmission line into the skeleton layer of the basic hierarchical knowledge base, divide the protection levels based on the risk section, record the path eigenvalue and the node risk value, establish a hierarchical association relationship including the edge cut set load ratio, and update the control strategy priority of the basic hierarchical knowledge base based on the hierarchical association relationship to obtain an optimized hierarchical knowledge base.

[0052] In a specific embodiment, a directed acyclic graph of the power line system is constructed. The directed acyclic graph consists of a set of operation characteristic nodes and a set of power transmission edges. The operation characteristic nodes represent various devices or nodes in the power system, such as transformers, switches, etc. The power transmission edges represent the power flow paths between the nodes. The operation characteristic values of each node can be obtained through historical data or real-time monitoring data. Common characteristic values include voltage, current, power, etc. For each edge, the power transmission efficiency is used as the characteristic value of the edge, which can be determined by the ratio of the rated capacity of the edge to the actual transmitted power.

[0053] Calculate the sum of the characteristic values of the connected edges of each operation characteristic node to obtain the importance value of the node. The importance value of the node reflects the criticality of the node in power transmission. The higher the importance value, the greater the impact of the node on the overall power system. According to the importance value of the node, determine the power transmission path. The power transmission path consists of a series of connected nodes and edges. The product of the edge characteristic values of each power transmission path can be used to represent the characteristic value of the path. The higher the characteristic value, the higher the transmission efficiency of the path.

[0054] In the process of identifying the main transmission lines, obtain the line load threshold of each operation characteristic node. The line load threshold refers to the maximum load that the node can withstand under normal operation. Identify the main transmission lines that meet the load threshold constraints through the maximum flow algorithm with iterative capacity adjustment. By continuously adjusting the flow of each edge, find the path that can maximize the flow. After identifying the main transmission lines, calculate the corresponding edge cut set load ratio. The edge cut set refers to the set of edges that can effectively separate the power flow under specific conditions. The load ratio is the ratio of the load of the edges in the set to their carrying capacity.

[0055] To evaluate the risk of the nodes, obtain the failure probability and the set of downstream affected nodes of each operation characteristic node. The failure probability refers to the possibility of the node failing within a specific time, and the set of downstream affected nodes refers to the other nodes affected by the failure of the node. The calculated node risk value can reflect the impact degree of the node on the entire system in case of failure. Correct the node risk value based on the edge cut set load ratio to more accurately reflect the risk situation of the node under different load conditions.

[0056] To divide the risk sections, use the spectral clustering algorithm. Through the clustering analysis of the node risk values, group the nodes with similar risks into one category, thus forming different risk sections. Corresponding protection strategies can be formulated for each risk section according to its characteristics to reduce the overall risk of the system.

[0057] Write the identified backbone transmission line into the skeleton layer of the basic hierarchical knowledge base, which is a multi-level knowledge management system. The skeleton layer contains the basic structure and main components of the system. Based on the risk section, divide the protection levels, record the path characteristic values and node risk values, and establish the hierarchical association relationship including the edge cut set load ratio. These association relationships help to understand the interaction between each node and edge in the power system.

[0058] Update the control strategy priority of the basic hierarchical knowledge base based on the hierarchical association relationship. The control strategy priority refers to the nodes and paths that are preferentially processed in case of faults or abnormalities. By optimizing the hierarchical knowledge base, the stability and reliability of the power system can be improved, ensuring a prompt response and corresponding measures can be taken when problems occur.

[0059] Exemplarily, taking a certain power company as an example, assume that the company has a power line system with ten nodes. The operating characteristic values of each node are as follows: Node A: Voltage 220V, current 100A, power 22kW; Node B: Voltage 220V, current 150A, power 33kW; Node C: Voltage 220V, current 80A, power 17.6kW; Node D: Voltage 220V, current 120A, power 26.4kW; Node E: Voltage 220V, current 90A, power 19.8kW; Node F: Voltage 220V, current 110A, power 24.2kW; Node G: Voltage 220V, current 130A, power 28.6kW; Node H: Voltage 220V, current 70A, power 15.4kW; Node I: Voltage 220V, current 140A, power 30.8kW; Node J: Voltage 220V, current 160A, power 35.2kW.

[0060] Based on the above data, the sum of the connected edge characteristic values of each node can be calculated, and then the importance value of the node can be obtained. Assume that node A is connected to nodes B and C, node B is connected to nodes D and E, and so on, and finally the power transmission path is determined.

[0061] When identifying the backbone transmission line, assume that the load threshold of node A is 150A and the load threshold of node B is 200A. After the maximum flow algorithm, the identified backbone line is A - B - D. The edge cut set load ratio is 0.8, indicating the stability of the line under load.

[0062] By obtaining the failure probability of nodes and the set of downstream affected nodes, assuming the failure probability of node A is 0.01 and the set of affected nodes is {B, C}, the risk value of node A is 0.02. After the spectral clustering algorithm, the nodes are divided into high-risk sections and low-risk sections.

[0063] The identified backbone transmission lines and risk section information are written into the basic hierarchical knowledge base, and the priority of the updated control strategy is optimized to ensure the efficient operation of the power system.

[0064] In an alternative embodiment, the line load threshold of each operating characteristic node is obtained, and the backbone transmission lines that meet the load threshold constraint are identified through the maximum flow algorithm of iterative capacity adjustment, and the corresponding edge cut set load ratio is calculated, including: Obtain the line load thresholds of each operating characteristic node in the power system network, construct an initial capacity matrix, and under the constraint that the actual load is less than or equal to the line load threshold, obtain the initial maximum flow value through the maximum flow algorithm, determine the corresponding minimum cut set based on the initial maximum flow value, and obtain the initial backbone transmission lines; Apply a perturbation coefficient to the initial backbone transmission lines, correct the initial capacity matrix to obtain a corrected capacity matrix, calculate the maximum flow value after perturbation, and calculate the ratio with the initial maximum flow value to obtain the network fitness index; Based on the difference between the network fitness index and the preset fitness threshold, update the corrected capacity matrix to obtain an iterative capacity matrix, calculate the iterative maximum flow value corresponding to the iterative capacity matrix, repeat the iteration until the change rate of the iterative maximum flow value between adjacent two iterations is less than the preset change rate threshold, obtain the final capacity matrix, screen the lines that meet the line load threshold from the final capacity matrix as the optimized backbone transmission lines, and calculate the corresponding edge cut set load ratio.

