A power distribution operation and maintenance strategy optimization method based on big data analysis

Abstract

The application discloses a power distribution operation and maintenance strategy optimization method based on big data analysis, which comprises the following steps: collecting multi-source operation data, preprocessing to generate a feature tensor and a resource constraint set; constructing a power distribution network heterogeneous high-order graph to represent the electrical and functional correlation of equipment; inputting an improved heterogeneous high-order graph attention network model to generate a risk score matrix and a health index matrix; constructing an equipment risk matrix to generate a preliminary strategy combination; generating an optimal strategy set through multi-objective hyper-heuristic algorithm and iterative optimization of the preliminary strategy combination; selecting a final operation and maintenance scheme, performing inspection and maintenance, and collecting feedback data; inputting the feedback data and the latest state data to update the risk score matrix and the optimal strategy set. Through the improved heterogeneous high-order graph attention network and the multi-objective hyper-heuristic optimization algorithm, the power distribution network node risk score calculation and the closed-loop adaptive optimization of the Pareto optimal operation and maintenance strategy are realized.

Description

Technical Field

[0001] This invention relates to the field of power distribution network operation and maintenance and intelligent dispatching technology, and in particular to a method for optimizing power distribution operation and maintenance strategies based on big data analysis. Background Technology

[0002] In the existing field of power distribution network operation and maintenance, with the increase in the number of devices and the complexity of network structures, traditional manual inspection and periodic maintenance methods are struggling to meet the requirements of high reliability and efficiency. Existing technologies largely rely on human experience or simple rules, limiting their ability to comprehensively analyze equipment operating status, load fluctuations, and environmental information, and making it difficult to fully consider the electrical connections and functional couplings between devices. Furthermore, existing strategies are mostly static planning, lacking dynamic updating capabilities and failing to adaptively adjust based on real-time data and node risks, leading to delays in the operation and maintenance of critical equipment or high-risk nodes. Existing methods typically remain at the single-channel or independent analysis stage in multi-source data processing, failing to uniformly integrate and standardize operating status, historical fault records, load data, and environmental information, and also failing to fully model high-order electrical structures and functional relationships between nodes, thus affecting the overall accuracy and reliability of operation and maintenance decisions.

[0003] In the strategy optimization and execution phases, existing technologies mostly employ traditional single-objective or multi-objective heuristic algorithms, which insufficiently consider complex constraints and functional dependencies between strategies. They lack analysis of historical strategy execution trajectories and node risk evolution, resulting in limitations in the feasibility, resource utilization, and risk control of the generated operation and maintenance strategies. Execution feedback data and device status information are not used for timely strategy iteration and updates, the operation and maintenance system lacks closed-loop adaptive capabilities, and it cannot form a Pareto optimal set of operation and maintenance strategies for multi-objective trade-offs. Furthermore, traditional methods lack interpretability in strategy generation and node risk assessment, making it difficult for operation and maintenance personnel to understand the decision-making basis. Therefore, existing technologies have significant shortcomings in multi-source data modeling, high-order topology analysis, strategy optimization, and closed-loop adaptive operation and maintenance.

[0004] Therefore, how to provide a method for optimizing power distribution operation and maintenance strategies based on big data analysis is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a distribution network operation and maintenance strategy optimization method based on big data analysis. This invention collects multi-source operational data from distribution network equipment and combines it with resource configuration data from the operation and maintenance system to construct a multi-channel feature tensor and a set of operation and maintenance resource constraints. This abstracts the equipment and its electrical and functional associations into a heterogeneous high-order graph of the distribution network. Using an improved heterogeneous high-order graph attention network model, heterogeneous semantic relationship attention calculations are performed on nodes, their neighboring nodes, and their respective high-order structures to generate node embedding representations. The node embeddings are iteratively updated through prediction-driven attention feedback. The embedding vectors are then interpreted to generate node risk score matrices and health index matrices. The node risk score matrices and operation and maintenance resource constraints are input into a multi-objective hyperheuristic optimization algorithm. Combined with a strategy collaboration metric matrix and Pareto potential field information, a preliminary strategy combination is generated. Through iterative optimization, combining the strategy mutual information matrix and two-level closed-loop feedback, a Pareto optimal operation and maintenance strategy set is formed. The execution feedback data and the latest equipment operating status are input into the model and algorithm for closed-loop updates, achieving dynamic adjustment of node risk scores and strategy combinations. This invention performs unified analysis of complex multi-source data and high-order topology structures, closely integrates node risk assessment with operation and maintenance strategy optimization, realizes dynamic strategy generation and closed-loop adaptive update, and has the advantages of comprehensive data processing, accurate strategy, closed-loop adaptiveness, strong interpretability and reasonable resource utilization.

[0006] A method for optimizing power distribution operation and maintenance strategies based on big data analysis according to an embodiment of the present invention includes: Collect multi-source operation data of distribution network equipment, perform preprocessing to generate multi-channel feature tensors, obtain resource configuration data of distribution network operation and maintenance system, and form a set of operation and maintenance resource constraints. A heterogeneous high-order graph of the distribution network is constructed based on the multi-channel feature tensor, where nodes represent devices and edges and high-order structures represent the electrical and functional relationships between devices. The heterogeneous high-order graph of the distribution network is input into the improved heterogeneous high-order graph attention network model. Heterogeneous semantic relationship attention calculation is performed to generate node embedding representations. Node risk prediction output is calculated, attention allocation is adjusted, node embedding representations are decomposed into interpretable factors, and node risk scoring matrix and health index matrix are generated. The device risk matrix is ​​constructed based on the node risk scoring matrix and the health index matrix. The device risk matrix and the set of operation and maintenance resource constraints are input into the multi-objective hyperheuristic optimization algorithm. The high-level strategy selector combines the strategy collaboration metric matrix and Pareto potential energy field information to generate a preliminary strategy combination. The initial strategy combination is iteratively optimized using a multi-objective hyperheuristic optimization algorithm. The strategy subgroup structure is dynamically adjusted based on the strategy mutual information matrix. The high-level strategy selector is updated by combining two-level closed-loop feedback to generate the Pareto optimal operation and maintenance strategy set. Select the final operation and maintenance solution from the Pareto optimal operation and maintenance strategy set, send it to the distribution network operation and maintenance system to perform inspection and maintenance operations, and collect execution feedback data; The execution feedback data and the latest equipment operating status data are input into the improved heterogeneous high-order graph attention network model and multi-objective hyperheuristic optimization algorithm to update the node risk score matrix and the Pareto optimal operation and maintenance strategy set.

[0007] Optionally, the generation of the multi-channel feature tensor and the set of operation and maintenance resource constraints includes: The multi-source operation data of the distribution network equipment are sorted according to the equipment number and the collection time, including operation status data, historical fault records, load data and environmental information, to generate a data sequence arranged by equipment number and time. Perform missing value imputation on the sorted data sequence, and use the historical average value to fill in the missing data items; Normalization is performed on the filled data sequence to convert the value of each feature, including voltage, current, load, and environment, to a uniform range, forming a standardized feature sequence. The normalized feature sequences are integrated according to device number, time order and data type to form a multi-channel feature tensor, where each channel corresponds to a data type; Obtain resource configuration data for the power distribution network operation and maintenance system, including the number of inspection personnel, the number of operation and maintenance equipment, and the available time periods, and organize them into a set of operation and maintenance resource constraints.

