Power distribution network automation operation and maintenance method and system based on artificial intelligence

Abstract

The present invention provides an artificial intelligence-based automated operation and maintenance method and system for an electric power distribution network, which relates to the technical field of electric power operation and maintenance, including real-time collection of operating data of the electric power distribution network through a distributed sensor network, transmitting the collected operating data to a cloud data center, designing a multi-agent reinforcement learning framework, dividing the distribution network into multiple sub-areas, and an intelligent operation and maintenance agent formulates a personalized fault handling plan by real-time monitoring and analysis of a digital twin model, and formulates an optimal execution plan based on the fault handling plan in combination with real-time road condition information and the location of maintenance personnel; the optimal execution plan is respectively sent to the automated equipment and the mobile terminals of the maintenance personnel through a secure encrypted channel, and preventive maintenance rules are extracted from historical operation and maintenance records, and a preventive maintenance plan is formulated based on the extracted preventive maintenance rules in combination with a pre-trained equipment life prediction model.

Description

Technical Field

[0001] The present invention relates to the field of electric power operation and maintenance technology, and in particular to an artificial intelligence-based automatic operation and maintenance method and system for an electric power distribution network. Background Art

[0002] With the rapid development of society and the continuous growth of economy, the demand for electricity is increasing day by day. As an important part of the power system, the operation efficiency and safety of the power distribution network are increasingly valued. The traditional operation and maintenance methods of the power distribution network mainly rely on manual experience and regular inspections, which have defects such as slow response speed, low fault diagnosis efficiency, and unscientific maintenance plan formulation. It is difficult to meet the requirements of modern power systems for real-time monitoring and automated management. Therefore, there is an urgent need for an automated operation and maintenance method and system for power distribution networks based on advanced technology to improve the operation efficiency and fault handling capabilities of power distribution networks. Summary of the invention

[0003] The embodiments of the present invention provide an artificial intelligence-based automatic operation and maintenance method and system for an electric power distribution network, which can solve the problems in the prior art.

[0004] According to a first aspect of the embodiments of the present invention,

[0005] Provided is an artificial intelligence-based automated operation and maintenance method for a power distribution network, comprising:

[0006] The operation data of the power distribution network is collected in real time through a distributed sensor network, and the collected operation data is transmitted to the cloud data center. The historical operation data and the real-time operation data are fused and analyzed using a multimodal deep learning framework. Based on the fusion analysis results, a hybrid power distribution network digital twin model containing a physical model and a data-driven model is constructed. Based on the constructed hybrid power distribution network digital twin model, a multi-agent reinforcement learning framework is designed to divide the distribution network into multiple sub-areas, each of which corresponds to an intelligent operation and maintenance sub-agent. A hierarchical reinforcement learning algorithm is used to train the intelligent operation and maintenance agents, in which the top-level agent is responsible for global strategy formulation and the bottom-level sub-agent is responsible for local execution.

[0007] The intelligent operation and maintenance agent monitors and analyzes the digital twin model in real time, and uses the attention mechanism and causal reasoning algorithm to identify potential fault risks and abnormal conditions. When an abnormality is detected, the intelligent operation and maintenance agent activates the multi-level fault diagnosis system, integrates the multi-level diagnosis results, and automatically generates a fault diagnosis report. According to the fault diagnosis report, the preset processing solution library is called, and combined with the current system status, a personalized fault handling solution is formulated. Based on the fault handling solution, combined with real-time road condition information and the location of maintenance personnel, the optimal execution plan is formulated;

[0008] The optimal execution plan is sent to the automation equipment and maintenance personnel's mobile terminals through a secure encrypted channel. During the execution process, the intelligent operation and maintenance agent collects equipment response data and maintenance personnel's feedback information in real time, and compares and analyzes them with the optimal execution plan. When an execution deviation is detected, the intelligent operation and maintenance agent dynamically adjusts the execution plan. After completing the fault handling, the intelligent operation and maintenance agent generates an operation and maintenance report based on the whole process data, stores the generated operation and maintenance report in the database, and extracts preventive maintenance rules from historical operation and maintenance records. Based on the extracted preventive maintenance rules and combined with the pre-trained equipment life prediction model, a preventive maintenance plan is formulated.

[0009] In an optional embodiment,

[0010] Based on the constructed digital twin model of hybrid power distribution network, a multi-agent reinforcement learning framework is designed to divide the distribution network into multiple sub-areas. Each sub-area corresponds to an intelligent operation and maintenance sub-agent. A hierarchical reinforcement learning algorithm is used to train the intelligent operation and maintenance agents. The top-level agent is responsible for global strategy formulation, and the bottom-level sub-agent is responsible for local execution, including:

[0011] The topological structure of the distribution network is represented as a time-varying graph, wherein the time-varying graph includes a fixed node set, a time-varying edge set, and a time-varying node feature matrix, wherein the fixed node set represents the distribution equipment, the time-varying edge set represents the power line, and the time-varying node feature matrix includes the equipment type, rated capacity, and operation status information;

[0012] Construct a multi-layer graph attention network, each layer of which includes multiple independent attention heads, weighted aggregation of neighbor node features, and concatenation of the outputs of multiple attention heads to obtain an embedded representation of the node, construct a similarity matrix based on the embedded representation of the node, use a Gaussian kernel function to calculate the similarity between nodes, calculate a standardized Laplacian matrix based on the similarity matrix, and obtain the eigenvector of the standardized Laplacian matrix, normalize the eigenvector to obtain a normalized eigenvector matrix, use a clustering algorithm to cluster each row of the normalized eigenvector matrix, and obtain a preliminary sub-region division result;

[0013] Based on the preliminary sub-region division results, an evaluation index based on electrical distance and load balance is introduced. The evaluation index considers the average electrical distance inside the sub-region and the total load of the sub-region. The optimal sub-region division result is determined by adjusting the number of clusters and minimizing the evaluation index.

[0014] An intelligent operation and maintenance subagent is allocated to each sub-region. The intelligent operation and maintenance subagent adopts a hierarchical reinforcement learning algorithm, including a top-level agent and a bottom-level subagent. The top-level agent is responsible for global strategy formulation. Its state space contains the global information of the entire distribution network, and its action space contains the resource allocation strategy and global optimization objectives between sub-regions. The reward function comprehensively considers the power supply reliability, economy and environmental protection of the entire distribution network.

[0015] The bottom-level subagent is responsible for local execution. Its state space includes the real-time operating parameters, historical fault records and maintenance plans of all equipment in the sub-area. Its action space includes equipment control instructions, fault diagnosis decisions and maintenance personnel scheduling strategies. The reward function considers the power supply reliability index, equipment health index and operation and maintenance cost in the sub-area.

[0016] Build a global coordinator to collect information from each intelligent operation and maintenance sub-agent, combine the global strategy of the top-level agent, solve the global optimal decision, and decompose it into sub-decisions and send them to each bottom-level sub-agent for execution.

[0017] In an optional embodiment,

[0018] Based on the preliminary sub-region division results, an evaluation index based on electrical distance and load balance is introduced. The evaluation index considers the average electrical distance within the sub-region and the total load of the sub-region. By adjusting the number of clusters and minimizing the evaluation index, the optimal sub-region division results are determined, including:

[0019] Based on the preliminary sub-region division results, a weighted adjacency graph between sub-regions is constructed, where the nodes of the graph represent sub-regions, and the weights of the edges comprehensively consider the electrical distance, power exchange, and topological connection strength between sub-regions.

[0020] Calculating the normalized Laplace matrix of the weighted adjacency graph and solving its eigenvalues ​​and eigenvectors, constructing a low-dimensional embedding space based on the eigenvectors, applying an improved spectral clustering algorithm in the low-dimensional embedding space, dynamically adjusting cluster centers and the number of clusters through an iterative optimization process, and obtaining optimized sub-region divisions;

[0021] Identify the boundary nodes between sub-areas, construct a boundary node impact assessment model, calculate the impact of the boundary nodes on the dynamic evaluation index function based on the boundary node impact assessment model, and adjust the attribution of the boundary nodes based on a greedy strategy to minimize the value of the dynamic evaluation index function;

[0022] The distribution network is divided into multiple overlapping local areas, each of which is assigned to a computing node. Clustering parameters and dynamic evaluation index weights are automatically adjusted on each computing node according to the characteristics of the local area. At the same time, a distributed consensus mechanism based on blockchain is designed to exchange boundary node information and local optimization results between computing nodes.

[0023] The federated learning method is used to integrate the local partitioning results of each computing node, construct a global sub-region partitioning scheme, and calculate the dynamic evaluation index function value of the global sub-region partitioning scheme. If it is better than the current optimal scheme, the optimal partitioning result is updated, and the iteration is repeated until the preset number of iterations is reached. The global sub-region partitioning scheme with the smallest dynamic evaluation index function value is output as the final result.

[0024] In an optional embodiment,

[0025] According to the fault diagnosis report, call the preset processing solution library, and combine the current system status to develop a personalized fault processing solution including:

[0026] Use natural language processing technology to extract key information from historical fault diagnosis reports, maintenance records and expert experience, and build a knowledge graph containing entities such as equipment, fault type and processing method;

[0027] Integrate new fault diagnosis reports and processing experience into the knowledge graph through an incremental learning framework, dynamically update the knowledge graph, and map the received fault diagnosis reports to entities and relationships in the updated knowledge graph;

[0028] Use the relational graph convolutional network to perform representation learning on the mapped knowledge graph to obtain the vector representation of the graph entity. The multi-hop reasoning algorithm based on the attention mechanism performs path search on the knowledge graph represented by the vector to obtain the initial set of fault handling solutions.