[0065] Construct the network topology structure of the power system, set the substations as network nodes and the transmission lines as network edges. For each transmission line, collect the following operating parameters: basic data such as rated capacity, real-time load, historical maximum load, voltage level, line length, conductor type, etc. Based on the historical operation data, calculate the average load level and load fluctuation range of each line. Considering the equipment safety margin and system stability requirements, set hierarchical load thresholds: for 500kV backbone lines, take 80% of the rated capacity as the threshold; for 220kV regional lines, take 75% of the rated capacity as the threshold; for 110kV and below distribution lines, take 70% of the rated capacity as the threshold.

[0066] When constructing the initial capacity matrix, the matrix dimension is the total number of nodes × the total number of nodes, and the matrix elements represent the load thresholds of the lines between the corresponding nodes. The matrix elements between non-connected nodes are 0. According to the power flow direction of the power system, convert the capacity matrix into a directed graph structure, and determine the power source side node as the source point and the load side node as the sink point.

[0067] The Ford-Fulkerson maximum flow algorithm is adopted to establish a residual network, and the flow of all edges is initialized to 0. The breadth-first search method is used to scan the network nodes layer by layer to find a feasible path from the source node to the sink node. For each feasible path, calculate the remaining capacity of all edges on the path, and select the minimum value as the flow for this augmentation. Update the remaining capacity of each edge in the residual network, and add the augmented flow to the total flow. Repeat the search-augmentation process until no new feasible path can be found. At this time, the total flow obtained is the initial maximum flow value.

[0068] Starting from the source node, mark all nodes reachable through the residual network as set S; the unmarked nodes are classified into set T. All edges connecting set S and set T form the minimum cut set, and the transmission lines corresponding to these edges are the initial backbone transmission lines. At the same time, record information such as the load threshold, actual load, and node connection relationship of these lines for subsequent optimization.

[0069] Design a multi-level perturbation test scheme, including single-line perturbation, associated line combination perturbation, regional perturbation, etc. For each backbone line, construct a perturbation coefficient sequence within the range of [-5%, +5%], and set the perturbation step size to 1%. For associated line groups, consider the load transfer effect and design a complementary perturbation scheme.

[0070] For each backbone line, multiply its initial capacity by the corresponding perturbation coefficient to obtain the corrected capacity value. At the same time, consider the load transfer constraint between lines to ensure that the corrected capacity distribution meets the system stability requirements. For non-backbone lines, set corresponding capacity adjustment coefficients according to their association degree with the backbone lines.

[0071] Under the corrected capacity constraint, repeat the execution of the maximum flow algorithm. Use the same Ford-Fulkerson algorithm as the initial calculation, but adopt the corrected capacity value. Record the maximum flow value in each perturbation case, compare it with the initial maximum flow value, and calculate the network fitness index. Conduct statistical analysis on the results of multiple perturbation tests to obtain the evaluation index of the system's adaptability to different types of perturbations.

[0072] Based on the network fitness evaluation results, establish a capacity optimization strategy. When the fitness index is lower than the threshold, give priority to increasing the capacity of key lines; when the index is higher than the threshold, consider reducing the capacity of non-key lines to achieve optimal resource allocation. The specific steps for each iterative adjustment include: calculating the fitness difference, determining the adjustment target, generating an adjustment plan, verifying the feasibility of the plan, and updating the capacity matrix.

[0073] Calculate the change rate of the maximum flow value between two adjacent iterations, and simultaneously monitor the convergence indicators in multiple dimensions such as the magnitude of capacity adjustment and the change trend of the network fitness index. When all indicators meet the requirements of the preset threshold, the optimization process is considered to have converged.

[0074] Based on the final capacity matrix, select the lines that meet the load threshold constraint and have high network importance. Calculate the characteristics of the edge cut sets formed by these lines, including: cut set capacity ratio, load distribution uniformity, reliability index, etc.

[0075] Exemplarily, take the 500kV backbone grid of a certain power grid as a specific case to illustrate the optimization process of the main transmission lines in detail. This power grid system consists of 12 500kV substations and 18 transmission lines, including three power source side nodes, namely S1 thermal power plant, S2 hydropower plant, and S3 nuclear power plant, four load center nodes from S9 to S12, and five relay transfer nodes from S4 to S8. In the initial operation state of the system, the operating parameters of the main transmission lines L1 to L5 connecting the power source side and the relay nodes are different: the rated capacity of the L1 line connecting the thermal power plant is 2400MW, and the actual operating load reaches 1850MW; the rated capacity of the L2 line connecting the hydropower plant is 2200MW, and the actual load is 1680MW; the rated capacity of the L3 line connecting the nuclear power plant is the largest, reaching 2600MW, and the actual load is also relatively high, maintaining at 2050MW; while the rated capacities of the L4 and L5 lines connecting the relay nodes are relatively small, 2000MW and 1800MW respectively, and the actual loads are also correspondingly low, 1520MW and 1380MW respectively.

[0076] In the first-round iterative optimization, the system first determined the source point set and the sink point set, calculated the initial maximum flow value of 5580MW through the maximum flow algorithm, and identified that L1, L2, and L3 form the key transmission channels. To test the adaptability of the system, different degrees of capacity perturbations were applied to these three key lines: L1 increased by 5%, L2 increased by 3%, and L3 increased by 4%. After the perturbation, the maximum flow of the system increased to 5750MW, and the network fitness index reached 1.03. Based on the test results, the capacity matrix was adjusted for the first time: while increasing the capacity of the key lines, appropriately reducing the capacity configuration of the non-main lines. Among them, the capacity of L1 was increased to 1968MW, L2 was increased to 1787MW, L3 was increased to 2122MW, while the capacities of L4 and L5 were reduced to 1584MW and 1426MW respectively.