[0008] Optionally, the construction of the heterogeneous high-order graph of the distribution network includes: Based on multi-channel feature tensors, distribution network equipment is converted into nodes. Equipment number, equipment type, equipment operating status data, historical fault records, load data and environmental information are written into the corresponding nodes to generate a node set and establish a unique identifier for each node. Based on the node set, edges are created, and the electrical parameters, line impedance, and load capacity between nodes are written into the corresponding edges to generate an edge set. The starting node number and ending node number of each edge are recorded. Generate an edge feature matrix based on the edge set, with rows corresponding to edge numbers and columns corresponding to edge attributes, including electrical parameters, line impedance, load capacity, and historical fault data; Identify higher-order structures in the edge set, including three-node triangles, four-node rings and bridging nodes, record the higher-order structure number and the set of constituent nodes, and generate a set of higher-order structures. The set of nodes, the set of edges, and the set of higher-order structures are encoded into node feature matrices, edge feature matrices, and higher-order structure feature matrices, respectively. These matrices are then mapped to different channels according to the categories of nodes, edges, and higher-order structures to generate graph structure tensors. The node feature matrix, edge feature matrix, and higher-order structure feature matrix are integrated to generate a heterogeneous higher-order graph of the distribution network, including a set of nodes, a set of edges, a set of higher-order structures, and the graph structure tensor.

[0009] Optionally, the generation of the node risk score and health index matrix includes: The heterogeneous high-order graph of the distribution network is input into the improved heterogeneous high-order graph attention network model, which includes a high-order attention calculation unit, a prediction-driven attention feedback unit, and an interpretable attention decomposition unit. The higher-order attention computation unit performs heterogeneous semantic relationship attention computation on each node, its neighboring nodes and its higher-order structure based on the node set, edge set and higher-order structure set. It generates attention weights and performs weighted aggregation to form a node embedding vector matrix, with rows corresponding to node numbers and columns corresponding to embedding feature dimensions. The prediction-driven attention feedback unit calculates the node risk prediction output based on the node embedding vector matrix, feeds the node risk prediction output back to the higher-order attention calculation unit, updates the node embedding vector matrix, and adjusts the aggregation ratio of node features in neighbors and higher-order structures in each iteration. The interpretable attention decomposition unit decomposes the updated node embedding vector matrix according to the feature dimension to generate an interpretable factor matrix. The component information is recorded according to the higher-order structure type, with rows corresponding to node numbers and columns corresponding to factor numbers. Each element represents the component of the node embedded in that factor. The node risk score matrix is ​​calculated based on the interpretable factor matrix. The rows correspond to the node numbers and the columns correspond to the risk dimensions. Each element is generated by the value of the node embedding vector on the corresponding factor. The health index matrix is ​​calculated based on the node embedding vector matrix and the node risk score matrix. The rows correspond to the node numbers and the columns correspond to the health index dimensions. Each element is generated by combining the node embedding vector and the risk score. Normalization is performed on the node risk score matrix and health index matrix to generate the final node risk score matrix and health index matrix; An improved heterogeneous high-order graph attention network model is trained. An optimization objective is constructed based on the joint deviation between the node embedding vector matrix and the node risk prediction output. The parameters in the high-order attention computation unit, the prediction-driven attention feedback unit, and the interpretable attention decomposition unit are continuously updated iteratively. During the training process, the aggregation weights of neighbor and high-order structural features are dynamically adjusted according to the mean and variance of node embeddings after each iteration. Training stops when the average change of the optimization objective in five consecutive iterations is less than 1% of the variance of each embedding feature.

[0010] Optionally, the generation of the initial strategy combination includes: The equipment risk matrix and the set of operation and maintenance resource constraints are input into a multi-objective hyperheuristic optimization algorithm to initialize the strategy set, generate a unique code for each strategy, form a strategy code set, and generate a strategy embedding vector matrix. Based on the policy encoding set and policy embedding vector matrix, the local density and Pareto potential value of each policy in the multi-objective index space are calculated to form a policy collaboration metric matrix. The rows correspond to the policy numbers, the columns correspond to the collaborative scores between policies, and the policy numbers that are related to the current policy in the multi-objective space are recorded to generate a neighborhood policy table. The low-level policy set is divided according to the policy collaboration metric matrix and the neighborhood policy table to form a policy subgroup structure matrix. The rows correspond to the subgroup numbers and the columns correspond to the policy numbers. Each element represents the correlation degree of the policy in the subgroup. During the division process, the internal order of the subgroup is determined by combining the policy embedding vector and the Pareto potential value. The strategy subgroup structure matrix is ​​combined with the strategy synergy metric matrix to generate a candidate strategy sequence according to the subgroup order and synergy score. Candidate policy sequences are mapped to form a preliminary policy combination matrix, with rows corresponding to policy combination numbers and columns corresponding to combination feature dimensions. Each element represents the index value of the policy in the combination, ultimately yielding the preliminary policy combination.

[0011] Optionally, generating the Pareto optimal operation and maintenance strategy set includes: The initial strategy combination, equipment risk matrix, and set of operation and maintenance resource constraints are input into a multi-objective hyperheuristic optimization algorithm to initialize the strategy iteration population, generate a unique code for each strategy, form a strategy code set, and construct a strategy embedding vector matrix, with rows corresponding to the strategy combination number and columns corresponding to the multi-objective indicator dimensions. The policy fitness value is calculated based on the policy encoding set and the policy embedding vector matrix to form a policy fitness matrix. The rows correspond to the policy combination number, and the columns correspond to multi-objective indicators, including node risk scores and resource consumption values. A policy mutual information matrix is ​​generated, and the elements record the correlation between each policy in the multi-objective indicator space. Based on the policy fitness matrix, non-dominated sorting and crowding calculation are performed to stratify the policies, filter and generate a set of candidate Pareto front policies, and iteratively update the subgroup structure of the candidate policies by combining the policy mutual information matrix. Perform crossover, mutation, and heuristic recombination operations on the candidate Pareto frontier policy set to update the policy combination feature values. Combine the two-level closed-loop feedback to update the state of the high-level policy selector and generate a new round of policy combination set. The iterative process of repeating fitness calculation, Pareto front screening and strategy combination update continues until the average change of each strategy in the multi-objective indicators of the candidate Pareto front strategy set is less than 1% in three consecutive iterations, and finally the Pareto optimal operation and maintenance strategy set is generated. The generated Pareto optimal operation and maintenance strategy set is normalized to form the final Pareto optimal operation and maintenance strategy set.

[0012] Optionally, the collected execution feedback data includes: The inspection and maintenance operation information performed by the distribution network operation and maintenance system is recorded as an operation log. The log includes the strategy number, the execution device number, the operation start time, and the operation end time. Collect operational status data of each device during inspection and maintenance operations, including device sensor readings, load data, environmental information and fault event records, and generate an operational status sequence by sorting by device number and timestamp; Node risk feedback data is generated based on the sequence of executed operations and running status. The strategy number corresponding to the executed operation is associated with the risk score and health index of each device to form a node risk feedback matrix, with rows corresponding to device numbers and columns corresponding to strategy numbers and risk indicators. The node risk feedback matrix is ​​integrated with operation logs and equipment operating status sequences to generate execution feedback data.

[0013] Optionally, the updated node risk scoring matrix and the Pareto optimal operation and maintenance strategy set include: The execution feedback data and the latest equipment operating status data are input into the improved heterogeneous high-order graph attention network model and multi-objective hyperheuristic optimization algorithm to update the node embedding vector and interpretable factor matrix; A new node risk score matrix and health index matrix are generated based on the updated node embedding vector and interpretable factor matrix; The updated node risk score matrix and the latest equipment operating status data are input into a multi-objective hyperheuristic optimization algorithm to iteratively optimize the strategy combination, update the node risk score matrix and the Pareto optimal operation and maintenance strategy set, and repeat the iteration until convergence to achieve closed-loop adaptive operation and maintenance.