[0029] Construct a sequence generation model based on the transformer architecture, including a multi-layer self-attention encoder and decoder, where the encoder encodes the fault diagnosis report and system status into a context vector sequence;

[0030] A pointer generation network with a copy mechanism is used in the decoder. The attention mechanism is used to dynamically select whether to copy information from the input or generate new words from the vocabulary. The encoded context vector sequence and the initial fault handling solution set are used as the input of the decoder. The sequence generation model is trained to maximize the likelihood probability of generating personalized fault handling solutions.

[0031] Use the trained sequence generation model to generate personalized fault handling solutions based on the current fault diagnosis report, system status, and potential handling solution set, including handling steps, required resources, and expected results.

[0032] In an optional embodiment,

[0033] Generate an operation and maintenance report based on the whole process data, store the generated operation and maintenance report in the database, and extract preventive maintenance rules from historical operation and maintenance records. Based on the extracted preventive maintenance rules and combined with the pre-trained equipment life prediction model, formulate a preventive maintenance plan including:

[0034] Collect equipment status data when a fault occurs, reasoning path during fault diagnosis, specific treatment measures taken, treatment results and effect evaluation to form full process data;

[0035] Process the whole process data using a generated pre-trained model based on a converter architecture to generate an operation and maintenance report, and store the generated operation and maintenance report in a document-based database to form a historical operation and maintenance record;

[0036] Using a named entity recognition model based on a bidirectional encoder representation to identify equipment names, fault types, and maintenance operations from the historical operation and maintenance records, and extract target information;

[0037] Applying association rule mining algorithm to process the extracted target information, mining the association between equipment status, operation behavior and fault, obtaining association rule set, performing time series analysis on the association rule set, identifying the time series event pattern leading to the fault, and forming preliminary preventive maintenance rules;

[0038] Evaluating the effectiveness of the preliminary preventive maintenance rules, screening out preventive maintenance rules with a credibility higher than a preset threshold, and forming a final preventive maintenance rule set;

[0039] Collect and analyze historical equipment operation data and input it into the pre-trained equipment life prediction model to calculate the expected remaining service life of each key component of the equipment and generate a life prediction report;

[0040] A preliminary preventive maintenance plan is formulated according to the final preventive maintenance rule set and the life prediction report, and the preliminary preventive maintenance plan is optimized by multi-objectives considering maintenance cost, equipment reliability and production influencing factors to obtain a final preventive maintenance plan.

[0041] In an optional embodiment,

[0042] Collect and analyze historical equipment operation data and input it into the pre-trained equipment life prediction model to calculate the expected remaining service life of each key component of the equipment and generate a life prediction report including:

[0043] Collecting and preprocessing historical operation data of the equipment, wherein the historical operation data includes time series measurement values ​​of multiple sensors;

[0044] Perform multi-scale feature extraction on the preprocessed data, respectively extract time domain features, frequency domain features and time-frequency domain features to obtain multi-scale features, and construct a multi-scale temporal convolutional network model based on the multi-scale features. The model includes multiple parallel one-dimensional convolution branches with different convolution kernel sizes, a scaled dot product attention mechanism, a residual connection and a temporal gating unit;

[0045] A multi-task learning framework is adopted, an adaptive moment estimation optimizer and a cosine annealing strategy are used to adjust the learning rate, and a combined loss function is used to train the multi-scale temporal convolutional network model, and the extracted multi-scale features are input into the trained multi-scale temporal convolutional network model to calculate the expected remaining service life of each key component of the equipment and its confidence interval;

[0046] Based on the expected remaining service life, generate a life prediction trend chart, analyze factors affecting the life based on the expected remaining service life, identify abnormal patterns that may cause a sharp decline in life based on the expected remaining service life, and conduct potential risk assessment;

[0047] Based on the expected remaining service life, life prediction trend chart, key factor analysis and potential risk assessment results, maintenance recommendations are generated, and the expected remaining service life and confidence interval of each key component, life prediction trend chart, key factor analysis affecting life, potential risk assessment and maintenance recommendations are integrated to generate a life prediction report.

[0048] According to a second aspect of the embodiments of the present invention,

[0049] Provided is an artificial intelligence-based power distribution network automation operation and maintenance system, including:

[0050] The first unit is used to collect the operation data of the power distribution network in real time through a distributed sensor network, transmit the collected operation data to the cloud data center, use a multimodal deep learning framework to fuse and analyze the historical operation data and the real-time operation data, and build a hybrid power distribution network digital twin model containing a physical model and a data-driven model based on the fusion analysis results. Based on the constructed hybrid power distribution network digital twin model, a multi-agent reinforcement learning framework is designed to divide the distribution network into multiple sub-areas, each of which corresponds to an intelligent operation and maintenance sub-agent. A hierarchical reinforcement learning algorithm is used to train the intelligent operation and maintenance agent, in which the top-level agent is responsible for global strategy formulation and the bottom-level sub-agent is responsible for local execution;

[0051] The second unit is used for the intelligent operation and maintenance agent to monitor and analyze the digital twin model in real time, and use the attention mechanism and causal reasoning algorithm to identify potential fault risks and abnormal conditions. When an abnormality is detected, the intelligent operation and maintenance agent activates the multi-level fault diagnosis system, integrates the multi-level diagnosis results, and automatically generates a fault diagnosis report. According to the fault diagnosis report, the preset processing solution library is called, and combined with the current system status, a personalized fault handling solution is formulated, and based on the fault handling solution, combined with real-time road condition information and the location of maintenance personnel, an optimal execution plan is formulated;

[0052] The third unit is used to send the optimal execution plan to the automation equipment and the maintenance personnel's mobile terminals through a secure encrypted channel. During the execution process, the intelligent operation and maintenance agent collects equipment response data and maintenance personnel feedback information in real time, and compares and analyzes them with the optimal execution plan. When an execution deviation is detected, the intelligent operation and maintenance agent dynamically adjusts the execution plan. After completing the fault handling, the intelligent operation and maintenance agent generates an operation and maintenance report based on the whole process data, stores the generated operation and maintenance report in the database, and extracts preventive maintenance rules from historical operation and maintenance records. Based on the extracted preventive maintenance rules and combined with the pre-trained equipment life prediction model, a preventive maintenance plan is formulated.

[0053] According to a third aspect of the embodiments of the present invention,

[0054] An electronic device is provided, comprising:

[0055] processor;

[0056] a memory for storing processor-executable instructions;

[0057] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0058] A fourth aspect of the embodiments of the present invention is:

[0059] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0060] In this embodiment, real-time monitoring of the distribution network is achieved through a distributed sensor network and a cloud data center, which improves the timeliness of anomaly detection; multimodal deep learning and a multi-level fault diagnosis system are used to improve the accuracy of fault diagnosis; intelligent decision support is provided based on a digital twin model and a multi-agent reinforcement learning framework to achieve personalized fault handling and dynamic adjustment. Through historical data analysis and equipment life prediction, proactive preventive maintenance is achieved; maintenance resource allocation is optimized to improve operation and maintenance efficiency. This solution realizes the systematic accumulation and inheritance of operation and maintenance knowledge, comprehensively manages operational risks, improves the level of automation and system reliability, and reduces maintenance costs. The multi-agent framework realizes inter-regional collaborative operation and maintenance and improves overall efficiency. This method has good scalability and security, realizes intelligent management of the entire life cycle of distribution network equipment, improves energy utilization efficiency, provides technical support for the digital transformation of the power industry, and greatly improves the operation and maintenance level of the power distribution network. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 A schematic diagram of a flow chart of an automatic operation and maintenance method for a power distribution network based on artificial intelligence according to an embodiment of the present invention;

[0062] Figure 2 It is a structural schematic diagram of an artificial intelligence-based power distribution network automation operation and maintenance system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0063] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0064] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0065] Figure 1 FIG. 1 is a flow chart of an automatic operation and maintenance method of a power distribution network based on artificial intelligence according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0066] S101. Collect the operating data of the power distribution network in real time through a distributed sensor network, transmit the collected operating data to the cloud data center, use a multimodal deep learning framework to fuse and analyze the historical operating data and the real-time operating data, and build a hybrid power distribution network digital twin model that includes a physical model and a data-driven model based on the fusion analysis results. Based on the constructed hybrid power distribution network digital twin model, design a multi-agent reinforcement learning framework, divide the distribution network into multiple sub-areas, each sub-area corresponds to an intelligent operation and maintenance sub-agent, and use a hierarchical reinforcement learning algorithm to train the intelligent operation and maintenance agents, in which the top-level agent is responsible for global strategy formulation, and the bottom-level sub-agent is responsible for local execution.

[0067] First, a distributed sensor network is deployed at each key node of the distribution network, including voltage sensors, current sensors, temperature sensors, etc., to collect real-time operation data of the distribution network. These sensors transmit the collected data to the nearest data aggregation node through wireless communication technology, and then the data aggregation node transmits the data to the cloud data center through the optical fiber network.