[0077] In the second iteration, the maximum system flow value is further increased to 5680 MW, which is a 1.79% increase compared to the initial state. Perturbation tests within the range of ±4% show that the maximum system flow value can fluctuate between 5580 MW and 5820 MW, and the network fitness remains at the level of 1.02. Based on this, the capacity configuration is further optimized: L1 is increased by 1.5% to reach 1998 MW, L2 and L3 are increased by 1% and 1.2% respectively to reach 1805 MW and 2147 MW, while the capacities of L4 and L5 are slightly decreased to 1571 MW and 1419 MW.

[0078] An important change occurred in the third iteration. The maximum system flow value reached 5720 MW, and the growth rate decreased to 0.70%. At the same time, the minimum cut set was extended to include the L6 line. The perturbation test range was narrowed to ±3%, and the fluctuation range of the maximum flow value was narrowed to between 5650 MW and 5780 MW. According to the test results, smaller adjustments were made to the main lines. The capacity growth rates of L1 to L3 were all controlled within 1%, and at the same time, the capacity of the newly included L6 line was increased by 1%.

[0079] After entering the fourth iteration, the system optimization began to show an obvious convergence trend. The maximum flow value was only slightly increased to 5735 MW, and the change rate decreased to 0.26%. The perturbation test showed that the system operation was more stable, and the fluctuation range of the maximum flow value was further narrowed to between 5700 MW and 5760 MW. At this time, the capacity adjustment range of each line did not exceed 0.5%. By the fifth iteration, the maximum system flow value reached 5742 MW, and the change rate between adjacent iterations decreased to 0.12%, which was lower than the preset threshold of 0.2%, indicating that the optimization process had reached convergence.

[0080] After five rounds of iterative optimization, the system formed a clear hierarchical transmission architecture: L1, L2, and L3 form the core backbone, L6 and L8 serve as important supports, and L12 and L15 play the role of local key channels. From the perspective of the load ratio of the edge cut set, L3 reached the highest value of 0.95, making it the most critical power transmission channel in the system; L1 was second with 0.92, and L2 was 0.88, both at relatively high levels; the load ratios of L6 and L8 were between 0.80 and 0.83, providing good support. Overall, the maximum transmission capacity of the optimized system increased by 2.9%, the load distribution of the key lines became more balanced, and the network fitness was stably maintained above 0.95, achieving the expected optimization goal.

[0081] The optimization analysis and enhancement technology of the main transmission lines in the power system are of great significance for the efficient operation of the current power grid. Traditional identification of the main grid lines mainly relies on static topology analysis and expert experience judgment, such as power flow marginal sensitivity analysis and N-1 security constraint analysis method, which are difficult to adapt to new complex scenarios such as dynamic fluctuations of grid loads and integration of new energy into the grid. The existing network flow theory is widely applied in communication and transportation networks, but its application in the power system is restricted by the special physical constraints of power flow.

[0082] Traditional methods usually regard the line capacity as a fixed parameter, ignoring the coupling relationship and dynamic adjustment potential among system operation parameters. For example, the analysis method based on node importance focuses on node connectivity and centrality, but it cannot accurately depict the load transfer relationship between lines; the method based on power flow sensitivity can identify key lines, but it is difficult to quantify the impact of line capacity adjustment on the overall system performance. In addition, most of the existing optimization methods adopt a single objective function, which is difficult to balance multi-dimensional requirements such as system security, economy and flexibility.

[0083] The steps of this embodiment are based on the network maximum flow-minimum cut theory, and introduce an iterative capacity matrix perturbation and adaptive optimization mechanism. Compared with traditional methods, the main improvements are reflected in three aspects: First, it realizes the setting of hierarchical load thresholds, and differentially configures safety margins according to the line voltage level and importance, which is more in line with the actual operation requirements; second, it designs a multi-level perturbation test system, and reveals the quantitative relationship between line capacity adjustment and the system's maximum transmission capacity through systematic perturbation analysis, overcoming the subjectivity of traditional methods in evaluating line importance; third, it establishes an iterative optimization framework based on network fitness indicators, realizing the adaptive adjustment and dynamic convergence of capacity allocation.

[0084] The starting point of the improvement stems from the in-depth understanding of the inherent characteristics of the power system. As a key infrastructure, the power system needs to maintain stable and reliable operation, and also needs to efficiently utilize existing resources to cope with load growth. Traditional static safety margin configuration often leads to waste of resources, while overly aggressive optimization may reduce system security. This method explores a technical path to improve the system's transmission capacity while ensuring safety through the organic combination of network flow theory and the characteristics of the power system.

[0085] Practice shows that this embodiment can effectively identify the key main lines in the power grid and scientifically and reasonably adjust the line load thresholds. In the test system, the maximum flow value after optimization increased by 2.96%, which means that under the same network topology, the power transmission capacity of the system has been significantly improved. The core backbone lines and load convergence lines are accurately identified as key edge cut sets, providing clear guidance for operation and maintenance and investment decisions. In addition, the network fitness index reaches 0.97, indicating that the optimized system has strong anti-perturbation ability and operation flexibility.

[0086] Compared with traditional methods, the implementation steps of this method have stronger practicability and adaptability. It is applicable not only to static planning scenarios but also supports real-time operation adjustment. It not only focuses on the optimization of a single line but also considers the overall coordination of the system. It not only provides the optimization results but also reveals the dynamic characteristics of the system during the optimization process. This comprehensive technical solution provides a new analysis tool and decision support for the planning and operation of power systems, and has important value for improving the utilization efficiency of grid resources and operation reliability.

