Power grid operation order generation system based on multi-agent cooperative work

Abstract

The invention discloses a power grid operation ticket generation system based on multi-agent cooperative work, and the system comprises a rule knowledge graph which stores a power grid operation ticket rule and a relation thereof; the graph neural network is used for matching rules in the rule knowledge graph; each agent is responsible for different subtasks, information is shared through reinforcement learning collaborative decision and an attention mechanism, and finally an operation ticket is generated; and the distributed workflow receives the operation order generated by the intelligent agent, converts the operation order into a modular workflow, stores the modular workflow in a code block or Prompt form and supports dynamic updating, and each module corresponds to a specific operation step or task. The system efficiently generates the power grid operation order by means of multi-agent cooperation and in combination with the knowledge graph and the graph neural network, the accuracy and intelligent degree of operation order generation are improved, the distributed workflow is convenient for dynamic management and updating, the system can better adapt to complex and changeable operation requirements of a power grid, and safe and stable operation of the power grid is guaranteed.

Description

Technical Field

[0001] This invention relates to a generation system, and more specifically to a power grid operation ticket generation system based on multi-agent collaborative work. Background Technology

[0002] Traditional power grid operation ticket generation relies on manual writing, which is inefficient and prone to rule omissions. Existing template-based generation methods lack the ability to dynamically adapt to complex scenarios. As the scale and complexity of the power grid increase, there is an urgent need for an intelligent system that can automatically generate operation tickets, dynamically adjust the rule knowledge graph, and achieve self-optimization. Summary of the Invention

[0003] In view of the shortcomings of the existing technology, the purpose of this invention is to provide a multi-agent driven power grid operation ticket generation system that can solve one or more of the above-mentioned problems.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a power grid operation ticket generation system based on multi-agent collaborative work, comprising: A rule knowledge graph, which stores the rules and relationships of power grid operation tickets; A graph neural network is used to match rules in a rule knowledge graph. Multiple intelligent agents, each responsible for different sub-tasks, collaborate in decision-making through reinforcement learning and share information through attention mechanisms, ultimately generating operation tickets; The distributed workflow receives operation tickets generated by multiple agents and transforms them into modular workflows. Each module corresponds to a specific operation step or task. The workflow is stored in the form of code blocks or prompts and supports dynamic updates.

[0005] As a further improvement of the present invention, the rule knowledge graph is constructed through the following steps: Step 1: Extract rules from historical operation tickets to construct an initial knowledge graph. Specifically, the historical operation tickets are first segmented, entity recognized, and semantically parsed. Then, key entities are extracted using named entity recognition technology. Next, logical relationships between entities are identified using relation extraction technology. Finally, the extracted entities and relationships are stored in the form of a graph structure to obtain the initial knowledge graph. Step 2: Extract rules from the newly added operation tickets and dynamically update the knowledge graph. The specific method of dynamic updating is as follows: First, perform consistency detection between the newly added rules and the rules in the existing graph to avoid rule conflicts. Then, for the newly added rules, use a deep learning model to generate the semantic representation of the rules and integrate it into the existing graph through vector similarity calculation. Finally, optimize the representation of the graph through graph embedding technology to ensure the usability and efficiency of the rule knowledge graph. Step 3: Dynamically expand the rule graph. Specifically, this involves first discovering latent rules through unsupervised learning, and then dynamically adjusting the priority of rules based on their usage frequency and importance in historical data.

[0006] As a further improvement of the present invention, the specific steps for rule matching by the graph neural network are as follows: Step 1: Convert the input entity (X) and rule knowledge graph (G) in the operation ticket into vector representations, and generate node feature matrix (H) and adjacency matrix (A) through graph embedding technology; Step 2: Update node features using a graph convolutional network to capture higher-order relationships; Step 3: Calculate the degree of matching between the input entity and the rule using the matching score function; In the matching process described above, a Transformer-based attention mechanism is introduced to enhance the context awareness of rule matching; at the same time, the rule matching and rule priority prediction tasks are optimized to improve the generalization ability of the model.

[0007] As a further improvement of the present invention, the plurality of intelligent agents include: Rule extraction agent: responsible for extracting rules from historical operation tickets; Rule-optimizing agent: used to dynamically adjust the rule knowledge graph; Operation ticket generation agent: used to generate operation tickets based on rule knowledge graph; Workflow generation agent: used to transform operation tickets into modular workflows; Module optimization agent: Used to optimize workflow modules themselves; Global optimization agent: Used to perform global optimization of the entire workflow.