[0068] After receiving the real-time operation data, the cloud data center combines the historical operation data and uses a multimodal deep learning framework to perform data fusion and analysis. Specifically, a convolutional neural network is used to process image data (such as infrared thermal imaging), a recurrent neural network is used to process time series data (such as voltage and current waveforms), and a fully connected neural network is used to process discrete data (such as switch status). Different modal features are fused through the attention mechanism to obtain a unified representation of the distribution network operation status.

[0069] Based on the fusion analysis results, a digital twin model of the hybrid power distribution network is constructed. The model consists of two parts: one is a physical model based on the physical laws of the power system, such as a power flow calculation model; the other is a data-driven model trained based on historical data, such as a transformer temperature rise prediction model. The two models complement each other and improve the accuracy and generalization ability of the digital twin model.

[0070] Next, a multi-agent reinforcement learning framework is designed based on the hybrid digital twin model. The distribution network is divided into multiple sub-areas, each of which corresponds to an intelligent operation and maintenance sub-agent. A hierarchical reinforcement learning algorithm is used to train the intelligent operation and maintenance agent, in which the top-level agent is responsible for formulating global strategies, such as load transfer plans; the bottom-level sub-agent is responsible for performing local tasks, such as equipment status monitoring. Specifically, the state space of the top-level agent includes global information such as the load distribution of the entire network and voltage margin, and the action space includes load transfer instructions, etc.; the state space of the bottom-level sub-agent includes local information such as equipment status and electrical quantity within the jurisdiction, and the action space includes equipment operation instructions, etc.

[0071] In an optional embodiment,

[0072] Based on the constructed digital twin model of hybrid power distribution network, a multi-agent reinforcement learning framework is designed to divide the distribution network into multiple sub-areas. Each sub-area corresponds to an intelligent operation and maintenance sub-agent. A hierarchical reinforcement learning algorithm is used to train the intelligent operation and maintenance agents. The top-level agent is responsible for global strategy formulation, and the bottom-level sub-agent is responsible for local execution, including:

[0073] The topological structure of the distribution network is represented as a time-varying graph, wherein the time-varying graph includes a fixed node set, a time-varying edge set, and a time-varying node feature matrix, wherein the fixed node set represents the distribution equipment, the time-varying edge set represents the power line, and the time-varying node feature matrix includes the equipment type, rated capacity, and operation status information;

[0074] Construct a multi-layer graph attention network, each layer of which includes multiple independent attention heads, weighted aggregation of neighbor node features, and concatenation of the outputs of multiple attention heads to obtain an embedded representation of the node, construct a similarity matrix based on the embedded representation of the node, use a Gaussian kernel function to calculate the similarity between nodes, calculate a standardized Laplacian matrix based on the similarity matrix, and obtain the eigenvector of the standardized Laplacian matrix, normalize the eigenvector to obtain a normalized eigenvector matrix, use a clustering algorithm to cluster each row of the normalized eigenvector matrix, and obtain a preliminary sub-region division result;

[0075] Based on the preliminary sub-region division results, an evaluation index based on electrical distance and load balance is introduced. The evaluation index considers the average electrical distance inside the sub-region and the total load of the sub-region. The optimal sub-region division result is determined by adjusting the number of clusters and minimizing the evaluation index.

[0076] An intelligent operation and maintenance subagent is allocated to each sub-region. The intelligent operation and maintenance subagent adopts a hierarchical reinforcement learning algorithm, including a top-level agent and a bottom-level subagent. The top-level agent is responsible for global strategy formulation. Its state space contains the global information of the entire distribution network, and its action space contains the resource allocation strategy and global optimization objectives between sub-regions. The reward function comprehensively considers the power supply reliability, economy and environmental protection of the entire distribution network.

[0077] The bottom-level subagent is responsible for local execution. Its state space includes the real-time operating parameters, historical fault records and maintenance plans of all equipment in the sub-area. Its action space includes equipment control instructions, fault diagnosis decisions and maintenance personnel scheduling strategies. The reward function considers the power supply reliability index, equipment health index and operation and maintenance cost in the sub-area.

[0078] Build a global coordinator to collect information from each intelligent operation and maintenance sub-agent, combine the global strategy of the top-level agent, solve the global optimal decision, and decompose it into sub-decisions and send them to each bottom-level sub-agent for execution.

[0079] Based on the digital twin model of hybrid power distribution network, a multi-agent reinforcement learning framework is designed to realize intelligent operation and maintenance of distribution network. The specific implementation is as follows:

[0080] First, the topological structure of the distribution network is represented as a time-varying graph. The time-varying graph contains three key components: a fixed node set, a time-varying edge set, and a time-varying node feature matrix. The fixed node set is used to represent various types of distribution equipment in the distribution network, such as transformers, switches, busbars, etc. The time-varying edge set is used to represent power lines, whose connection status may change over time; the time-varying node feature matrix contains detailed information of each node (i.e., distribution equipment), such as equipment type, rated capacity, current operating status, etc. For example, for a distribution network containing 100 nodes, a 100x100 adjacency matrix can be used to represent the edge set, and a 100xm matrix can be used to represent the node features, where m is the feature dimension of each node.

[0081] Next, a multi-layer graph attention network is constructed to extract the topological features of the distribution network. Each layer of the graph attention network contains multiple independent attention heads, and each attention head performs weighted aggregation on the features of neighboring nodes through different weight matrices. For example, using 8 attention heads, the output dimension of each head is 64, and the output dimension of a single-layer graph attention network is 512. By stacking multiple layers (such as 3 layers) of graph attention networks, a wider range of node relationships can be captured. Finally, the embedded representation of each node is obtained with a dimension of 512.

[0082] Based on the node embedding representation, the similarity matrix between nodes is constructed. The Gaussian kernel function is used to calculate the similarity between node pairs, and the bandwidth parameter of the kernel function can be determined by cross-validation. The obtained similarity matrix is ​​used to calculate the standardized Laplace matrix. The eigenvectors are extracted from the standardized Laplace matrix and the eigenvectors are normalized to obtain the normalized eigenvector matrix.

[0083] A clustering algorithm (such as K-means) is used to cluster each row of the normalized feature vector matrix to obtain a preliminary sub-region division result. The initial number of clusters can be set to 10, and the optimal number of clusters is determined through iterative optimization.

[0084] In order to further optimize the sub-region division results, an evaluation index based on electrical distance and load balance is introduced. This index comprehensively considers the average electrical distance within the sub-region and the total load of the sub-region. The electrical distance can be calculated through the impedance matrix between nodes, and the total load is based on the power demand of each node. The optimal sub-region division result is determined by adjusting the number of clusters (such as varying between 5-15) and minimizing the evaluation index.

[0085] For each divided sub-area, an intelligent operation and maintenance sub-agent is assigned. Each intelligent operation and maintenance sub-agent adopts a hierarchical reinforcement learning algorithm, including a top-level agent and a bottom-level sub-agent. The top-level agent is responsible for global strategy formulation, and its state space contains global information of the entire distribution network, such as power supply reliability indicators, economic indicators, and environmental protection indicators of each sub-area. The action space contains resource allocation strategies between sub-areas (such as maintenance personnel scheduling) and global optimization objectives (such as minimizing the overall failure rate). The reward function comprehensively considers the power supply reliability (such as the system average outage frequency index SAIFI), economy (such as operation and maintenance costs) and environmental protection (such as carbon emissions) of the entire distribution network.

[0086] The bottom-level subagent is responsible for local execution, and its state space is more refined, including the real-time operating parameters (such as voltage, current, temperature, etc.), historical fault records and maintenance plans of all equipment in the sub-area. The action space contains specific equipment control instructions (such as adjusting transformer taps), fault diagnosis decisions and maintenance personnel scheduling strategies. The reward function considers the power supply reliability indicators (such as the customer average outage time index CAIDI), equipment health indicators (such as remaining service life) and operation and maintenance costs in the sub-area.

[0087] In order to coordinate the behavior of each intelligent operation and maintenance sub-agent, a global coordinator is built. The global coordinator collects the status information and local decisions of each intelligent operation and maintenance sub-agent, combines the global strategy formulated by the top-level agent, and solves the global optimal decision through a multi-objective optimization algorithm (such as NSGA-II). Then the global optimal decision is decomposed into sub-decisions and sent to each bottom-level sub-agent for execution.

[0088] Through the above steps, a distribution network intelligent operation and maintenance method based on multi-agent reinforcement learning is realized. This method makes full use of the topological characteristics of the distribution network, realizes reasonable sub-area division and hierarchical decision-making, and effectively improves the operation and maintenance efficiency and reliability of the distribution network.

[0089] By representing the topological structure of the distribution network as a time-varying graph, the dynamic characteristics and complex structure of the distribution network can be accurately captured. This representation method not only considers the static equipment layout, but also includes the time-varying line connection and equipment operation status, providing a comprehensive and real-time network view for subsequent analysis and decision-making. The sub-area division using a multi-layer graph attention network and a spectral clustering algorithm can fully consider the topological relationship and electrical characteristics between nodes. Compared with the traditional division method, this method is more intelligent and adaptive, and can automatically adjust the division results according to the dynamic changes of the network, improving the accuracy and flexibility of the division. By introducing evaluation indicators based on electrical distance and load balance, electrical characteristics and load distribution can be considered simultaneously when dividing sub-areas. The state space design of the top-level agent and the bottom-level sub-agent covers multiple levels from global network information to local equipment parameters, realizing a comprehensive perception of the operation status of the distribution network. By incorporating the maintenance personnel scheduling strategy into the action space of the bottom-level sub-agent, the intelligent management of operation and maintenance human resources is realized, and the utilization efficiency of human resources and the operation and maintenance response speed are improved.