[0087] As Figure 3 shown, it presents the optimized network topology structure and load distribution of the 500kV power grid system trunk transmission lines. The entire power system consists of 12 nodes and 18 lines, including 3 power source side nodes (represented by circles), 5 relay transfer nodes (represented by diamonds), and 4 load center nodes (represented by squares). The thickness of the lines in the figure represents their importance and load level in the system, and the line labels show the "actual load / optimized threshold (MW)" information. Through optimization, the system forms a clear hierarchical transmission architecture: Lines L1, L2, and L3 form the core backbone (the thickest lines), connecting the three power source side nodes (S1 thermal power plant, S2 hydropower plant, and S3 nuclear power plant) to the relay substations respectively, and their load ratios reach 92.6% (1850 / 1998MW), 93.1% (1680 / 1805MW), and 95.5% (2050 / 2147MW) respectively, fully exerting the transmission channel capacity. Five lines, L6, L8, L10, L12, and L15, form an important support layer (medium lines), with load ratios ranging from 83.8% to 95.1%. Among them, L6 (1450 / 1730MW) and L8 (1320 / 1580MW), as the key connections between relay nodes, support the efficient allocation of energy in the middle of the system. The remaining lines form the general transmission layer and the low-load standby layer, providing diverse path options for the system. Overall, the optimized system shows an obvious "center-radiation" structural feature, with the S8 substation as the core hub, connecting the upstream power sources and downstream loads. Through this optimized configuration, the maximum flow value of the system increases from the initial 5580MW to 5745MW, with a growth rate of 2.96%, and the network fitness index reaches 0.97. At the same time, the main edge cut set composed of L1+L2+L3 and the secondary edge cut set composed of L9+L10+L11 both reach a relatively high load balancing level, indicating that this optimization scheme not only improves the overall transmission capacity of the system but also enhances the stability of the system in the face of complex operating conditions.

[0088] In this embodiment, through the maximum flow algorithm and minimum cut set analysis, the backbone transmission lines in the power network can be accurately identified, providing a clear basis for subsequent optimization; by introducing a perturbation coefficient and a fitness index, iterative update of the capacity matrix is realized, enabling the optimization process to adaptively adjust the network state, improving the transmission efficiency and reliability; under the constraint of meeting the line load threshold, by repeatedly iteratively calculating the maximum flow value, the network load distribution is optimized, thereby reducing the risk of local overload and enhancing the system stability; by calculating the load ratio of the edge cut set for the backbone lines, the load conditions of the key transmission paths in the network are further quantified, providing data support for the operation and maintenance decision-making of the power system.

[0089] In an alternative embodiment, based on the optimized hierarchical knowledge base and the backbone transmission lines, a patrol strategy model is established using the reinforcement learning method, and the adaptive generation of the patrol execution plan includes: Based on the fault characteristics in the optimized hierarchical knowledge base, the mapping relationship with the operating state, and the transmission path eigenvalue of the backbone transmission line, a regional feature function is constructed, and the boundary of the dynamically associated sub-region is determined according to the regional feature function; A local state evaluator is constructed for the dynamically associated sub-region, the risk factor score of the local state evaluator is multiplied by the weight coefficient to obtain the regional risk degree, and the regional risk degree is combined with the patrol time interval and the load level to calculate the regional patrol urgency; A feature similarity matrix of the dynamically associated sub-region is constructed, a regional knowledge transfer function is generated according to the feature similarity matrix, and the regional knowledge transfer function is fused with the historical patrol experience to obtain the regional shared experience; The regional patrol urgency, the regional risk degree, and the regional shared experience are used to construct a state vector, the patrol priority parameter is used to construct an action vector, and a reward function is established according to the state vector and the action vector; The reinforcement learning method is used to optimize the reward function to obtain the policy model parameters, and the patrol execution plan is generated according to the policy model parameters.

[0090] An optimized hierarchical knowledge base is constructed, which contains the mapping relationship between fault characteristics and operating states, as well as the transmission path eigenvalues of the backbone transmission lines. By analyzing historical fault data, the characteristics of different fault types are extracted and associated with the operating states of the equipment. The operating states include parameters such as the temperature, pressure, and current of the equipment. At this time, the construction of the regional feature function is based on these characteristics, which can effectively reflect the operating states and fault risks of different regions.

[0091] Based on the regional feature function, determine the boundaries of the dynamically associated sub-regions. The dynamically associated sub-regions refer to the regions where the fault characteristics are highly correlated with the operating state under specific conditions. By real-time monitoring of the operating data and combining with the changes in fault characteristics, dynamically adjust the boundaries of the sub-regions to ensure the effectiveness of the inspection strategy.

[0092] For the determined dynamically associated sub-regions, construct a local state evaluator. The main function of the local state evaluator is to evaluate the risks of the local regions. By combining the scores of risk factors and the weight coefficients, obtain the regional risk degree. Risk factors include the fault history of the equipment, the degree of abnormality of the current operating state, etc. The regional risk degree reflects the inspection priority of the region in the current state.

[0093] On this basis, combine the regional risk degree with the inspection time interval and the load level to calculate the urgency of regional inspection. The inspection time interval refers to the time from the last inspection to the current moment, and the load level refers to the working load of the equipment during inspection. By comprehensively considering these factors, it is possible to arrange the inspection plan more reasonably to ensure that high-risk regions are inspected in a timely manner.

[0094] Further construct the feature similarity matrix of the dynamically associated sub-regions. The feature similarity matrix is obtained by comparing the characteristics of each region to evaluate their similarity degree. Regions with high similarity can share inspection experiences, thereby improving the inspection efficiency. Generate the regional knowledge transfer function, fuse the feature similarity matrix with the historical inspection experiences to obtain the regional shared experiences. This process can effectively utilize the existing inspection data, reduce repetitive labor, and improve the scientificity and accuracy of inspections.

[0095] Integrate the regional inspection urgency, regional risk degree, and regional shared experiences to construct a state vector. The state vector contains all the important information of the current inspection task and can provide a basis for subsequent decisions. At the same time, the inspection priority parameter is constructed as an action vector, indicating different inspection strategies that can be adopted in the current state.