[0014] The beneficial effects of this invention are: This invention proposes a power distribution operation and maintenance strategy optimization method based on big data analysis. By uniformly collecting and processing multi-source operation data of power distribution network equipment, and combining it with the resource allocation of the operation and maintenance system, a multi-channel feature tensor and a set of operation and maintenance resource constraints are constructed to realize the systematic mapping between equipment status and resource information. This enables the generation of node embedding representation, node risk scoring matrix and health index matrix to have consistency and quantifiability, effectively solving the problem that traditional operation and maintenance strategies are difficult to integrate multi-source data and high-order topology structures.

[0015] This invention constructs a heterogeneous high-order graph of a distribution network and introduces an improved heterogeneous high-order graph attention network model. It employs high-order attention computation, prediction-driven attention feedback, and interpretable attention decomposition to iteratively aggregate and decompose node and neighboring node features and high-order structural characteristics, enabling the dynamic generation and updating of node risk scores and health indices. Simultaneously, the node risk score matrix and O&M resource constraints are input into a multi-objective hyperheuristic optimization algorithm. Combined with the policy collaboration metric matrix and Pareto potential value, a preliminary policy combination is generated. Through iterative optimization, incorporating the policy mutual information matrix and two-level closed-loop feedback, a Pareto-optimal O&M policy set is formed.

[0016] This invention enables accurate assessment of node risks, optimized generation of preliminary strategy combinations, and closed-loop adaptive updating of Pareto optimal strategies. It closely integrates node risk assessment with operation and maintenance strategy optimization, realizing dynamic generation of operation and maintenance strategies, reasonable resource scheduling, and risk control, thereby improving the overall accuracy of strategies, closed-loop adaptive capabilities, comprehensiveness of data processing, interpretability, and resource utilization efficiency. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a power distribution operation and maintenance strategy optimization method based on big data analysis proposed in this invention; Figure 2 This is a schematic diagram of the structure of an improved heterogeneous high-order graph attention network model for a power distribution operation and maintenance strategy optimization method based on big data analysis proposed in this invention. Figure 3 This is a schematic diagram of the iterative process of a multi-objective hyperheuristic optimization algorithm for a power distribution operation and maintenance strategy optimization method based on big data analysis proposed in this invention. Detailed Implementation

[0018] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0019] refer to Figure 1 , Figure 2 and Figure 3 A method for optimizing power distribution operation and maintenance strategies based on big data analysis includes: Collect multi-source operation data of distribution network equipment, perform preprocessing to generate multi-channel feature tensors, obtain resource configuration data of distribution network operation and maintenance system, and form a set of operation and maintenance resource constraints. A heterogeneous high-order graph of the distribution network is constructed based on the multi-channel feature tensor, where nodes represent devices and edges and high-order structures represent the electrical and functional relationships between devices. The heterogeneous high-order graph of the distribution network is input into the improved heterogeneous high-order graph attention network model. Heterogeneous semantic relationship attention calculation is performed to generate node embedding representations. Node risk prediction output is calculated, attention allocation is adjusted, node embedding representations are decomposed into interpretable factors, and node risk scoring matrix and health index matrix are generated. The device risk matrix is ​​constructed based on the node risk scoring matrix and the health index matrix. The device risk matrix and the set of operation and maintenance resource constraints are input into the multi-objective hyperheuristic optimization algorithm. The high-level strategy selector combines the strategy collaboration metric matrix and Pareto potential energy field information to generate a preliminary strategy combination. The initial strategy combination is iteratively optimized using a multi-objective hyperheuristic optimization algorithm. The strategy subgroup structure is dynamically adjusted based on the strategy mutual information matrix. The high-level strategy selector is updated by combining two-level closed-loop feedback to generate the Pareto optimal operation and maintenance strategy set. Select the final operation and maintenance solution from the Pareto optimal operation and maintenance strategy set, send it to the distribution network operation and maintenance system to perform inspection and maintenance operations, and collect execution feedback data; The execution feedback data and the latest equipment operating status data are input into the improved heterogeneous high-order graph attention network model and multi-objective hyperheuristic optimization algorithm to update the node risk score matrix and the Pareto optimal operation and maintenance strategy set.

[0020] In this embodiment, generating the multi-channel feature tensor and the set of operation and maintenance resource constraints includes: The multi-source operation data of the distribution network equipment are sorted according to the equipment number and the collection time, including operation status data, historical fault records, load data and environmental information, to generate a data sequence arranged by equipment number and time. Perform missing value imputation on the sorted data sequence, and use the historical average value to fill in the missing data items; Normalization is performed on the filled data sequence to convert the value of each feature, including voltage, current, load, and environment, to a uniform range, forming a standardized feature sequence. The normalized feature sequences are integrated according to device number, time order and data type to form a multi-channel feature tensor, where each channel corresponds to a data type; Obtain resource configuration data for the power distribution network operation and maintenance system, including the number of inspection personnel, the number of operation and maintenance equipment, and the available time periods, and organize them into a set of operation and maintenance resource constraints.

[0021] In this embodiment, the construction of the heterogeneous high-order graph of the distribution network includes: Based on multi-channel feature tensors, distribution network equipment is converted into nodes. Equipment number, equipment type, equipment operating status data, historical fault records, load data and environmental information are written into the corresponding nodes to generate a node set and establish a unique identifier for each node. Based on the node set, edges are created, and the electrical parameters, line impedance, and load capacity between nodes are written into the corresponding edges to generate an edge set. The starting node number and ending node number of each edge are recorded. Generate an edge feature matrix based on the edge set, with rows corresponding to edge numbers and columns corresponding to edge attributes, including electrical parameters, line impedance, load capacity, and historical fault data; Identify higher-order structures in the edge set, including three-node triangles, four-node cycles, and bridging nodes. Record the higher-order structure number and the set of nodes that make up the structure, and generate a set of higher-order structures, where: Identifying higher-order structures within an edge set specifically involves: Iterate through each node and its neighboring nodes in ascending order of node number, and check whether any three nodes are connected in pairs. If the condition is met, mark it as a three-node triangle. Traverse each combination of four nodes in ascending order and determine whether there is a closed cycle with the first and last nodes connected and the edge being unique. If the condition is met, mark it as a four-node cycle. For bridging nodes, identify nodes that simultaneously connect two non-directly connected subgraphs and mark them as bridging nodes; During the traversal, a unique number is assigned to each higher-order structure, and the corresponding set of constituent nodes is recorded in the set of higher-order structures. The node set, edge set, and higher-order structure set are encoded into node feature matrices, edge feature matrices, and higher-order structure feature matrices, respectively. These matrices are then mapped to different channels according to the node, edge, and higher-order structure categories to generate a graph structure tensor. Specifically, the generated graph structure tensor is as follows: The rows of the node feature matrix correspond to the node number of each device, and the columns correspond to the node feature dimensions, including device type, operating status, load data, historical fault statistics and environmental information, forming a total of 128-dimensional node feature vector; The rows of the edge feature matrix correspond to the edge numbers, and the columns correspond to the edge attribute dimensions, including electrical parameters, line impedance, load capacity, and historical fault data, forming a total of 64-dimensional edge feature vectors. The rows of the higher-order structure feature matrix correspond to the higher-order structure number, and the columns correspond to the higher-order structure combination feature dimension. By weighted average aggregation of the constituent node features and edge features, topological attributes are introduced to form a total of 32-dimensional higher-order structure feature vector. Nodes, edges, and higher-order structure matrices are mapped to different channels. Node features occupy the first 128 channels, edge features occupy 64 channels, and higher-order structure features occupy 32 channels. The node feature matrix, edge feature matrix, and higher-order structure feature matrix are integrated to form a graph structure tensor. The tensor has the dimension of [number of nodes + number of edges + number of higher-order structures × corresponding feature dimension × number of channels], and is uniformly encoded using row order and channel order. By integrating the node feature matrix, edge feature matrix, and higher-order structure feature matrix, a heterogeneous higher-order graph of the distribution network is generated, including a set of nodes, a set of edges, a set of higher-order structures, and the graph structure tensor, wherein: Generate a heterogeneous high-order graph of the distribution network, specifically as follows: The node feature matrix, edge feature matrix, and higher-order structure feature matrix are arranged in ascending order by node number, edge number, and higher-order structure number, respectively. The three matrices are concatenated along the row direction in the tensor dimension to form a unified tensor. The matrix columns correspond to their respective feature dimensions. The channels are divided according to the category: node features occupy the first 128 channels, edge features occupy the middle 64 channels, and high-order structure features occupy the last 32 channels. The tensor dimension of the integrated graph structure is [total number of nodes + total number of edges + total number of higher-order structures × feature dimension × number of channels]. The node set, edge set, and higher-order structure set numbers, constituent node indexes, and edge indexes are recorded in the graph structure, generating a heterogeneous higher-order graph of the distribution network.