[0090] In an optional embodiment,

[0091] Based on the preliminary sub-region division results, an evaluation index based on electrical distance and load balance is introduced. The evaluation index considers the average electrical distance within the sub-region and the total load of the sub-region. By adjusting the number of clusters and minimizing the evaluation index, the optimal sub-region division results are determined, including:

[0092] Based on the preliminary sub-region division results, a weighted adjacency graph between sub-regions is constructed, where the nodes of the graph represent sub-regions, and the weights of the edges comprehensively consider the electrical distance, power exchange, and topological connection strength between sub-regions.

[0093] Calculating the normalized Laplace matrix of the weighted adjacency graph and solving its eigenvalues ​​and eigenvectors, constructing a low-dimensional embedding space based on the eigenvectors, applying an improved spectral clustering algorithm in the low-dimensional embedding space, dynamically adjusting cluster centers and the number of clusters through an iterative optimization process, and obtaining optimized sub-region divisions;

[0094] Identify the boundary nodes between sub-areas, construct a boundary node impact assessment model, calculate the impact of the boundary nodes on the dynamic evaluation index function based on the boundary node impact assessment model, and adjust the attribution of the boundary nodes based on a greedy strategy to minimize the value of the dynamic evaluation index function;

[0095] The distribution network is divided into multiple overlapping local areas, each of which is assigned to a computing node. Clustering parameters and dynamic evaluation index weights are automatically adjusted on each computing node according to the characteristics of the local area. At the same time, a distributed consensus mechanism based on blockchain is designed to exchange boundary node information and local optimization results between computing nodes.

[0096] The federated learning method is used to integrate the local partitioning results of each computing node, construct a global sub-region partitioning scheme, and calculate the dynamic evaluation index function value of the global sub-region partitioning scheme. If it is better than the current optimal scheme, the optimal partitioning result is updated, and the iteration is repeated until the preset number of iterations is reached. The global sub-region partitioning scheme with the smallest dynamic evaluation index function value is output as the final result.

[0097] The specific implementation is as follows:

[0098] First, a weighted adjacency graph between sub-regions is constructed based on the preliminary sub-region division results. The nodes of the graph represent each sub-region, and the weight of the edge comprehensively considers the electrical distance, power exchange, and topological connection strength between sub-regions. For example, the electrical distance can be defined as the sum of the impedances on the shortest path between sub-regions, the power exchange can be defined as the amount of active power and reactive power exchanged between sub-regions, and the topological connection strength can be defined as the number of lines connecting two sub-regions. The comprehensive weight of the edge is obtained by weighted summing these three factors.

[0099] Next, the normalized Laplace matrix of the constructed weighted adjacency graph is calculated. The normalized Laplace matrix can be calculated by the adjacency matrix and the degree matrix. Then the eigenvalues ​​and eigenvectors of the matrix are solved, and the k eigenvectors corresponding to the minimum non-zero eigenvalue are selected to construct a k-dimensional low-dimensional embedding space. In this low-dimensional space, the improved spectral clustering algorithm is applied to re-divide the sub-regions. The improved spectral clustering algorithm dynamically adjusts the cluster center and the number of clusters through an iterative optimization process. For example, an adaptive K-means algorithm can be used to automatically determine the optimal number of clusters according to the data distribution characteristics.

[0100] After that, the boundary nodes between sub-areas are identified and the boundary node impact assessment model is constructed. The model takes into account the electrical characteristics, load characteristics and connection relationship of the boundary nodes with the adjacent sub-areas. Based on the model, the impact of each boundary node on the dynamic evaluation index function is calculated. At the same time, the greedy strategy is used to adjust the attribution of the boundary nodes, that is, try to attribute the boundary nodes to different adjacent sub-areas one by one, calculate the change of the dynamic evaluation index function value, and select the attribution scheme that minimizes the index function value.

[0101] Then, the distribution network is divided into multiple overlapping local areas, each of which is assigned to a computing node. On each computing node, the clustering parameters (such as the number of clusters, distance measurement method, etc.) and the weights of dynamic evaluation indicators are automatically adjusted according to the network topology, load distribution and other characteristics of the local area. At the same time, a distributed consensus mechanism based on blockchain is designed to exchange boundary node information and local optimization results between computing nodes. Specifically, consensus algorithms such as proof of work (PoW) or proof of stake (PoS) can be used to ensure the consistency and security of information exchange between nodes.

[0102] Finally, the federated learning method is used to fuse the local partitioning results of each computing node to construct a global sub-region partitioning scheme. The federated learning framework allows each computing node to collaboratively train the global model without sharing the original data. In this scheme, the FedAvg algorithm can be used. Each node uploads the local model parameters to the central server. The server performs weighted averaging on the parameters to obtain the global model, and then sends the global model to each node for local update. After multiple rounds of iterations, a global sub-region partitioning scheme that integrates all local information is finally obtained.

[0103] The dynamic evaluation index function value is calculated for the global sub-region division scheme, which comprehensively considers factors such as the average electrical distance within the sub-region, the total load balance of the sub-region, and the number of boundary nodes. If the obtained index function value is better than the current optimal solution, the optimal division result is updated. The above process is repeated until the preset number of iterations (such as 100 times) is reached, and finally the global sub-region division scheme with the smallest dynamic evaluation index function value is output as the final result.

[0104] In practical applications, appropriate parameters can be set according to the specific conditions of the distribution network. For example, for a medium-sized distribution network containing 1,000 nodes, it can be initially divided into 10 sub-areas, the low-dimensional embedding space selects 3 dimensions, and the number of iterations is set to 50 times. In each iteration, the range of dynamically adjusted cluster numbers can be set to 8-12. The boundary node impact assessment model can adopt a neural network-based regression model, and the input features include the node's voltage level, load size, number of connecting lines, etc. Distributed computing can deploy 20 computing nodes, each of which is responsible for the local optimization of about 100 distribution network nodes. In this way, the computing efficiency can be improved while ensuring the optimization quality, which is suitable for the dynamic division of large-scale distribution networks.

[0105] In this embodiment, by constructing a weighted adjacency graph between sub-regions, the electrical distance, power exchange and topological connection strength are comprehensively considered, and the relationship between sub-regions can be described more accurately. This method can capture the complex interdependence in the distribution network and provide a more accurate basis for subsequent optimization. The low-dimensional embedding space is constructed using the eigenvalues ​​and eigenvectors of the normalized Laplace matrix, which effectively reduces the complexity of the problem. The improved spectral clustering algorithm is applied in the low-dimensional space to more efficiently optimize the sub-region partitioning while retaining the key features of the original high-dimensional data. The cluster center and the number of clusters are dynamically adjusted through the iterative optimization process, so that the sub-region partitioning can be adaptively adjusted to adapt to changes in the network structure. This dynamicity improves the flexibility and adaptability of the partitioning results. By iteratively optimizing the global sub-region partitioning scheme and continuously updating the optimal partitioning results, it is ensured that the final partitioning scheme is the optimal solution in the global scope. The distributed computing architecture and iterative optimization process enable the system to quickly respond to changes in the network structure or operating state, and realize near-real-time sub-region partitioning adjustment. This enables the solution to efficiently, accurately and safely complete the sub-area division task in a complex, large-scale and dynamically changing distribution network environment, providing strong technical support for the intelligent operation, maintenance and management of the distribution network.

[0106] S102. The intelligent operation and maintenance agent monitors and analyzes the digital twin model in real time, and uses the attention mechanism and causal reasoning algorithm to identify potential fault risks and abnormal conditions. When an abnormality is detected, the intelligent operation and maintenance agent activates the multi-level fault diagnosis system, integrates the multi-level diagnosis results, and automatically generates a fault diagnosis report. According to the fault diagnosis report, the preset processing solution library is called, and combined with the current system status, a personalized fault handling solution is formulated, and based on the fault handling solution, combined with real-time road conditions and the location of maintenance personnel, an optimal execution plan is formulated.

[0107] The intelligent operation and maintenance agent monitors and analyzes the digital twin model in real time, uses the attention mechanism to identify abnormal conditions of key equipment, and uses the causal reasoning algorithm to analyze the root cause of the abnormality. For example, if the temperature of a transformer rises abnormally, the attention mechanism will focus on the status of the transformer and its related equipment, and the causal reasoning algorithm will analyze that the temperature rise may be caused by overload or cooling system failure.

[0108] When an anomaly is detected, the intelligent operation and maintenance agent activates a multi-level fault diagnosis system. The system includes three levels: device level, feeder level, and system level. Device-level diagnosis is based on the historical data and expert rules of a single device; feeder-level diagnosis considers the correlation between upstream and downstream devices on the feeder; and system-level diagnosis analyzes the scope and root cause of the fault from the perspective of the entire distribution network. Combining the diagnostic results of the three levels, a fault diagnosis report containing information such as fault type, severity, and possible causes is automatically generated.

[0109] According to the fault diagnosis report, the intelligent operation and maintenance agent calls the preset processing solution library. The solution library contains the processing flow and operation steps for various common faults. The intelligent operation and maintenance agent dynamically adjusts the preset solution based on the current system status, such as load conditions and weather conditions, and formulates a personalized fault handling solution. For example, for a transformer overload fault, a load transfer solution can be used when the load is light, and a standby transformer can be considered when the load is heavy.