[0096] According to the constructed state vector and action vector, establish a reward function. The reward function is used to evaluate the effects of different inspection strategies and can guide the reinforcement learning algorithm to optimize the inspection strategy. By feedback from the historical inspection data, adjust the parameters of the reward function so that it can more accurately reflect the advantages and disadvantages of the inspection strategy.

[0097] Adopt the reinforcement learning method to optimize the reward function to obtain the policy model parameters. The reinforcement learning algorithm adjusts the policy model parameters through continuous trial and error to achieve the optimal selection of the inspection strategy. Through multiple rounds of learning and optimization, finally generate an inspection execution plan with strong adaptability and high efficiency.

[0098] Exemplarily, there have been multiple equipment failures in a substation of a power company in the past year. The types of failures include transformer overheating, switchgear failures, etc. By analyzing historical failure data, an optimized hierarchical knowledge base containing failure characteristics and operating states was constructed. According to this knowledge base, the boundary of the dynamically associated sub-region was determined as a specific area around the transformer.

[0099] During actual inspections, the local state evaluator scores the risk factors in this area, obtaining a regional risk level of 0.8, an inspection time interval of 30 days, and a load level of 70%. Combining these data, it is calculated that the urgency of regional inspections is high. The feature similarity matrix shows that the similarity between this area and adjacent areas is relatively high, so historical inspection experiences can be shared.

[0100] Based on the state vector and action vector, the established reward function guides the reinforcement learning algorithm to optimize the policy model parameters, and the generated inspection execution plan is to conduct a special inspection of this area once a week to ensure that high-risk equipment receives timely attention.

[0101] In an alternative implementation, constructing a feature similarity matrix for the dynamically associated sub-region, generating a regional knowledge transfer function according to the feature similarity matrix, and integrating the regional knowledge transfer function with historical inspection experiences to obtain regional shared experiences includes: Construct the dynamically associated sub-region as a graph network model. Each dynamically associated sub-region in the graph network model corresponds to a regional node, and the regional node contains failure characteristics and operating state data. Based on the physical connection relationship between the dynamically associated sub-regions, edge connections are constructed, and the topological association features between the regional nodes are extracted through a graph convolutional network to obtain a feature similarity matrix; Calculate the similarity of adjacent regional nodes to obtain a gating coefficient, and construct a regional knowledge transfer function according to the gating coefficient and the feature similarity matrix. The regional knowledge transfer function characterizes the knowledge transfer relationship between regional nodes; Obtain positive sample pairs by constructing the features of the same dynamically associated sub-region within different time windows, obtain negative sample pairs by constructing the features of different dynamically associated sub-regions, and train a feature extraction model through a contrastive loss function to obtain historical inspection experiences; Integrate the regional knowledge transfer function with the historical inspection experiences to obtain regional shared experiences.

[0102] In a specific embodiment, based on the physical connection relationship and electrical characteristics of the device, an adaptive clustering algorithm is used to divide the dynamically associated sub-regions. First, calculate the electrical distance between any two monitoring points, including impedance distance and power flow coupling degree; then construct an affinity matrix based on the electrical distance and use the spectral clustering method for preliminary partitioning; finally, considering factors such as geographical location and operating environment, manually correct and optimize the partitioning results. The boundary of each sub-region is dynamically adjusted according to the real-time operating state to ensure that the regional division always reflects the current system operating characteristics.

[0103] When constructing the graph network model, each dynamically associated sub-region corresponds to a regional node. The feature vector of the node consists of multiple parts: the operating state data includes the statistical characteristics of electrical parameters such as voltage level, load rate, and power factor; the device state data includes the time-series characteristics of physical parameters such as temperature distribution, vibration characteristics, and insulation characteristics; the fault feature data includes historical statistical information such as fault frequency, type distribution, and severity. The edges between nodes are established based on the physical topology relationship, and the attributes of the edges include information in multiple dimensions such as line impedance, power flow magnitude, and fault correlation.

[0104] Use a multi-layer graph convolutional network to extract the topological association features between nodes. The first layer of graph convolution performs spatial feature aggregation, and combines the features of each node with those of its first-order neighbor nodes through weighted combination. The weight coefficient is calculated through an attention mechanism, considering the similarity of node features and the importance of edge attributes. The second layer of graph convolution expands the receptive field and aggregates the information of second-order neighbor nodes to capture the association patterns in a larger range. The third layer of graph convolution introduces skip connections to fuse the feature representations at different levels. After multiple non-linear transformations and feature aggregations, a feature similarity matrix reflecting the deep association relationship between regional nodes is finally obtained.

[0105] Calculating the knowledge transfer intensity between regional nodes requires considering multiple factors. Calculate the Euclidean distance and cosine similarity between node pairs based on the feature similarity matrix to obtain the initial similarity score. Then, introduce expert knowledge rules to weighted adjust the similarity score according to factors such as the similarity of the operating environment of the region, the consistency of device types, and the relevance of fault modes. For adjacent regional node pairs, calculate the gating coefficient between them, which determines the degree of knowledge transfer.

[0106] When constructing the regional knowledge transfer function, use the form of a dynamic weight matrix. Each element of the matrix represents the knowledge transfer intensity from the source region to the target region, which consists of three parts: the basic transfer weight reflects the physical connection strength between regions; the similarity weight reflects the similarity degree of regional features; the gating weight controls the selectivity of knowledge transfer. Through the combination of these three types of weights, the adaptive transfer of knowledge between regions is realized. At the same time, design a time-series attenuation mechanism so that newer knowledge has a higher transfer weight to ensure the timeliness of knowledge transfer.

[0107] In the time dimension, for each dynamically associated sub-region, characteristic sequences of different time windows are selected to construct positive sample pairs. The length of the time window is determined according to the periodic characteristics of the data and can be at the hourly, daily, or weekly level. For each time window, multi-dimensional feature vectors including equipment status, operating parameters, and environmental conditions are extracted. The construction of positive sample pairs takes into account time correlation, and the characteristic sequences of adjacent time windows are more likely to form positive sample pairs.