[0022] In this embodiment, the generation of the node risk score and health index matrix includes: The heterogeneous high-order graph of the distribution network is input into the improved heterogeneous high-order graph attention network model, which includes a high-order attention computation unit, a prediction-driven attention feedback unit, and an interpretable attention decomposition unit, wherein: The improved heterogeneous high-order graph attention network model is as follows: Based on the node embedding computation structure of the traditional heterogeneous high-order graph attention network model, the high-order attention computation unit introduces heterogeneous semantic relation attention, which performs weighted aggregation of the features of nodes and their neighboring nodes in different types of high-order structures, encodes the attention weights into channels according to the high-order structure type, and generates a node embedding vector matrix. The prediction-driven attention feedback unit inputs the risk prediction output of the node embedding vectors into the high-order attention computation unit to update the node embedding vector matrix. The interpretable attention decomposition unit decomposes the updated node embedding vector matrix according to the feature dimension to generate an interpretable factor matrix, which records the factor weights corresponding to different high-order structure types. The high-order attention computation unit, the prediction-driven attention feedback unit, and the interpretable attention decomposition unit are connected in the order of data flow, and the output of each module serves as the input of the next module, forming an improved heterogeneous high-order graph attention network model. The higher-order attention computation unit, based on the node set, edge set, and higher-order structure set, performs heterogeneous semantic relationship attention computation on each node, its neighboring nodes, and its associated higher-order structure. This generates attention weights, which are then weighted and aggregated to form a node embedding vector matrix. Rows correspond to node numbers, and columns correspond to embedding feature dimensions. Where: The higher-order attention computation unit includes: Node aggregation table: Records the feature index and feature value of each node and its neighboring nodes by node number; Structural Feature Encoding Table: Stores the constituent node indices, edge indices, and encoded feature vectors for each higher-order structure; Heterogeneous attention weight table: contains the attention weight values ​​between different types of higher-order structures and the features of neighboring nodes; Weighted computation array: Based on the heterogeneous attention weights, the node feature vectors and higher-order structure feature vectors are linearly superimposed and weighted to generate a node embedding vector matrix; Channel mapping matrix: maps the node and higher-order structure embedding vectors to tensor channels; Skip-level embedding cache table: records embedding information of multi-level nodes and higher-order structures; Embedded Vector Cache Table: Stores the node embedding vector matrix after higher-order attention calculation, with rows corresponding to node numbers and columns corresponding to embedding feature dimensions; In the higher-order attention computation unit, the information recorded in the node aggregation table and the structural feature encoding table, together with the weight values ​​in the heterogeneous attention weight table, are input into the weighted computation array to generate a node embedding vector matrix. This matrix is ​​then mapped to a tensor channel through a channel mapping matrix. The skip-level embedding cache table integrates the node embedding vectors and higher-order structural features from different levels, and passes the integrated embedding information to the embedding vector cache table to store the final node embedding vector matrix. Heterogeneous semantic relation attention computation is performed on each node, its neighboring nodes, and its corresponding higher-order structure, specifically as follows: Each node is taken out sequentially from the node set, and the neighbor node embedding vectors are obtained according to the skip-level embedding cache table to form a neighbor feature submatrix; Based on the set of higher-order structures, the embedding vectors of the higher-order structures to which the nodes belong are extracted from the structural feature encoding table to form a higher-order structure feature submatrix, with each row corresponding to a feature vector of a higher-order structure. Obtain the attention weight vector of the node's neighboring nodes and higher-order structures from the heterogeneous attention weight table; The neighbor feature submatrix is ​​weighted and summed with the corresponding attention weight vector for each row to obtain the neighbor weight vector. The higher-order structure feature submatrix is ​​weighted and summed in the same way with respect to the attention weight vector to obtain the higher-order structure weight vector. The node embedding vector, the neighbor weighted vector, and the higher-order structure weighted vector are linearly combined to generate the initial embedding vector of the node. Following the node numbering order, the embedding vector of each node is sequentially filled into the corresponding row of the node embedding vector matrix to form a complete node embedding vector matrix. The prediction-driven attention feedback unit calculates the node risk prediction output based on the node embedding vector matrix, feeds the node risk prediction output back to the higher-order attention calculation unit, updates the node embedding vector matrix, and adjusts the aggregation ratio of node features in neighbors and higher-order structures in each iteration, where: The prediction-driven attention feedback unit includes: Node risk prediction matrix: Stores the risk prediction value for each node, with rows corresponding to node numbers and columns corresponding to risk dimensions; Embedded update coefficient table: records the update coefficients of each node in its neighboring nodes and its corresponding higher-order structure aggregation; Risk feedback mapping table: Establishes the correspondence between the node risk prediction matrix and the rows of the node embedding vector matrix; Iterative array update: Update the node embedding vector matrix row by row according to the embedding update coefficient table and the risk feedback mapping table; Update the cache table: record the node embedding vector matrix after each iteration; Neighbor higher-order feature weight table: stores the feature aggregation ratio of a node in its neighboring nodes and its corresponding higher-order structure in each iteration; In the prediction-driven attention feedback unit, the node risk prediction matrix and the embedding update coefficient table are applied to the node embedding vector matrix through the risk feedback mapping table; the iterative update array generates a new node embedding vector matrix, which is written into the update cache table, and the aggregation ratio for the next round is updated according to the neighbor high-order feature weight table; The output of the computation node risk prediction is as follows: Each node extracts its embedding row vectors from the node embedding vector matrix to form a set of node feature vectors; Input the set of node feature vectors into the corresponding row of the node risk prediction matrix, and obtain the predicted values ​​of the node in each risk dimension through linear mapping; The generated node risk prediction values ​​are arranged by node number and combined to form a complete node risk prediction matrix. In each iteration, the aggregation ratio of node features in neighbors and higher-order structures is adjusted, specifically as follows: In the iterative update array, each node sequentially reads the neighbor feature submatrix and the higher-order structure feature submatrix, and takes the corresponding rows of the embedding update coefficient table and the neighbor higher-order feature weight table; Based on the values ​​in the node risk prediction matrix, update the weight values ​​in the embedding update coefficient table and the neighbor higher-order feature weight table to generate new neighbor weight vectors and higher-order structure weight vectors. The neighbor feature submatrix is ​​weighted and summed with the corresponding updated neighbor weight vector for each row to generate a neighbor weight vector. The higher-order structure feature submatrix is ​​weighted and summed with the updated higher-order structure weight vector to generate a higher-order structure weight vector. The node embedding vector is linearly combined with the neighbor weighted vector and the higher-order structure weighted vector to form the updated node embedding vector; Write each updated node embedding vector into the update cache table in the order of node number, with the row corresponding to the node number and the column corresponding to the embedding feature dimension. The interpretable attention decomposition unit decomposes the updated node embedding vector matrix along the feature dimension, generating an interpretable factor matrix. Component information is recorded according to the higher-order structure type, with rows corresponding to node numbers and columns corresponding to factor numbers. Each element represents the component of a node embedded in that factor, where: Interpretable attention decomposition units include: Node embedding vector cache table: Stores updated node embedding vectors, with rows corresponding to node numbers and columns corresponding to embedding feature dimensions; Factor index list: Records the set of embedded feature indexes corresponding to each interpretable factor; Higher-order structure type mapping dictionary: maps factor numbers to corresponding higher-order structure types; Sparse matrix of factor components: intermediate storage of the value of each node on each factor during the decomposition process; Decomposition result cache queue: Stores the factor component matrix output of each round of decomposition in iterative order; Mapping index matrix: During the decomposition process, the feature columns of the node embedding vector matrix are mapped to the columns of the factor component matrix; Factor-Neighborhood Association Table: Records the neighboring nodes and higher-order structure set of each factor and its influence; Interpretable factor matrix: The final output matrix, with rows corresponding to node numbers and columns corresponding to factor numbers. Each element is the component of the node embedded in the factor, which is generated by integrating the factor component sparse matrix and the mapping index matrix. In the interpretable attention decomposition unit, the node embedding vector matrix is ​​input into the node embedding cache table. The node embedding vector is decomposed into a sparse matrix of factor components through the factor index list and the mapping index matrix. The decomposition result cache queue records the output of each iteration. The factor-neighborhood association table provides high-order structure reference information. The sparse matrix and the mapping index matrix are integrated to generate an interpretable factor matrix. The node risk score matrix is ​​calculated based on the interpretable factor matrix. Rows correspond to node numbers, and columns correspond to risk dimensions. Each element is generated by the value of the node embedding vector on the corresponding factor, where: The node risk scoring matrix is ​​calculated as follows: Each node corresponds to a row vector in the interpretable factor matrix. The components of the node on each factor are arranged in columns to form a set of node factor vectors. Based on the mapping relationship between risk dimensions and factor numbers, the factor vectors of each node are linearly combined to generate the node's value on each risk dimension. Combine the risk vectors of all nodes in order of node number to form a node risk scoring matrix, with rows corresponding to node numbers and columns corresponding to risk dimensions; The health index matrix is ​​calculated based on the node embedding vector matrix and the node risk score matrix. Rows correspond to node numbers, and columns correspond to health index dimensions. Each element is generated by combining the node embedding vector and the risk score. Specifically, the calculation of the health index matrix is ​​as follows: Each node corresponds to a row vector in the node embedding vector matrix and a row vector in the node risk scoring matrix; By using a mapping matrix, the feature dimensions of the node embedding vector are calculated with the corresponding risk dimension values ​​of the node risk scoring matrix to generate a node health vector. Arrange the health vectors of all nodes in order of node number to form a health index matrix, with rows corresponding to node numbers and columns corresponding to health index dimensions. Normalization is performed on the node risk score matrix and health index matrix to generate the final node risk score matrix and health index matrix; An improved heterogeneous high-order graph attention network model is trained. An optimization objective is constructed based on the joint deviation between the node embedding vector matrix and the node risk prediction output. The parameters in the high-order attention computation unit, the prediction-driven attention feedback unit, and the interpretable attention decomposition unit are continuously updated iteratively. During the training process, the aggregation weights of neighbor and high-order structural features are dynamically adjusted according to the mean and variance of node embeddings after each iteration. Training stops when the average change of the optimization objective in five consecutive iterations is less than 1% of the variance of each embedding feature.