[0110] Based on the fault handling plan, the intelligent operation and maintenance agent further combines real-time traffic information and maintenance personnel location information to formulate the optimal execution plan. Specifically, the path planning algorithm is used to plan the shortest path for maintenance personnel, and dynamic adjustments are made considering traffic congestion; at the same time, maintenance tasks are reasonably allocated according to the maintenance personnel's professional skills and current workload to ensure the optimization of personnel scheduling.

[0111] In an optional embodiment,

[0112] According to the fault diagnosis report, call the preset processing solution library, and combine the current system status to develop a personalized fault processing solution including:

[0113] Use natural language processing technology to extract key information from historical fault diagnosis reports, maintenance records and expert experience, and build a knowledge graph containing entities such as equipment, fault type and processing method;

[0114] Integrate new fault diagnosis reports and processing experience into the knowledge graph through an incremental learning framework, dynamically update the knowledge graph, and map the received fault diagnosis reports to entities and relationships in the updated knowledge graph;

[0115] Use the relational graph convolutional network to perform representation learning on the mapped knowledge graph to obtain the vector representation of the graph entity. The multi-hop reasoning algorithm based on the attention mechanism performs path search on the knowledge graph represented by the vector to obtain the initial set of fault handling solutions.

[0116] Construct a sequence generation model based on the transformer architecture, including a multi-layer self-attention encoder and decoder, where the encoder encodes the fault diagnosis report and system status into a context vector sequence;

[0117] A pointer generation network with a copy mechanism is used in the decoder. The attention mechanism is used to dynamically select whether to copy information from the input or generate new words from the vocabulary. The encoded context vector sequence and the initial fault handling solution set are used as the input of the decoder. The sequence generation model is trained to maximize the likelihood probability of generating personalized fault handling solutions.

[0118] Use the trained sequence generation model to generate personalized fault handling solutions based on the current fault diagnosis report, system status, and potential handling solution set, including handling steps, required resources, and expected results.

[0119] In order to formulate personalized fault handling solutions, this implementation provides a method based on knowledge graph and deep learning. The method first builds and updates the knowledge graph, and then uses the graph neural network and sequence generation model to generate personalized handling solutions. The specific implementation steps are as follows:

[0120] First, natural language processing technology is used to extract key information from historical fault diagnosis reports, maintenance records, and expert experience to build an initial knowledge graph. The knowledge graph contains three types of entities: equipment, fault type, and treatment method, as well as the relationships between them. For example, the fault type "transformer overheating" can be extracted and connected to the "transformer" equipment entity, and also to treatment method entities such as "check cooling system" and "replace insulating oil". In practical applications, named entity recognition and relationship extraction can be used to identify entities and relationships from texts, and to build a large-scale knowledge graph containing tens of thousands of nodes and edges.

[0121] Secondly, an incremental learning framework is used to continuously integrate new fault diagnosis reports and processing experience into the knowledge graph. Specifically, when a new fault report is received, the key entities and relationships in the report are first identified and then added to the existing knowledge graph. If a conflict with an existing entity or relationship is found, it is updated based on information such as timestamps. For example, the recommended processing method for a certain device may change with technological advances, in which case the corresponding graph relationships need to be updated. In this way, the knowledge graph can evolve dynamically and continuously accumulate new domain knowledge.

[0122] Next, the received fault diagnosis report is mapped to the updated knowledge graph. Specifically, keywords and phrases are first extracted from the report, and then matching entities and relationships are searched in the knowledge graph. For example, if the report mentions "abnormal increase in the temperature of the main transformer", it can be mapped to the "transformer overheating" fault type node in the graph. The purpose of this step is to convert unstructured text information into a structured graph representation, laying the foundation for subsequent processing.

[0123] Then, a relational graph convolutional network is used to learn the representation of the mapped knowledge graph. The network aggregates the features of each node with the features of its neighboring nodes through multi-layer graph convolution operations, and finally obtains a low-dimensional vector representation of each entity. These vectors not only encode the attributes of the entity itself, but also contain its structural information in the graph. For example, the vector representations of similar fault types will be close, while the vectors of different devices will be more scattered.

[0124] Based on the learned vector representation, a multi-hop reasoning algorithm based on the attention mechanism is used to perform path search on the knowledge graph to obtain an initial set of fault handling solutions. The algorithm starts from the fault type node and finally reaches the handling method node through multiple jumps. During this period, the attention mechanism is used to dynamically adjust the importance of different paths. For example, for the "transformer overheating" fault, a series of related processing steps such as "check the cooling system" - "clean the radiator" - "replace the cooling fan" may be found. In this way, various associated information in the graph can be comprehensively utilized to generate a preliminary handling solution.

[0125] Next, a sequence generation model based on the transformer architecture is constructed to generate the final personalized processing solution. The model includes a multi-layer self-attention encoder and decoder. The encoder encodes the fault diagnosis report and system status into a context vector sequence. For example, for input information such as "the temperature of the main transformer is abnormally high, and the current load is 80%", the encoder generates a series of vectors to represent the semantics of this information. The decoder uses a pointer generation network with a copy mechanism, which can choose to copy information from the input or generate new words from a predefined vocabulary according to the specific situation.

[0126] During the training phase, the encoded context vector sequence and the initial fault handling solution set are used as inputs to the decoder, and optimization is performed with the goal of maximizing the likelihood of generating personalized fault handling solutions. Specifically, a large number of historical fault cases and their corresponding handling solutions can be collected as training data to allow the model to learn how to generate appropriate handling steps based on specific circumstances. During the training process, the model will gradually learn how to combine input information with the initial solution to generate a more personalized and specific handling solution.

[0127] Finally, the trained sequence generation model is used to generate personalized fault handling solutions based on the current fault diagnosis report, system status, and potential handling solution set. The generated solution includes specific handling steps, required resources, and expected results.

[0128] Through the above steps, this method realizes the whole process from constructing knowledge graphs from historical data to generating personalized solutions using deep learning models. This method can not only effectively utilize historical experience, but also formulate flexible processing strategies according to specific circumstances, which is of great significance for improving fault handling efficiency.

[0129] In this embodiment, natural language processing technology is used to extract key information from historical data, build a knowledge graph, and realize the intelligent extraction and structured representation of professional knowledge implicit in unstructured text. This greatly improves the efficiency and accuracy of knowledge acquisition. Through the incremental learning framework, the dynamic update of the knowledge graph is realized. This ensures that the knowledge base is always kept up to date and can reflect the latest fault diagnosis experience and processing methods in a timely manner. Using a relational graph convolutional network for representation learning, combined with a multi-hop reasoning algorithm with an attention mechanism, it is possible to quickly and accurately search for paths in complex knowledge graphs. This greatly improves the efficiency and accuracy of fault handling solution generation. Through the knowledge graph and multi-hop reasoning, the system can provide a reasoning path for generating solutions and enhance the interpretability of decisions. This helps to improve users' trust in the system. By combining advanced technologies such as natural language processing, knowledge graphs, and deep learning, the transition from historical experience to intelligent fault handling is achieved. It not only improves the efficiency and accuracy of fault handling, but also provides a powerful decision support tool for the intelligent operation and maintenance of distribution networks, which is expected to significantly improve the overall operation and maintenance level and system reliability.

[0130] S103. The optimal execution plan is sent to the automation equipment and the maintenance personnel's mobile terminals through a secure encrypted channel. During the execution process, the intelligent operation and maintenance agent collects equipment response data and maintenance personnel feedback information in real time, and compares and analyzes them with the optimal execution plan. When an execution deviation is detected, the intelligent operation and maintenance agent dynamically adjusts the execution plan. After completing the fault handling, the intelligent operation and maintenance agent generates an operation and maintenance report based on the whole process data, stores the generated operation and maintenance report in the database, and extracts preventive maintenance rules from historical operation and maintenance records. Based on the extracted preventive maintenance rules and combined with the pre-trained equipment life prediction model, a preventive maintenance plan is formulated.

[0131] The intelligent operation and maintenance agent sends the optimal execution plan to the automation equipment and the maintenance personnel's mobile terminal through a secure encrypted channel. During the execution process, the equipment response data (such as switch action time) and maintenance personnel feedback information (such as arrival time on site, fault handling progress, etc.) are continuously collected. These real-time data are compared and analyzed with the optimal execution plan. When execution deviations are detected, such as abnormal operation of a certain device or delayed arrival of maintenance personnel, the intelligent operation and maintenance agent will dynamically adjust the execution plan, such as reallocating tasks or adjusting the operation sequence.

[0132] After the fault handling is completed, the intelligent operation and maintenance agent generates an operation and maintenance report based on the whole process data, including fault description, handling process, resource consumption, repair time and other information. The generated operation and maintenance report is stored in the database for subsequent analysis and optimization. At the same time, preventive maintenance rules are extracted from historical operation and maintenance records. If the failure rate of a certain type of equipment is found to be high, the maintenance frequency of this type of equipment will be increased.

[0133] Finally, based on the extracted preventive maintenance rules and combined with the pre-trained equipment life prediction model, a preventive maintenance plan is formulated. The equipment life prediction model considers factors such as equipment type, operating environment, and load conditions, and uses a deep learning algorithm to predict the remaining service life of the equipment. According to the prediction results and preventive maintenance rules, the equipment maintenance and replacement time is reasonably arranged to achieve active prevention and reduce the probability of failure.