[0108] In the space dimension, characteristic sequences of different sub-regions are selected to construct negative sample pairs. The selection of negative samples needs to ensure sufficient distinctiveness between samples. Therefore, regions with a large physical distance and significant differences in operating characteristics are preferably selected. At the same time, considering the periodic characteristics of the power system, a sufficient number of negative sample pairs should be constructed at different time scales to ensure that the model can learn stable feature representations.

[0109] The training of the feature extraction model adopts a contrastive learning framework. The designed loss function consists of two parts: a pulling term between positive sample pairs and a pushing term between negative sample pairs. By minimizing this loss function, the feature representations learned by the model can not only maintain time continuity but also reflect spatial differences. During the training process, a hard sample mining strategy is adopted to dynamically adjust the weights of samples, enabling the model to pay more attention to those sample pairs that are difficult to distinguish.

[0110] The trained feature extraction model is combined with the regional knowledge transfer function to generate regional shared experience. This process first uses the feature extraction model to encode the historical data of each region to obtain an experience vector reflecting the characteristics of that region. Then, through the knowledge transfer function, the experience transfer intensity between different regions is calculated. The finally obtained regional shared experience not only contains the unique operating characteristics of each region but also realizes the experience complementarity between similar regions, providing a reliable knowledge basis for subsequent inspection decisions.

[0111] Exemplarily, a 500 kV transmission line system is 180 kilometers long and has 6 substations along the line. According to the equipment distribution characteristics and operating characteristics, it is divided into 4 dynamically associated sub-regions. Region A contains 2 substations and 50 kilometers of line, mainly located in mountainous terrain; Region B contains 1 substation and 40 kilometers of line, spanning a plain area; Region C contains 2 substations and 50 kilometers of line, passing through a coastal area; Region D contains 1 substation and 40 kilometers of line, located in an industrial intensive area.

[0112] Intelligent sensing devices are installed in each sub-region to continuously collect operation data. Taking Region A as an example, its node features include operation parameters such as the current load rate (75%-85%), wire temperature (45℃-55℃), insulator pollution degree (0.15-0.25) at 5 key monitoring points, as well as 12 equipment failure records that occurred in the past 3 years. When constructing the graph network, Region A establishes edge connections with adjacent Regions B and C, and the edge weights are 0.8 and 0.6 respectively, reflecting the electrical distance relationship between them.

[0113] Through graph convolutional network analysis, it is found that the feature similarity between Region A and Region B is 0.85. This is because both regions face common challenges brought by mountainous terrain, such as high lightning strike risk and high mechanical stress. The calculated gating coefficient is 0.82, indicating that strong knowledge transfer can occur between these two regions. In contrast, the feature similarity between Region A and Region C is only 0.45, and the gating coefficient is 0.38, which reflects the significant differences in the operation environment and fault characteristics between mountainous and coastal areas.

[0114] In the model training stage, the operation data of Region A in different seasons are selected to construct positive sample pairs, such as the feature sequences in winter and spring, to capture the seasonal change patterns. At the same time, the contemporaneous data of Region A and Region C are selected to construct negative sample pairs to help the model identify the feature differences under different geographical environments. Through contrastive learning, the model successfully extracts the historical inspection experience reflecting the characteristics of each region.

[0115] The finally generated regional shared experience shows that Region A and Region B can share the experience of lightning protection and mechanical stress monitoring to optimize the inspection strategy; the anti-salt fog corrosion experience of Region C is mainly used within the region and only migrates to Region D in a small range; the industrial pollution protection experience of Region D can provide some reference for other regions. This differential knowledge sharing mechanism significantly improves the inspection efficiency, and the average accuracy of fault early warning in each region is increased by 15%.

[0116] Such as Figure 4As shown, the knowledge transfer process and its effect evolution between dynamically associated sub-regions are visually demonstrated through a tree-like knowledge flow structure. The thick black line in the center represents the time axis of system experience. From the initial state to 60 weeks, the size of the circular nodes intuitively reflects the growth process of the cumulative experience of this technical solution, gradually increasing from the initial 57.2% to the final 96.3%. At the same time, the square nodes on the right side of the figure show the cumulative experience of the traditional method at the same time points. It is obvious that its growth is slow and it only reaches 77% finally. The four branches extending from the main trunk represent four groups of typical inter-regional knowledge transfer paths, where the thickness of the lines directly reflects the intensity of knowledge transfer. The A→B regional transfer path is the thickest, with a transfer efficiency as high as 0.90. This is mainly attributed to the high similarity of the mountain terrain features between the two regions, which ultimately contributed to a significant achievement of a 19.8% improvement in fault warning ability. The C→D regional transfer path is the second, with a transfer efficiency of 0.75, mainly reflected in a 15.4% improvement in salt spray corrosion detection ability. In contrast, the A→C transfer path is thinner, reflecting the differences in environmental characteristics between mountainous and coastal areas. The transfer efficiency is only 0.45, but even so, it still brings a 12.3% improvement in inspection optimization. As time goes by, the figure visually shows the accelerated accumulation process of system experience. Especially after 36 weeks, the gap between this technical solution and the traditional method further widens, highlighting the cumulative advantage of the knowledge transfer mechanism based on graph convolutional network in long-term operation. It can also be clearly seen from the figure that there is a positive correlation between the higher knowledge transfer efficiency and the significant performance improvement, confirming the effectiveness of the design of the regional knowledge transfer function in this technical solution.

[0117] In this embodiment, by constructing a graph network model and a graph convolutional network, the physical connections and topological features between dynamically associated sub-regions are effectively extracted, enhancing the understanding and modeling ability of the relationships between regions; a knowledge transfer function is constructed using the gating coefficient and feature similarity to achieve the effective transfer of experience and information between regions, providing support for fault prediction and operation decision-making; a feature extraction model is trained using positive and negative sample pairs and a contrast loss function to learn and refine key experience from historical inspection data, improving the model's ability to identify anomalies and faults; the regional knowledge transfer function is integrated with historical inspection experience to generate regional shared experience, realizing cross-regional information complementarity and enhancing the intelligent decision-making ability of the overall system.