[0023] In this embodiment, generating the initial strategy combination includes: The equipment risk matrix and the set of operation and maintenance resource constraints are input into a multi-objective hyperheuristic optimization algorithm to initialize the policy set. A unique code is generated for each policy, forming a policy code set. Finally, a policy embedding vector matrix is ​​generated, where: Initialize the policy set, specifically as follows: A list of low-level strategies is generated based on the equipment risk matrix and the set of operation and maintenance resource constraints. Each strategy includes a strategy number, strategy type, node coverage range, and initial parameters. Each strategy is assigned a unique code, forming a set of strategy codes; The generated strategy embedding vector matrix is ​​as follows: Each strategy's feature value in the multi-objective indicator space is used to construct a one-dimensional strategy vector, with the vector length equal to the indicator dimension. All one-dimensional strategy vectors are arranged in order of strategy number to form a two-dimensional embedding vector matrix, with rows corresponding to strategy numbers and columns corresponding to each indicator dimension. Each element records the strategy's value in the corresponding metric dimension, including node risk score, resource consumption value, and device priority; Standardize indicators with different dimensions to form a unified strategy embedding vector matrix; Based on the policy encoding set and policy embedding vector matrix, the local density and Pareto potential value of each policy in the multi-objective index space are calculated to form a policy cooperation metric matrix. Rows correspond to policy numbers, and columns correspond to inter-policy cooperation scores. A neighborhood policy table is generated by recording the policy numbers associated with the current policy in the multi-objective space. Calculate the local density of each strategy in the multi-objective metric space, specifically as follows: For each policy embedding vector, count the number of neighboring policies in the multi-objective index space to form a local density vector. Each element of the local density vector corresponds to the local density value of a policy. The Pareto potential energy is calculated as follows: For each strategy, perform non-dominated ranking in the multi-objective indicator space to determine the strategy level; The non-dominant superiority relationship between each policy and its neighboring policies is statistically analyzed to form the potential vector of each policy; Combine and summarize the potential vectors of all strategies to obtain the Pareto potential value for each strategy; The low-level policy set is partitioned based on the policy collaboration metric matrix and the neighborhood policy table, forming a policy subgroup structure matrix. Rows correspond to subgroup numbers, columns to policy numbers, and each element represents the correlation degree of a policy within a subgroup. During the partitioning process, the internal ranking of subgroups is determined by combining the policy embedding vector and Pareto potential value. The strategy subgroup structure matrix is ​​formed as follows: Based on the policy collaboration metric matrix and the neighborhood policy table, the low-level policy set is divided into several subgroups; Extract the corresponding policy embedding vectors and Pareto potential values ​​for policies within the subgroup, and sort them according to the feature distance of the embedding vectors and the magnitude of the potential values ​​to determine the order of the policies in the subgroup; Fill the strategy IDs in each subgroup into a matrix, with rows corresponding to subgroup IDs and columns corresponding to strategy IDs. Each element records the correlation and ranking information of the strategy in the subgroup, forming a strategy subgroup structure matrix. The policy subgroup structure matrix is ​​combined with the policy collaboration metric matrix to generate a candidate policy sequence according to the subgroup order and collaboration score, where: Candidate policy sequences are generated based on subgroup order and collaborative scoring, specifically as follows: Traverse the strategy subgroup structure matrix and process each subgroup row by row. For strategies within a subgroup, extract the collaboration score from the corresponding strategy collaboration metric matrix, and use the collaboration score as the sorting criterion to arrange the strategies in descending order. Arrange the sorted strategy numbers in order of subgroups in the matrix to form a candidate strategy sequence; Candidate policy sequences are mapped to form a preliminary policy combination matrix, with rows corresponding to policy combination numbers and columns corresponding to combination feature dimensions. Each element represents the index value of the policy in the combination, ultimately yielding the preliminary policy combination.