[0134] In an optional embodiment,

[0135] Generate an operation and maintenance report based on the whole process data, store the generated operation and maintenance report in the database, and extract preventive maintenance rules from historical operation and maintenance records. Based on the extracted preventive maintenance rules and combined with the pre-trained equipment life prediction model, formulate a preventive maintenance plan including:

[0136] Collect equipment status data when a fault occurs, reasoning path during fault diagnosis, specific treatment measures taken, treatment results and effect evaluation to form full process data;

[0137] Process the whole process data using a generated pre-trained model based on a converter architecture to generate an operation and maintenance report, and store the generated operation and maintenance report in a document-based database to form a historical operation and maintenance record;

[0138] Using a named entity recognition model based on a bidirectional encoder representation to identify equipment names, fault types, and maintenance operations from the historical operation and maintenance records, and extract target information;

[0139] Applying association rule mining algorithm to process the extracted target information, mining the association between equipment status, operation behavior and fault, obtaining association rule set, performing time series analysis on the association rule set, identifying the time series event pattern leading to the fault, and forming preliminary preventive maintenance rules;

[0140] Evaluating the effectiveness of the preliminary preventive maintenance rules, screening out preventive maintenance rules with a credibility higher than a preset threshold, and forming a final preventive maintenance rule set;

[0141] Collect and analyze historical equipment operation data and input it into the pre-trained equipment life prediction model to calculate the expected remaining service life of each key component of the equipment and generate a life prediction report;

[0142] A preliminary preventive maintenance plan is formulated according to the final preventive maintenance rule set and the life prediction report, and the preliminary preventive maintenance plan is optimized by multi-objectives considering maintenance cost, equipment reliability and production influencing factors to obtain a final preventive maintenance plan.

[0143] First, collect the equipment status data when the fault occurs, the reasoning path during the fault diagnosis process, the specific treatment measures taken, the treatment results and effect evaluation to form the whole process data. For example, for an industrial robot, the joint angle, motor current, vibration frequency and other status data when the fault occurs can be collected; record the reasoning process of the fault diagnosis system, such as "detecting abnormal vibration of joint 2 → checking the transmission system → finding gear wear"; record the treatment measures for replacing gears; evaluate whether the robot's operating status has returned to normal after replacement. These data together constitute a complete fault handling process data.

[0144] Next, the whole process data is processed using a generative pre-trained model based on the converter architecture to generate an operation and maintenance report. The model can be a large language model based on the GPT series, which is fine-tuned to adapt to the text generation task in the field of industrial equipment maintenance. The model input is the above-mentioned whole process data, and the output is a structured operation and maintenance report, which contains chapters such as fault description, cause analysis, processing process, and processing results. The generated operation and maintenance report is stored in a document-based database such as MongoDB to form a retrievable historical operation and maintenance record.

[0145] Then, a named entity recognition model based on bidirectional encoder representation (BERT) is used to identify the equipment name, fault type, and maintenance operation from the historical operation and maintenance records to extract the target information. For example, from the sentence "Replace the robot joint gear No. 2 to eliminate abnormal vibration", information such as "robot" (equipment name), "abnormal vibration" (fault type), and "replace gear" (maintenance operation) can be extracted. This step converts unstructured text information into structured data for subsequent analysis.

[0146] Subsequently, an association rule mining algorithm (such as the Apriori algorithm) is applied to process the extracted target information, mining the associations between the equipment status, operating behavior and faults to obtain an association rule set. For example, it may be found that there is a strong association between "the robot continuously works for more than 8 hours" and "abnormal vibration of the joints". The association rule set is subjected to a time series analysis to identify the time series event pattern that causes the fault and form a preliminary preventive maintenance rule. For example, "If the robot continuously works for nearly 8 hours, lubrication maintenance should be performed to prevent abnormal vibration of the joints."

[0147] The effectiveness of the preliminary preventive maintenance rules is evaluated, and preventive maintenance rules with a credibility higher than a preset threshold are screened out to form a final preventive maintenance rule set. The credibility can be measured by the applicability of the rules in historical data. For example, setting the threshold to 0.8 means that only rules that can effectively prevent failures in more than 80% of similar situations are retained.

[0148] At the same time, the historical operation data of the equipment is collected and analyzed and input into the pre-trained equipment life prediction model to calculate the expected remaining service life of each key component of the equipment and generate a life prediction report. The model can be a time series prediction model based on the long short-term memory network (LSTM), with the input being the historical operation parameters of the equipment (such as working hours, load, etc.), and the output being the expected remaining life of each key component. For example, the prediction report may show that "the expected remaining life of joint gear No. 2 is 500 hours".

[0149] Finally, a preliminary preventive maintenance plan is formulated based on the final preventive maintenance rule set and the life prediction report. For example, combining the rule of "the robot needs maintenance after working continuously for 8 hours" and the prediction of "the remaining life of the gear is 500 hours", a maintenance plan of "lubricating the joints after work every day and replacing the gears after 500 hours" can be formulated. Considering maintenance costs, equipment reliability and production influencing factors, the preliminary preventive maintenance plan is optimized by multiple objectives to obtain the final preventive maintenance plan. The optimization process can use a genetic algorithm, with minimizing maintenance costs, maximizing equipment reliability and minimizing production interruption time as the objective functions, and obtain the Pareto optimal solution set through iterative calculation, from which the solution that best suits the actual situation is selected as the final plan.

[0150] In this embodiment, by collecting data of the entire fault process, a complete record of the entire fault handling process is achieved. This provides a comprehensive and detailed data basis for subsequent analysis and optimization. The use of a pre-trained model based on the converter architecture to automatically generate an operation and maintenance report greatly improves the efficiency and consistency of report generation and reduces manual errors. Through the association rule mining algorithm, the potential association between equipment status, operating behavior and faults is automatically discovered, which helps to deeply understand the cause of the fault and preventive measures. Using the pre-trained life prediction model, an accurate prediction of the remaining service life of key components of the equipment is achieved, providing an important reference for maintenance decisions. Based on accurate life prediction and optimized maintenance plans, the problems of over-maintenance and under-maintenance are avoided, and maintenance efficiency and resource utilization are improved. Improve the intelligence level of equipment operation and maintenance, realize the transformation from passive response to active prevention, significantly improve equipment reliability, optimize resource allocation, reduce operation and maintenance costs, and create greater economic benefits for enterprises. At the same time, this intelligent, data-driven operation and maintenance method also provides strong support for the digital transformation of enterprises.

[0151] In an optional embodiment,

[0152] Collect and analyze historical equipment operation data and input it into the pre-trained equipment life prediction model to calculate the expected remaining service life of each key component of the equipment and generate a life prediction report including:

[0153] Collecting and preprocessing historical operation data of the equipment, wherein the historical operation data includes time series measurement values ​​of multiple sensors;

[0154] Perform multi-scale feature extraction on the preprocessed data, respectively extract time domain features, frequency domain features and time-frequency domain features to obtain multi-scale features, and construct a multi-scale temporal convolutional network model based on the multi-scale features. The model includes multiple parallel one-dimensional convolution branches with different convolution kernel sizes, a scaled dot product attention mechanism, a residual connection and a temporal gating unit;

[0155] A multi-task learning framework is adopted, an adaptive moment estimation optimizer and a cosine annealing strategy are used to adjust the learning rate, and a combined loss function is used to train the multi-scale temporal convolutional network model, and the extracted multi-scale features are input into the trained multi-scale temporal convolutional network model to calculate the expected remaining service life of each key component of the equipment and its confidence interval;

[0156] Based on the expected remaining service life, generate a life prediction trend chart, analyze factors affecting the life based on the expected remaining service life, identify abnormal patterns that may cause a sharp decline in life based on the expected remaining service life, and conduct potential risk assessment;

[0157] Based on the expected remaining service life, life prediction trend chart, key factor analysis and potential risk assessment results, maintenance recommendations are generated, and the expected remaining service life and confidence interval of each key component, life prediction trend chart, key factor analysis affecting life, potential risk assessment and maintenance recommendations are integrated to generate a life prediction report.

[0158] First, collect and preprocess the historical operation data of the equipment. The historical operation data contains time series measurement values ​​of multiple sensors, such as temperature, pressure, vibration, etc. Clean the original data, remove outliers and missing values, and normalize the data to make the data distribution consistent. For example, a device has 10 sensors, each sensor collects data once an hour, and a year of data collection can obtain a data matrix of 87600×10.

[0159] Then, multi-scale feature extraction is performed on the preprocessed data. In the time domain, statistical features such as mean, standard deviation, kurtosis, etc. are extracted; in the frequency domain, Fourier transform is performed to extract spectral features; in the time-frequency domain, wavelet transform is performed to extract wavelet coefficients. For example, for temperature data, time domain features such as the highest temperature, lowest temperature, and average temperature within 24 hours, frequency domain features such as the periodicity of temperature changes, and time-frequency domain features such as temperature change characteristics at different time scales can be extracted.

[0160] Next, a multi-scale temporal convolutional network model is constructed based on the extracted multi-scale features. The model includes multiple parallel one-dimensional convolution branches, each with a different convolution kernel size to capture features at different time scales. For example, three parallel branches can be set with convolution kernel sizes of 3, 7, and 15, respectively, to capture hourly, daily, and weekly features. The model also includes a scaled dot product attention mechanism to learn the correlation between features; residual connections to alleviate the gradient vanishing problem; and a temporal gating unit to capture long-term dependencies.