[0118] The power line health status assessment and prediction system based on big data in the embodiment of the present invention includes: The first unit is used to obtain the historical operation data of the power line, preprocess the historical operation data according to the preset data quality rules, and generate standardized operation data; The second unit is used to extract fault features from the standardized operation data based on the spatio-temporal heterogeneous graph network, establish the mapping relationship between the features and the operation state, construct a basic hierarchical knowledge base, automatically annotate the standardized operation data, and generate a training sample set; The third unit is used to model the power line system as a directed acyclic graph, calculate the transmission path eigenvalue, identify the main transmission lines, divide the risk sections, and integrate them into the basic hierarchical knowledge base to obtain an optimized hierarchical knowledge base; The fourth unit is used to establish an inspection strategy model based on the optimized hierarchical knowledge base and the main transmission lines by means of reinforcement learning method, and adaptively generate an inspection execution plan; execute the inspection execution plan and evaluate the health state of the power line to obtain a health state evaluation result; The fifth unit is used to generate a power line health state report based on the health state evaluation result.

[0119] In the third aspect of the embodiments of the present invention, There is provided an electronic device, including: A processor; A memory for storing instructions executable by the processor; Wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0120] In the fourth aspect of the embodiments of the present invention, There is provided a computer-readable storage medium, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0121] The present invention may be a method, an apparatus, a system, and / or a computer program product. The computer program product may include a computer-readable storage medium, on which computer-readable program instructions for executing various aspects of the present invention are uploaded.

[0122] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A power line health status assessment and prediction method based on big data, characterized in that: include: Obtain historical operation data of power lines, pre-process the historical operation data according to preset data quality rules, and generate standardized operation data; Extract fault features from standardized operation data based on spatiotemporal heterogeneous graph networks, establish a mapping relationship between features and operation status, build a basic hierarchical knowledge base, automatically annotate standardized operation data, and generate a training sample set; The power line system is modeled as a directed acyclic graph, the transmission path characteristic values ​​are calculated, the trunk transmission lines are identified, the risk sections are divided, and integrated into the basic hierarchical knowledge base to obtain an optimized hierarchical knowledge base; According to the optimized hierarchical knowledge base and trunk transmission lines, the inspection strategy model is established based on the reinforcement learning method, and the inspection execution plan is adaptively generated; Execute the inspection execution plan and evaluate the health status of the power lines to obtain the health status evaluation results; Generate a power line health status report based on the health status assessment results.

2. The method according to claim 1, characterized in that Based on the spatiotemporal heterogeneous graph network, fault features are extracted from the standardized operation data, the mapping relationship between features and operation status is established, and the basic hierarchical knowledge base is constructed, including: The standardized operation data includes equipment data, environmental data and historical fault data; Based on the standardized operation data, they are respectively constructed into equipment status nodes, environmental feature nodes and historical fault nodes; Based on the physical connection relationship between device status nodes, the topological association edge between devices is established. Based on the influence relationship between device status nodes and environmental feature nodes, the device-environment association edge is established. Based on the corresponding relationship between device status nodes and historical fault nodes, the device-fault association edge is established, and together they form a heterogeneous information network. For each node in the heterogeneous information network, type-aware weighted aggregation is performed based on a multi-head attention mechanism to extract fault features, and a fault feature vector is obtained. Feature aggregation is performed through differentiable pooling guided by node importance to obtain a subgraph feature representation; Performing sliding convolution operation on the subgraph feature representation using a causal convolution kernel, generating a time series feature vector through time series-based information processing, calculating the contrast loss between normal operation state samples and the boundary loss between abnormal operation state samples, and obtaining a comprehensive abnormality score of the fault feature; The time series feature vector, the comprehensive anomaly score and the corresponding fault case are stored in the instance layer, and a mapping relationship between the fault feature and the operating status is established; the fault cases whose Euclidean distance between the time series feature vectors corresponding to the fault cases is less than a first preset threshold are aggregated into fault modes and stored in the pattern layer; the evolution relationship between the fault modes whose time series correlation degree corresponding to the fault modes is greater than a second preset threshold is stored in the rule layer; the instance layer, the pattern layer and the rule layer constitute a basic hierarchical knowledge base.

3. The method according to claim 2, characterized in that For each node in the heterogeneous information network, type-aware weighted aggregation is performed based on the multi-head attention mechanism to extract fault features, and a fault feature vector is obtained. Feature aggregation is performed through differentiable pooling guided by node importance, and the subgraph feature representation is obtained, including: According to different types of neighbor nodes of each node in the heterogeneous information network, a multi-head attention weight matrix is ​​constructed, wherein the multi-head attention weight matrix is ​​used to calculate the dynamic association strength between the node and the different types of neighbor nodes; Multiplying the original feature vector of the node by the conversion matrix corresponding to the type of the node to obtain a type-aware feature vector, and performing a weighted aggregation operation on the type-aware feature vector based on the dynamic association strength to generate a fault feature vector; The fault feature vector is concatenated with the global feature vector of the heterogeneous information network, and a node importance score is calculated through a preset multi-layer perceptron, and nodes are sampled based on the node importance score to construct a local network structure; For the fault feature vector in the local network structure, a differentiable pooling operation is used to perform feature aggregation, and the aggregated features are updated through a gated recurrent unit to obtain a subgraph feature representation.