[0024] In this embodiment, generating the Pareto optimal operation and maintenance strategy set includes: The initial strategy combination, equipment risk matrix, and set of operation and maintenance resource constraints are input into a multi-objective hyperheuristic optimization algorithm to initialize the strategy iteration population. A unique code is generated for each strategy, forming a strategy code set. A strategy embedding vector matrix is ​​constructed, with rows corresponding to strategy combination numbers and columns corresponding to multi-objective indicator dimensions, where: Initialize the strategy iterative population as follows: The initial strategy combinations are loaded one by one into the multi-objective hyperheuristic optimization algorithm to form an initial set of individual strategies; Each strategy is assigned a unique code, the code set corresponds to the strategy combination number, and the strategy attributes and initial iteration state are recorded. The initial iterative population size is set to 180 policy combinations to form a complete policy iterative population; Construct the policy embedding vector matrix as follows: Based on the initial policy iterative population, the numerical features of each policy combination in the multi-objective index space are extracted to form a policy feature vector. Arrange all policy feature vectors in the order of policy encoding to generate a policy embedding vector matrix. The rows correspond to the policy combination number, the columns correspond to the multi-objective index dimension, and each element represents the numerical feature of the policy on the corresponding target index. Based on the policy encoding set and policy embedding vector matrix, the policy fitness value is calculated to form a policy fitness matrix. Rows correspond to policy combination numbers, and columns correspond to multi-objective indicators, including node risk scores and resource consumption values. A policy mutual information matrix is ​​then generated, with elements recording the correlation between policies in the multi-objective indicator space. The fitness value of the strategy is calculated as follows: Based on the policy encoding set and the policy embedding vector matrix, the feature values ​​of each policy combination in each target indicator dimension are summarized and calculated to form a policy fitness vector. Arrange all policy fitness vectors in the order of policy encoding to generate a policy fitness matrix. Rows correspond to policy combination numbers, columns correspond to multi-objective indicators, and each element represents the fitness value of the policy on the corresponding objective indicator. The strategy mutual information matrix is ​​generated as follows: Based on the strategy fitness matrix, the correlation of each strategy combination in the multi-objective index space is calculated to form the strategy mutual information value. The strategy mutual information value is arranged according to the strategy combination number to generate a strategy mutual information matrix, with rows and columns corresponding to the strategy combination number, and each element recording the multi-objective index correlation value between the corresponding strategy combinations. Based on the policy fitness matrix, non-dominated ranking and crowding calculation are performed to stratify the policies, generate a candidate Pareto front policy set, and iteratively update the subgroup structure of the candidate policies using the policy mutual information matrix. The non-dominated ranking and crowding calculation are performed based on the policy fitness matrix, specifically as follows: Traverse each strategy combination in the strategy fitness matrix, compare the objective values ​​of each strategy combination in the multi-objective index space with other strategy combinations, count the number of times it is dominated by other strategies, and determine the non-dominated level; among strategy combinations with the same non-dominated level, calculate the local density according to the numerical interval of each objective index to obtain the crowding value. Generate a set of candidate Pareto frontier policies, specifically as follows: Select the strategy combination with the lowest non-dominated level and arrange them in descending order of crowding value to form a candidate Pareto front strategy list. Each strategy combination in the candidate Pareto front strategy list contains a feature vector of the multi-objective indicator space, with the row corresponding to the strategy combination number and the column corresponding to the multi-objective indicator dimension. The subgroup structure of candidate strategies is iteratively updated as follows: Read the candidate Pareto frontier policy portfolios, and calculate the mutual information score between each policy portfolio and other policy portfolios within the subgroup, using the policy mutual information matrix. Calculate the mutual information score between each strategy combination and other strategy combinations within the subgroup, specifically as follows: For each subgroup of candidate Pareto front strategy combinations, extract the feature vector in the multi-objective index space; Extract feature vectors sequentially from other strategy combinations within the same subgroup; Statistically analyze the value distribution of the current strategy portfolio and other strategy portfolios within the subgroup across various target indicator dimensions; Based on the value distribution, calculate the mutual information score between the current strategy combination and each other strategy combination; The combined mutual information score of the current strategy combination is obtained by summing the mutual information scores of the current strategy combination with all other strategy combinations within the subgroup. All strategy combinations within a subgroup undergo the mutual information calculation process sequentially to form a complete set of mutual information scores; The strategy combinations within a subgroup are sorted from high to low according to their mutual information scores to form an updated subgroup structure. Crossover, mutation, and heuristic recombination operations are performed on the candidate Pareto frontier policy set to update the policy combination feature values. Combined with two-level closed-loop feedback, the state of the high-level policy selector is updated to generate a new set of policy combinations, where: The state of the high-level policy selector is updated by combining two-level closed-loop feedback, specifically as follows: The first-level closed loop reads the multi-objective indicator values ​​of each strategy combination in the current candidate Pareto frontier strategy set, including node risk score and resource consumption value; Statistical analysis is performed on the index values ​​of other strategy combinations within the same subgroup, the differences in each objective dimension are calculated, and the results are summarized to form a local deviation vector. By combining the strategy mutual information matrix, the local bias vector is mapped to the adaptive change score; The second-level closed loop iteratively updates the policy selection parameters and subgroup structure adjustment coefficients in the high-level policy selector based on the adaptive change score output by the first-level closed loop. The iterative process of repeating fitness calculation, Pareto front screening and strategy combination update continues until the average change of each strategy in the multi-objective indicators of the candidate Pareto front strategy set is less than 1% in three consecutive iterations, and finally the Pareto optimal operation and maintenance strategy set is generated. The generated Pareto optimal operation and maintenance strategy set is normalized to form the final Pareto optimal operation and maintenance strategy set.

[0025] In this embodiment, the collection and execution feedback data includes: The inspection and maintenance operation information executed by the distribution network operation and maintenance system is recorded as an operation log, including strategy number, execution device number, operation start time and operation end time; Collect operational status data of each device during inspection and maintenance operations, including device sensor readings, load data, environmental information and fault event records, and generate a sequence of device operational statuses by sorting them by device number and timestamp; Node risk feedback data is generated based on operation logs and device operation status sequences. The strategy number corresponding to the executed operation is associated with the risk score and health index of each device to form a node risk feedback matrix, with rows corresponding to device numbers and columns corresponding to strategy numbers and risk indicators. The node risk feedback matrix is ​​integrated with operation logs and equipment operating status sequences to generate execution feedback data.

[0026] In this embodiment, the updated node risk scoring matrix and the Pareto optimal operation and maintenance strategy set include: The execution feedback data and the latest equipment operating status data are input into the improved heterogeneous high-order graph attention network model and multi-objective hyperheuristic optimization algorithm to update the node embedding vector and interpretable factor matrix; A new node risk score matrix and health index matrix are generated based on the updated node embedding vector and interpretable factor matrix; The updated node risk score matrix and the latest equipment operating status data are input into a multi-objective hyperheuristic optimization algorithm to iteratively optimize the strategy combination, update the Pareto optimal operation and maintenance strategy set, and repeat the iteration until convergence to achieve closed-loop adaptive operation and maintenance.

[0027] Example 1: To verify the feasibility of this invention in practice, it was applied to a large-scale power distribution network operation and maintenance scenario. The system received operating status data, historical fault records, load data, and environmental information from 1800 devices. The raw data included approximately 5.4 million records, with an average missing rate of 4.7% for voltage, current, load, and environmental features, and approximately 620 historical fault event records. The system cleaned and normalized the data, scaling each feature value to the 0–1 range, while filling in missing data to form a multi-channel feature tensor with dimensions of 1800×100×6. Each device corresponded to 100 time steps, and each step contained 6 feature channels. The set of operation and maintenance resource constraints included 72 inspection personnel, 120 sets of operation and maintenance tools, and 48 schedulable time periods.