[0161] Subsequently, the multi-task learning framework is used to train the multi-scale temporal convolutional network model. The adaptive moment estimation optimizer is used to adaptively adjust the learning rate of each parameter according to the first-order moment and second-order moment of the parameter. At the same time, the cosine annealing strategy is used to adjust the global learning rate. A larger learning rate is used for rapid convergence in the early stage of training, and a smaller learning rate is used for fine-tuning in the later stage. The loss function uses a combination of mean square error loss and smooth L1 loss, taking into account both regression accuracy and robustness. For example, the total loss function can be set to 0.7×mean square error loss+0.3×smooth L1 loss.

[0162] After the training is completed, the extracted multi-scale features are input into the trained model to calculate the expected remaining service life of each key component of the equipment and its confidence interval. For example, the expected remaining service life of a bearing is 500 hours, the confidence interval is [450 hours, 550 hours], and the confidence level is 95%.

[0163] Based on the expected remaining service life, a life prediction trend chart is generated, with time on the horizontal axis and remaining life on the vertical axis, which can intuitively display the trend of life changes. At the same time, key factors affecting life, such as operating load and ambient temperature, are analyzed, and their importance is determined by calculating the correlation between each factor and life. In addition, abnormal patterns that may cause a sharp decline in life, such as sudden abnormality in sensor data, are identified to conduct potential risk assessments.

[0164] Finally, maintenance recommendations are generated based on the above analysis results. For example, for components with a short expected life, it is recommended to arrange spare parts in advance; for key factors affecting life, it is recommended to optimize operating parameters; for potential risks, it is recommended to strengthen monitoring. The expected remaining service life and confidence interval of each key component, life prediction trend chart, analysis of key factors affecting life, potential risk assessment and maintenance recommendations are integrated to generate a complete life prediction report to provide a basis for equipment maintenance decisions.

[0165] In this embodiment, by collecting and preprocessing the time series measurement values ​​of multiple sensors, comprehensive monitoring and data utilization of the equipment operation status are achieved. By extracting time domain, frequency domain and time-frequency domain features, multi-dimensional and in-depth analysis of equipment operation data is achieved, and the expression ability and information volume of features are improved. The multi-scale time convolution network model can effectively capture feature patterns on different time scales and improve the model's learning ability for complex time series patterns. The multi-task learning framework, adaptive optimizer and learning rate adjustment strategy are adopted to improve the efficiency and effect of model training. Through the trained model, the expected remaining service life of each key component of the equipment can be accurately calculated, providing a reliable basis for maintenance decisions. The accuracy, reliability and intelligence level of equipment life prediction and management are improved. It not only provides enterprises with a powerful decision support tool, but also provides a technical basis for the innovation and optimization of equipment management mode.

[0166] Figure 2 FIG. 1 is a schematic diagram of the structure of an automatic operation and maintenance system for a power distribution network based on artificial intelligence according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0167] The first unit is used to collect the operation data of the power distribution network in real time through a distributed sensor network, transmit the collected operation data to the cloud data center, use a multimodal deep learning framework to fuse and analyze the historical operation data and the real-time operation data, and build a hybrid power distribution network digital twin model containing a physical model and a data-driven model based on the fusion analysis results. Based on the constructed hybrid power distribution network digital twin model, a multi-agent reinforcement learning framework is designed to divide the distribution network into multiple sub-areas, each of which corresponds to an intelligent operation and maintenance sub-agent. A hierarchical reinforcement learning algorithm is used to train the intelligent operation and maintenance agent, in which the top-level agent is responsible for global strategy formulation and the bottom-level sub-agent is responsible for local execution;

[0168] The second unit is used for the intelligent operation and maintenance agent to monitor and analyze the digital twin model in real time, and use the attention mechanism and causal reasoning algorithm to identify potential fault risks and abnormal conditions. When an abnormality is detected, the intelligent operation and maintenance agent activates the multi-level fault diagnosis system, integrates the multi-level diagnosis results, and automatically generates a fault diagnosis report. According to the fault diagnosis report, the preset processing solution library is called, and combined with the current system status, a personalized fault handling solution is formulated, and based on the fault handling solution, combined with real-time road condition information and the location of maintenance personnel, an optimal execution plan is formulated;

[0169] The third unit is used to send the optimal execution plan to the automation equipment and the maintenance personnel's mobile terminals through a secure encrypted channel. During the execution process, the intelligent operation and maintenance agent collects equipment response data and maintenance personnel feedback information in real time, and compares and analyzes them with the optimal execution plan. When an execution deviation is detected, the intelligent operation and maintenance agent dynamically adjusts the execution plan. After completing the fault handling, the intelligent operation and maintenance agent generates an operation and maintenance report based on the whole process data, stores the generated operation and maintenance report in the database, and extracts preventive maintenance rules from historical operation and maintenance records. Based on the extracted preventive maintenance rules and combined with the pre-trained equipment life prediction model, a preventive maintenance plan is formulated.

[0170] According to a third aspect of the embodiments of the present invention,

[0171] An electronic device is provided, comprising:

[0172] processor;

[0173] a memory for storing processor-executable instructions;

[0174] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0175] A fourth aspect of the embodiments of the present invention is:

[0176] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0177] 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 carrying computer-readable program instructions for executing various aspects of the present invention.

[0178] 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 it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An artificial intelligence-based automatic operation and maintenance method for power distribution network, characterized in that: include: The operation data of the power distribution network is collected in real time through a distributed sensor network, and the collected operation data is transmitted to the cloud data center. The historical operation data and the real-time operation data are fused and analyzed using a multimodal deep learning framework. Based on the fusion analysis results, a hybrid power distribution network digital twin model containing a physical model and a data-driven model is constructed. Based on the constructed hybrid power distribution network digital twin model, a multi-agent reinforcement learning framework is designed to divide the distribution network into multiple sub-areas, each of which corresponds to an intelligent operation and maintenance sub-agent. A hierarchical reinforcement learning algorithm is used to train the intelligent operation and maintenance agents, in which the top-level agent is responsible for global strategy formulation and the bottom-level sub-agent is responsible for local execution. The intelligent operation and maintenance agent monitors and analyzes the digital twin model in real time, and uses the attention mechanism and causal reasoning algorithm to identify potential fault risks and abnormal conditions. When an abnormality is detected, the intelligent operation and maintenance agent activates the multi-level fault diagnosis system, integrates the multi-level diagnosis results, and automatically generates a fault diagnosis report. According to the fault diagnosis report, the preset processing solution library is called, and combined with the current system status, a personalized fault handling solution is formulated. Based on the fault handling solution, combined with real-time road condition information and the location of maintenance personnel, the optimal execution plan is formulated; The optimal execution plan is sent to the automation equipment and maintenance personnel's mobile terminals through a secure encrypted channel. During the execution process, the intelligent operation and maintenance agent collects equipment response data and maintenance personnel's feedback information in real time, and compares and analyzes them with the optimal execution plan. When an execution deviation is detected, the intelligent operation and maintenance agent dynamically adjusts the execution plan. After completing the fault handling, the intelligent operation and maintenance agent generates an operation and maintenance report based on the whole process data, stores the generated operation and maintenance report in the database, and extracts preventive maintenance rules from historical operation and maintenance records. Based on the extracted preventive maintenance rules and combined with the pre-trained equipment life prediction model, a preventive maintenance plan is formulated; Based on the constructed digital twin model of hybrid power distribution network, a multi-agent reinforcement learning framework is designed to divide the distribution network into multiple sub-areas. Each sub-area corresponds to an intelligent operation and maintenance sub-agent. A hierarchical reinforcement learning algorithm is used to train the intelligent operation and maintenance agents. The top-level agent is responsible for global strategy formulation, and the bottom-level sub-agent is responsible for local execution, including: The topological structure of the distribution network is represented as a time-varying graph, wherein the time-varying graph includes a fixed node set, a time-varying edge set, and a time-varying node feature matrix, wherein the fixed node set represents the distribution equipment, the time-varying edge set represents the power line, and the time-varying node feature matrix includes the equipment type, rated capacity, and operation status information; Construct a multi-layer graph attention network, each layer of which includes multiple independent attention heads, weighted aggregation of neighbor node features, and concatenation of the outputs of multiple attention heads to obtain an embedded representation of the node, construct a similarity matrix based on the embedded representation of the node, use a Gaussian kernel function to calculate the similarity between nodes, calculate a standardized Laplacian matrix based on the similarity matrix, and obtain the eigenvector of the standardized Laplacian matrix, normalize the eigenvector to obtain a normalized eigenvector matrix, use a clustering algorithm to cluster each row of the normalized eigenvector matrix, and obtain a preliminary sub-region division result; Based on the preliminary sub-region division results, an evaluation index based on electrical distance and load balance is introduced. The evaluation index considers the average electrical distance inside the sub-region and the total load of the sub-region. The optimal sub-region division result is determined by adjusting the number of clusters and minimizing the evaluation index. An intelligent operation and maintenance subagent is allocated to each sub-region. The intelligent operation and maintenance subagent adopts a hierarchical reinforcement learning algorithm, including a top-level agent and a bottom-level subagent. The top-level agent is responsible for global strategy formulation. Its state space contains the global information of the entire distribution network, and its action space contains the resource allocation strategy and global optimization objectives between sub-regions. The reward function comprehensively considers the power supply reliability, economy and environmental protection of the entire distribution network. The bottom-level subagent is responsible for local execution. Its state space includes the real-time operating parameters, historical fault records and maintenance plans of all equipment in the sub-area. Its action space includes equipment control instructions, fault diagnosis decisions and maintenance personnel scheduling strategies. The reward function considers the power supply reliability index, equipment health index and operation and maintenance cost in the sub-area. Build a global coordinator to collect information from each intelligent operation and maintenance sub-agent, combine the global strategy of the top-level agent, solve the global optimal decision, and decompose it into sub-decisions and send them to each bottom-level sub-agent for execution.