4. The method according to claim 1, characterized in that: The power line system is modeled as a directed acyclic graph, the transmission path characteristic values ​​are calculated, the trunk transmission lines are identified, the risk sections are divided, and integrated into the basic hierarchical knowledge base. The optimized hierarchical knowledge base includes: Constructing a directed acyclic graph of the power line system, the directed acyclic graph including an operation feature node set and an electric energy transmission edge set, obtaining an operation feature value of each node in a training sample set, and setting the electric energy transmission efficiency corresponding to each edge as the edge feature value; Calculate the sum of the connected edge characteristic values ​​of each running characteristic node in the directed acyclic graph to obtain a node importance value, determine the power transmission path based on the node importance value, and calculate the product of the edge characteristic values ​​of each power transmission path to obtain a path characteristic value; Obtain the line load threshold of each operating characteristic node, identify the trunk transmission line that meets the load threshold constraint through the maximum flow algorithm with iterative capacity adjustment, and calculate the corresponding edge cut set load ratio; Obtain the failure probability of each running characteristic node and the set of downstream impact nodes, calculate the node risk value, correct the node risk value based on the edge cut set load ratio, and divide the risk section by a spectral clustering algorithm; The trunk transmission line is written into the skeleton layer of the basic hierarchical knowledge base, the protection level is divided based on the risk section, the path characteristic value and the node risk value are recorded, a hierarchical association relationship including the edge cut set load ratio is established, and the control strategy priority of the basic hierarchical knowledge base is updated based on the hierarchical association relationship to obtain an optimized hierarchical knowledge base.

5. The method according to claim 4, characterized in that Obtain the line load threshold of each running characteristic node, identify the trunk transmission line that meets the load threshold constraint through the maximum flow algorithm with iterative capacity adjustment, and calculate the corresponding edge cut set load ratio, including: Obtain the line load threshold of each operating characteristic node in the power system network, construct the initial capacity matrix, and obtain the initial maximum flow value through the maximum flow algorithm under the constraint that the actual load is less than or equal to the line load threshold. Based on the initial maximum flow value, determine the corresponding minimum cut set to obtain the initial trunk transmission line; Apply a disturbance coefficient to the initial trunk transmission line, modify the initial capacity matrix to obtain a modified capacity matrix, calculate the maximum flow value after the disturbance, and calculate the ratio to the initial maximum flow value to obtain the network fitness index; Based on the difference between the network fitness index and the preset fitness threshold, the modified capacity matrix is ​​updated to obtain the iterative capacity matrix, and the iterative maximum flow value corresponding to the iterative capacity matrix is ​​calculated. The iteration is repeated until the rate of change of the maximum flow values ​​of two adjacent iterations is less than the preset change rate threshold. The final capacity matrix is ​​obtained, and the lines that meet the line load threshold are selected from the final capacity matrix as the trunk transmission optimization lines, and the corresponding edge cut set load ratio is calculated.

6. The method according to claim 1, characterized in that According to the optimized hierarchical knowledge base and trunk transmission lines, the inspection strategy model is established based on the reinforcement learning method, and the inspection execution plan is adaptively generated, including: Based on the fault characteristics in the optimized hierarchical knowledge base, the mapping relationship with the operating status and the transmission path characteristic value of the trunk transmission line, a regional characteristic function is constructed, and the boundary of the dynamic associated sub-region is determined according to the regional characteristic function; A local state evaluator is constructed for the dynamically associated sub-areas, and the risk factor score of the local state evaluator is multiplied by the weight coefficient to obtain the regional risk degree. The regional risk degree is combined with the inspection time interval and load level to calculate the regional inspection urgency. Constructing a feature similarity matrix of dynamically associated sub-regions, generating a regional knowledge transfer function based on the feature similarity matrix, and integrating the regional knowledge transfer function with historical inspection experience to obtain regional shared experience; The regional inspection urgency, regional risk, and regional shared experience are used to construct a state vector, the inspection priority parameter is used to construct an action vector, and a reward function is established based on the state vector and the action vector; The reward function is optimized by a reinforcement learning method to obtain the policy model parameters, and an inspection execution plan is generated according to the policy model parameters.

7. The method according to claim 6, characterized in that Constructing a feature similarity matrix of dynamically associated sub-regions, generating a regional knowledge transfer function based on the feature similarity matrix, and integrating the regional knowledge transfer function with historical inspection experience to obtain regional shared experience includes: The dynamically associated sub-regions are constructed as a graph network model, each of the dynamically associated sub-regions in the graph network model corresponds to a region node, the region node contains fault characteristics and operating status data, edge connections are constructed based on the physical connection relationship between the dynamically associated sub-regions, and the topological association features between the region nodes are extracted through a graph convolutional network to obtain a feature similarity matrix; Calculating the similarity of the adjacent regional nodes to obtain a gating coefficient, and constructing a regional knowledge transfer function according to the gating coefficient and the feature similarity matrix, wherein the regional knowledge transfer function represents the knowledge transfer relationship between regional nodes; Acquire features of the same dynamically associated sub-region in different time windows to construct positive sample pairs, acquire features of different dynamically associated sub-regions to construct negative sample pairs, and train a feature extraction model through a comparative loss function to obtain historical inspection experience; The regional knowledge transfer function is integrated with the historical inspection experience to obtain regional shared experience.

8. A power line health status assessment and prediction system based on big data, used to implement the method described in any one of claims 1 to 7, characterized in that: include: The first unit is used to obtain historical operation data of the power line, pre-process the historical operation data according to a preset data quality rule, and generate standardized operation data; The second unit is used to extract fault features from standardized operation data based on the spatiotemporal heterogeneous graph network, establish a mapping relationship between features and operation status, build a basic hierarchical knowledge base, automatically annotate the standardized operation data, and generate a training sample set; The third unit is used to model the power line system as a directed acyclic graph, calculate the transmission path characteristic value, identify the trunk transmission line, divide the risk section, and integrate it into the basic hierarchical knowledge base to obtain an optimized hierarchical knowledge base; The fourth unit is used to establish an inspection strategy model based on the optimization layered knowledge base and the trunk transmission line based on the reinforcement learning method, and adaptively generate an inspection execution plan; Execute the inspection execution plan and evaluate the health status of the power lines to obtain the health status evaluation results; The fifth unit is used to generate a power line health status report based on the health status assessment result.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images