[0028] After data processing, the system constructs a heterogeneous high-order graph of the distribution network, mapping 1800 devices to nodes. Node numbers, types, and multi-channel features are written into the node matrix, generating a node feature matrix of 1800×128. The system identifies approximately 4200 high-order structures, including three-node triangles and four-node rings, and generates a high-order structure feature matrix of 1800×128×3. Node, edge, and high-order structure features are integrated to form a graph structure tensor, which is used as input to the improved heterogeneous high-order graph attention network model. In the high-order attention calculation unit, heterogeneous semantic relationship attention calculation is performed on each node, its neighboring nodes, and its associated high-order structure, generating a node embedding vector matrix of 1800×128. The prediction-driven attention feedback unit uses the node embedding vectors to generate node risk prediction output and feeds back to update the embeddings. The interpretable attention decomposition unit decomposes the embedding matrix according to the feature dimensions to generate an interpretable factor matrix of 1800×16. After calculating the node risk score matrix and health index matrix, the risk prediction error decreases from the initial 0.19 to 0.04, and the health index coverage reaches 100%.

[0029] The equipment risk matrix and the set of O&M resource constraints are input into a multi-objective hyperheuristic optimization algorithm. An initial set of 180 strategies is generated, a strategy encoding set is produced, and a strategy embedding vector matrix is ​​constructed. Based on the strategy embedding vectors, the system calculates the local density and Pareto potential value of each strategy in the multi-objective indicator space, forming a strategy collaboration metric matrix. The strategy subgroup structure is divided according to the strategy collaboration metric matrix and embedding vectors, and candidate strategy sequences are generated by combining the subgroup order and collaboration score, mapping to form preliminary strategy combinations. During iterative optimization, the system performs non-dominated sorting, congestion calculation, and strategy crossover and mutation operations, dynamically adjusting the strategy subgroup structure. After three consecutive iterations, the Pareto optimal O&M strategy set is generated, including 36 strategy combinations. The average node risk score of the strategy combinations in the multi-objective indicators is 0.067, and the average resource utilization rate is 88%. The final strategy is sent to the operation and maintenance system for inspection. After collecting feedback data, the node risk score matrix is ​​updated. After closed-loop iteration, the average fault response time is reduced from 12 hours to 3.3 hours, and the strategy repetition rate is reduced from 21% to 5%, which verifies the feasibility and efficiency of the present invention in closed-loop adaptive operation and maintenance based on strategy collaboration and Pareto potential.

[0030] Table 1. Comparison of key indicators of the power distribution operation and maintenance strategy optimization method between the method of the present invention and the traditional method.

[0031] As can be seen from the statistical results in Table 1, this invention achieves significant improvements over traditional methods in optimizing power distribution operation and maintenance strategies. Regarding node risk prediction, the prediction error decreased from 0.19 in the traditional method to 0.04, and the health index coverage increased from 82% to 100%, demonstrating that this invention can more accurately assess equipment status and comprehensively cover critical nodes. In strategy optimization, the number of Pareto optimal strategy combinations increased from 24 in the traditional method to 36, the average node risk score decreased from 0.12 to 0.067, and the strategy duplication rate decreased from 21% to 5%, reflecting enhanced diversity and reduced redundancy in strategy combinations, effectively improving the problems of single strategy and uneven resource allocation in traditional methods.

[0032] In terms of operational efficiency, this invention significantly improves overall response efficiency and critical node coverage. The average fault response time is reduced from 12 hours using traditional methods to 3.3 hours, and critical node coverage is increased from 79% to 98%, indicating that this invention is more efficient in critical node risk identification and inspection task allocation. By calculating the policy collaboration metric matrix and Pareto potential value using policy embedding vectors, this invention can dynamically adjust the policy subgroup structure and combination order during iteration, achieving closed-loop adaptive optimization in a multi-objective space and ensuring that the policy combination remains optimal and stable.

[0033] Comprehensive analysis shows that this invention not only outperforms traditional methods in terms of node risk prediction accuracy, strategy diversity, and resource utilization, but also significantly shortens fault response time and improves critical node coverage and closed-loop adaptive capability. These data fully verify the engineering feasibility and technical advantages of this invention in implementing closed-loop adaptive operation and maintenance in large-scale distribution networks, effectively solving problems such as single strategy, delayed critical node risk identification, and resource waste in traditional operation and maintenance methods.

[0034] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for optimizing power distribution operation and maintenance strategies based on big data analysis, characterized in that, include: Collect multi-source operation data of distribution network equipment, perform preprocessing to generate multi-channel feature tensors, obtain resource configuration data of distribution network operation and maintenance system, and form a set of operation and maintenance resource constraints. A heterogeneous high-order graph of the distribution network is constructed based on the multi-channel feature tensor, where nodes represent devices and edges and high-order structures represent the electrical and functional relationships between devices. The heterogeneous high-order graph of the distribution network is input into the improved heterogeneous high-order graph attention network model. Heterogeneous semantic relationship attention calculation is performed to generate node embedding representations, calculate node risk prediction outputs, adjust attention allocation, decompose node embedding representations into interpretable factors, and generate node risk score matrices and health index matrices. The device risk matrix is ​​constructed based on the node risk scoring matrix and the health index matrix. The device risk matrix and the set of operation and maintenance resource constraints are input into the multi-objective hyperheuristic optimization algorithm. The high-level strategy selector combines the strategy collaboration metric matrix and Pareto potential energy field information to generate a preliminary strategy combination. The initial strategy combination is iteratively optimized using a multi-objective hyperheuristic optimization algorithm. The strategy subgroup structure is dynamically adjusted based on the strategy mutual information matrix. The high-level strategy selector is updated by combining two-level closed-loop feedback to generate the Pareto optimal operation and maintenance strategy set. Select the final operation and maintenance solution from the Pareto optimal operation and maintenance strategy set, send it to the distribution network operation and maintenance system to perform inspection and maintenance operations, and collect execution feedback data; The execution feedback data and the latest equipment operating status data are input into the improved heterogeneous high-order graph attention network model and multi-objective hyperheuristic optimization algorithm to update the node risk score matrix and the Pareto optimal operation and maintenance strategy set.

2. The power distribution operation and maintenance strategy optimization method based on big data analysis according to claim 1, characterized in that, The generation of the multi-channel feature tensor and the set of operation and maintenance resource constraints includes: The multi-source operation data of the distribution network equipment are sorted according to the equipment number and the collection time, including operation status data, historical fault records, load data and environmental information, to generate a data sequence arranged by equipment number and time. Perform missing value imputation on the sorted data sequence, and use the historical average value to fill in the missing data items; Normalization is performed on the filled data sequence to convert the value of each feature, including voltage, current, load, and environment, to a uniform range, forming a standardized feature sequence. The normalized feature sequences are integrated according to device number, time order and data type to form a multi-channel feature tensor, where each channel corresponds to a data type; Obtain resource configuration data for the power distribution network operation and maintenance system, including the number of inspection personnel, the number of operation and maintenance equipment, and the available time periods, and organize them into a set of operation and maintenance resource constraints.

3. The method for optimizing power distribution operation and maintenance strategies based on big data analysis according to claim 1, characterized in that, The construction of the heterogeneous high-order graph of the distribution network includes: Based on multi-channel feature tensors, distribution network equipment is converted into nodes. Equipment number, equipment type, equipment operating status data, historical fault records, load data and environmental information are written into the corresponding nodes to generate a node set and establish a unique identifier for each node. Based on the node set, edges are created, and the electrical parameters, line impedance, and load capacity between nodes are written into the corresponding edges to generate an edge set. The starting node number and ending node number of each edge are recorded. Generate an edge feature matrix based on the edge set, with rows corresponding to edge numbers and columns corresponding to edge attributes, including electrical parameters, line impedance, load capacity, and historical fault data; Identify higher-order structures in the edge set, including three-node triangles, four-node rings and bridging nodes, record the higher-order structure number and the set of constituent nodes, and generate a set of higher-order structures. The set of nodes, the set of edges, and the set of higher-order structures are encoded into node feature matrices, edge feature matrices, and higher-order structure feature matrices, respectively. These matrices are then mapped to different channels according to the categories of nodes, edges, and higher-order structures to generate graph structure tensors. The node feature matrix, edge feature matrix, and higher-order structure feature matrix are integrated to generate a heterogeneous higher-order graph of the distribution network, including a set of nodes, a set of edges, a set of higher-order structures, and the graph structure tensor.