2. The method according to claim 1, characterized in that Based on the preliminary sub-region division results, an evaluation index based on electrical distance and load balance is introduced. The evaluation index considers the average electrical distance within the sub-region and the total load of the sub-region. By adjusting the number of clusters and minimizing the evaluation index, the optimal sub-region division results are determined, including: Based on the preliminary sub-region division results, a weighted adjacency graph between sub-regions is constructed, where the nodes of the graph represent sub-regions, and the weights of the edges comprehensively consider the electrical distance, power exchange, and topological connection strength between sub-regions. Calculating the normalized Laplace matrix of the weighted adjacency graph and solving its eigenvalues ​​and eigenvectors, constructing a low-dimensional embedding space based on the eigenvectors, applying an improved spectral clustering algorithm in the low-dimensional embedding space, dynamically adjusting cluster centers and the number of clusters through an iterative optimization process, and obtaining optimized sub-region divisions; Identify the boundary nodes between sub-areas, construct a boundary node impact assessment model, calculate the impact of the boundary nodes on the dynamic evaluation index function based on the boundary node impact assessment model, and adjust the attribution of the boundary nodes based on a greedy strategy to minimize the value of the dynamic evaluation index function; The distribution network is divided into multiple overlapping local areas, each of which is assigned to a computing node. Clustering parameters and dynamic evaluation index weights are automatically adjusted on each computing node according to the characteristics of the local area. At the same time, a distributed consensus mechanism based on blockchain is designed to exchange boundary node information and local optimization results between computing nodes. The federated learning method is used to integrate the local partitioning results of each computing node, construct a global sub-region partitioning scheme, and calculate the dynamic evaluation index function value of the global sub-region partitioning scheme. If it is better than the current optimal scheme, the optimal partitioning result is updated, and the iteration is repeated until the preset number of iterations is reached. The global sub-region partitioning scheme with the smallest dynamic evaluation index function value is output as the final result.

3. The method according to claim 1, characterized in that According to the fault diagnosis report, call the preset processing solution library, and combine the current system status to develop a personalized fault processing solution including: Use natural language processing technology to extract key information from historical fault diagnosis reports, maintenance records and expert experience, and build a knowledge graph containing entities such as equipment, fault type and processing method; Integrate new fault diagnosis reports and processing experience into the knowledge graph through an incremental learning framework, dynamically update the knowledge graph, and map the received fault diagnosis reports to entities and relationships in the updated knowledge graph; Use the relational graph convolutional network to perform representation learning on the mapped knowledge graph to obtain the vector representation of the graph entity. The multi-hop reasoning algorithm based on the attention mechanism performs path search on the knowledge graph represented by the vector to obtain the initial set of fault handling solutions. Construct a sequence generation model based on the transformer architecture, including a multi-layer self-attention encoder and decoder, where the encoder encodes the fault diagnosis report and system status into a context vector sequence; A pointer generation network with a copy mechanism is used in the decoder. The attention mechanism is used to dynamically select whether to copy information from the input or generate new words from the vocabulary. The encoded context vector sequence and the initial fault handling solution set are used as the input of the decoder. The sequence generation model is trained to maximize the likelihood probability of generating personalized fault handling solutions. Use the trained sequence generation model to generate personalized fault handling solutions based on the current fault diagnosis report, system status, and potential handling solution set, including handling steps, required resources, and expected results.

4. The method according to claim 1, characterized in that Generate an operation and maintenance report based on the whole process data, store the generated operation and maintenance report in the database, and extract preventive maintenance rules from historical operation and maintenance records. Based on the extracted preventive maintenance rules and combined with the pre-trained equipment life prediction model, formulate a preventive maintenance plan including: Collect equipment status data when a fault occurs, reasoning path during fault diagnosis, specific treatment measures taken, treatment results and effect evaluation to form full process data; Process the whole process data using a generated pre-trained model based on a converter architecture to generate an operation and maintenance report, and store the generated operation and maintenance report in a document-based database to form a historical operation and maintenance record; Using a named entity recognition model based on a bidirectional encoder representation to identify equipment names, fault types, and maintenance operations from the historical operation and maintenance records, and extract target information; Applying association rule mining algorithm to process the extracted target information, mining the association between equipment status, operation behavior and fault, obtaining association rule set, performing time series analysis on the association rule set, identifying the time series event pattern leading to the fault, and forming preliminary preventive maintenance rules; Evaluating the effectiveness of the preliminary preventive maintenance rules, screening out preventive maintenance rules with a credibility higher than a preset threshold, and forming a final preventive maintenance rule set; Collect and analyze historical equipment operation data and input it into the pre-trained equipment life prediction model to calculate the expected remaining service life of each key component of the equipment and generate a life prediction report; A preliminary preventive maintenance plan is formulated according to the final preventive maintenance rule set and the life prediction report, and the preliminary preventive maintenance plan is optimized by multi-objectives considering maintenance cost, equipment reliability and production influencing factors to obtain a final preventive maintenance plan.

5. The method according to claim 4, characterized in that Collect and analyze historical equipment operation data and input it into the pre-trained equipment life prediction model to calculate the expected remaining service life of each key component of the equipment and generate a life prediction report including: Collecting and preprocessing historical operation data of the equipment, wherein the historical operation data includes time series measurement values ​​of multiple sensors; Perform multi-scale feature extraction on the preprocessed data, respectively extract time domain features, frequency domain features and time-frequency domain features to obtain multi-scale features, and construct a multi-scale temporal convolutional network model based on the multi-scale features. The model includes multiple parallel one-dimensional convolution branches with different convolution kernel sizes, a scaled dot product attention mechanism, a residual connection and a temporal gating unit; A multi-task learning framework is adopted, an adaptive moment estimation optimizer and a cosine annealing strategy are used to adjust the learning rate, and a combined loss function is used to train the multi-scale temporal convolutional network model, and the extracted multi-scale features are input into the trained multi-scale temporal convolutional network model to calculate the expected remaining service life of each key component of the equipment and its confidence interval; Based on the expected remaining service life, generate a life prediction trend chart, analyze factors affecting the life based on the expected remaining service life, identify abnormal patterns that may cause a sharp decline in life based on the expected remaining service life, and conduct potential risk assessment; Based on the expected remaining service life, life prediction trend chart, key factor analysis and potential risk assessment results, maintenance recommendations are generated, and the expected remaining service life and confidence interval of each key component, life prediction trend chart, key factor analysis affecting life, potential risk assessment and maintenance recommendations are integrated to generate a life prediction report.

6. An artificial intelligence-based power distribution network automation operation and maintenance system, used to implement the method described in any one of claims 1 to 5, characterized in that: include: The first unit is used to collect the operation data of the power distribution network in real time through a distributed sensor network, transmit the collected operation data to the cloud data center, use a multimodal deep learning framework to fuse and analyze the historical operation data and the real-time operation data, and build a hybrid power distribution network digital twin model containing a physical model and a data-driven model based on the fusion analysis results. Based on the constructed hybrid power distribution network digital twin model, a multi-agent reinforcement learning framework is designed to divide the distribution network into multiple sub-areas, each of which corresponds to an intelligent operation and maintenance sub-agent. A hierarchical reinforcement learning algorithm is used to train the intelligent operation and maintenance agent, in which the top-level agent is responsible for global strategy formulation and the bottom-level sub-agent is responsible for local execution; The second unit is used for the intelligent operation and maintenance agent to monitor and analyze the digital twin model in real time, and use the attention mechanism and causal reasoning algorithm to identify potential fault risks and abnormal conditions. When an abnormality is detected, the intelligent operation and maintenance agent activates the multi-level fault diagnosis system, integrates the multi-level diagnosis results, and automatically generates a fault diagnosis report. According to the fault diagnosis report, the preset processing solution library is called, and combined with the current system status, a personalized fault handling solution is formulated, and based on the fault handling solution, combined with real-time road condition information and the location of maintenance personnel, an optimal execution plan is formulated; The third unit is used to send the optimal execution plan to the automation equipment and the maintenance personnel's mobile terminals through a secure encrypted channel. During the execution process, the intelligent operation and maintenance agent collects equipment response data and maintenance personnel feedback information in real time, and compares and analyzes them with the optimal execution plan. When an execution deviation is detected, the intelligent operation and maintenance agent dynamically adjusts the execution plan. After completing the fault handling, the intelligent operation and maintenance agent generates an operation and maintenance report based on the whole process data, stores the generated operation and maintenance report in the database, and extracts preventive maintenance rules from historical operation and maintenance records. Based on the extracted preventive maintenance rules and combined with the pre-trained equipment life prediction model, a preventive maintenance plan is formulated.

7. 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 5.

8. 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 5 is implemented.

Images