4. The method for optimizing power distribution operation and maintenance strategies based on big data analysis according to claim 1, characterized in that, The generated node risk score and health index matrix includes: The heterogeneous high-order graph of the distribution network is input into the improved heterogeneous high-order graph attention network model, which includes a high-order attention calculation unit, a prediction-driven attention feedback unit, and an interpretable attention decomposition unit. The higher-order attention computation unit performs heterogeneous semantic relationship attention computation on each node, its neighboring nodes and its higher-order structure based on the node set, edge set and higher-order structure set. It generates attention weights and performs weighted aggregation to form a node embedding vector matrix, with rows corresponding to node numbers and columns corresponding to embedding feature dimensions. The prediction-driven attention feedback unit calculates the node risk prediction output based on the node embedding vector matrix, feeds the node risk prediction output back to the higher-order attention calculation unit, updates the node embedding vector matrix, and adjusts the aggregation ratio of node features in neighbors and higher-order structures in each iteration. The interpretable attention decomposition unit decomposes the updated node embedding vector matrix according to the feature dimension to generate an interpretable factor matrix. The component information is recorded according to the higher-order structure type, with rows corresponding to node numbers and columns corresponding to factor numbers. Each element represents the component of the node embedded in that factor. The node risk score matrix is ​​calculated based on the interpretable factor matrix. The rows correspond to the node numbers and the columns correspond to the risk dimensions. Each element is generated by the value of the node embedding vector on the corresponding factor. The health index matrix is ​​calculated based on the node embedding vector matrix and the node risk score matrix. The rows correspond to the node numbers and the columns correspond to the health index dimensions. Each element is generated by combining the node embedding vector and the risk score. Normalization is performed on the node risk score matrix and health index matrix to generate the final node risk score matrix and health index matrix; An improved heterogeneous high-order graph attention network model is trained. An optimization objective is constructed based on the joint deviation between the node embedding vector matrix and the node risk prediction output. The parameters in the high-order attention computation unit, the prediction-driven attention feedback unit, and the interpretable attention decomposition unit are continuously updated iteratively. During the training process, the aggregation weights of neighbor and high-order structural features are dynamically adjusted according to the mean and variance of node embeddings after each iteration. Training stops when the average change of the optimization objective in five consecutive iterations is less than 1% of the variance of each embedding feature.

5. The method for optimizing power distribution operation and maintenance strategies based on big data analysis according to claim 1, characterized in that, The generation of the initial strategy combination includes: The equipment risk matrix and the set of operation and maintenance resource constraints are input into a multi-objective hyperheuristic optimization algorithm to initialize the strategy set, generate a unique code for each strategy, form a strategy code set, and generate a strategy embedding vector matrix. Based on the policy encoding set and policy embedding vector matrix, the local density and Pareto potential value of each policy in the multi-objective index space are calculated to form a policy collaboration metric matrix. The rows correspond to the policy numbers, the columns correspond to the collaborative scores between policies, and the policy numbers that are related to the current policy in the multi-objective space are recorded to generate a neighborhood policy table. The low-level policy set is divided according to the policy collaboration metric matrix and the neighborhood policy table to form a policy subgroup structure matrix. The rows correspond to the subgroup numbers and the columns correspond to the policy numbers. Each element represents the correlation degree of the policy in the subgroup. During the division process, the internal order of the subgroup is determined by combining the policy embedding vector and the Pareto potential value. The strategy subgroup structure matrix is ​​combined with the strategy synergy metric matrix to generate a candidate strategy sequence according to the subgroup order and synergy score. Candidate policy sequences are mapped to form a preliminary policy combination matrix, with rows corresponding to policy combination numbers and columns corresponding to combination feature dimensions. Each element represents the index value of the policy in the combination, ultimately yielding the preliminary policy combination.

6. The method for optimizing power distribution operation and maintenance strategies based on big data analysis according to claim 1, characterized in that, The set of optimal Pareto operation and maintenance strategies includes: The initial strategy combination, equipment risk matrix, and set of operation and maintenance resource constraints are input into a multi-objective hyperheuristic optimization algorithm to initialize the strategy iteration population, generate a unique code for each strategy, form a strategy code set, and construct a strategy embedding vector matrix, with rows corresponding to the strategy combination number and columns corresponding to the multi-objective indicator dimensions. The policy fitness value is calculated based on the policy encoding set and the policy embedding vector matrix to form a policy fitness matrix. The rows correspond to the policy combination number, and the columns correspond to multi-objective indicators, including node risk scores and resource consumption values. A policy mutual information matrix is ​​generated, and the elements record the correlation between each policy in the multi-objective indicator space. Based on the policy fitness matrix, non-dominated sorting and crowding calculation are performed to stratify the policies, filter and generate a set of candidate Pareto front policies, and iteratively update the subgroup structure of the candidate policies by combining the policy mutual information matrix. Perform crossover, mutation, and heuristic recombination operations on the candidate Pareto frontier policy set to update the policy combination feature values. Combine the two-level closed-loop feedback to update the state of the high-level policy selector and generate a new round of policy combination set. The iterative process of repeating fitness calculation, Pareto front screening and strategy combination update continues until the average change of each strategy in the multi-objective indicators of the candidate Pareto front strategy set is less than 1% in three consecutive iterations, and finally the Pareto optimal operation and maintenance strategy set is generated. The generated Pareto optimal operation and maintenance strategy set is normalized to form the final Pareto optimal operation and maintenance strategy set.

7. The method for optimizing power distribution operation and maintenance strategies based on big data analysis according to claim 1, characterized in that, The collected execution feedback data includes: The inspection and maintenance operation information performed by the distribution network operation and maintenance system is recorded as an operation log. The log includes the strategy number, the execution device number, the operation start time, and the operation end time. Collect operational status data of each device during inspection and maintenance operations, including device sensor readings, load data, environmental information and fault event records, and generate an operational status sequence by sorting by device number and timestamp; Node risk feedback data is generated based on the sequence of executed operations and running status. The strategy number corresponding to the executed operation is associated with the risk score and health index of each device to form a node risk feedback matrix, with rows corresponding to device numbers and columns corresponding to strategy numbers and risk indicators. The node risk feedback matrix is ​​integrated with operation logs and equipment operating status sequences to generate execution feedback data.

8. The method for optimizing power distribution operation and maintenance strategies based on big data analysis according to claim 1, characterized in that, The updated node risk scoring matrix and Pareto optimal operation and maintenance strategy set include: The execution feedback data and the latest equipment operating status data are input into the improved heterogeneous high-order graph attention network model and multi-objective hyperheuristic optimization algorithm to update the node embedding vector and interpretable factor matrix; A new node risk score matrix and health index matrix are generated based on the updated node embedding vector and interpretable factor matrix; The updated node risk score matrix and the latest equipment operating status data are input into a multi-objective hyperheuristic optimization algorithm to iteratively optimize the strategy combination, update the node risk score matrix and the Pareto optimal operation and maintenance strategy set, and repeat the iteration until convergence to achieve closed-loop adaptive operation and maintenance.

Images