[0008] As a further improvement to the present invention, the specific method of collaborative decision-making among the multiple intelligent agents is as follows: Step 4: Represent the state of each agent as (s_i^t), and the global state as (S^t = {s_1^t, s_2^t, ..., s_n^t}); Step 5: Obtain the policy of each agent as (\pi_i), and optimize it through deep reinforcement learning and a global reward function; The global reward function (R (S^t, a^t)) combines the accuracy, security, and efficiency of operation ticket generation.

[0009] As a further improvement of the present invention, the distributed workflow is dynamically updated using module optimization and global optimization.

[0010] As a further improvement of the present invention, a personalized recommendation and privacy protection module is also included. This module captures user behavior to build a user preference model, and also avoids data leakage by prioritizing the local processing of sensitive data and training the model through federated learning, and protects individual characteristics by adding noise to the data. The beneficial effects of this invention are: Deep learning-driven rule knowledge graph update: By combining deep learning technology, the system can automatically extract rules from historical operation records and dynamically optimize the knowledge graph through machine learning algorithms.

[0011] Rule matching based on graph neural networks: Graph neural networks (GNNs) are used to achieve efficient matching of rule knowledge graphs, thereby improving the accuracy and intelligence of operation ticket generation.

[0012] Multi-agent reinforcement learning collaborative decision-making: Agents make collaborative decisions through reinforcement learning algorithms to optimize the operation ticket generation process and ensure that the generated operation tickets conform to the real-time power grid status.

[0013] Distributed workflow optimization architecture: It adopts a distributed computing architecture to achieve real-time optimization of workflows and supports modular design and dynamic updates.

[0014] Personalized recommendations and privacy protection integration: Personalized recommendations are made through user behavior analysis, and differential privacy technology is used to ensure user data security. Attached Figure Description

[0015] Figure 1 The block diagram of the improved multi-agent driven power grid operation ticket generation system; Figure 2 Flowchart for updating rule-based knowledge graphs driven by deep learning; Figure 3 This is a flowchart of rule matching based on graph neural networks; Figure 4 A flowchart for multi-agent reinforcement learning collaborative decision-making; Figure 5 A schematic diagram of an optimized architecture for distributed workflows. Detailed Implementation

[0016] The present invention will now be described in further detail with reference to the embodiments shown in the accompanying drawings.

[0017] Reference Figures 1 to 5 As shown in the figure, a power grid operation ticket generation system based on multi-agent collaborative work in this embodiment specifically includes the following: 1. Rule-based knowledge graphs driven by deep learning 1.1 Construction of the rule knowledge graph The rule knowledge graph is the core of the entire system, used to store the rules and relationships of power grid operation tickets. The initial knowledge graph (G^0) is constructed by extracting rules from historical operation tickets, specifically through the following steps: Data preprocessing: Historical operation tickets are segmented into words, identified as entities, and semantically parsed; Rule-based entity extraction: Extract key entities (such as device name, operating steps, precautions, etc.) using Named Entity Recognition (NER) technology. Rule-based relationship identification: Using relationship extraction techniques to identify logical relationships (such as causal relationships, sequential relationships, etc.) between entities; Graph construction: The extracted entities and relations are stored in the form of a graph structure. The initial graph is (G^0 = (V, E)), where (V) represents the set of entities and (E) represents the set of relations.

[0018] 1.2 Dynamically Updated Rule Knowledge Graph During operation, the system extracts rules from newly added operation tickets and dynamically updates the knowledge graph (G). The update process includes the following steps: Rule conflict detection: Perform consistency checks between newly added rules and existing rules in the graph to avoid rule conflicts; Rule fusion: For new rules, a deep learning model (such as Transformer) is used to generate the semantic representation of the rule, and then the representation is fused into the existing graph through vector similarity calculation; Graph optimization: Optimize the representation of the graph through graph embedding techniques (such as graph embedding) to ensure the usability and efficiency of the rule knowledge graph.

[0019] The mathematical representation of the update process is: [G^{t+1} = \text{Update}(G^t, \text{DeepLearn}(D^t))] where (D^t) represents the new dataset, (\text{DeepLearn}) is the deep learning model, and (\text{Update}) is the graph update function.

[0020] 1.3 Dynamic Expansion of Rule Knowledge Graph To adapt to the complexity of power grid operation, the system supports dynamically expanding the rule graph: New rule discovery: Discovering latent rules through unsupervised learning methods (such as cluster analysis); Rule priority adjustment: The priority of rules is dynamically adjusted based on the frequency and importance of rule usage in historical data.

[0021] 2. Graph Neural Networks for Rule Matching 2.1 Introduction to Graph Neural Networks Graph Neural Networks (GNNs) are used to efficiently match rules in rule-based knowledge graphs. GNNs can capture higher-order relationships between nodes and edges in graph-structured data, making them suitable for handling complex rule matching problems.

[0022] 2.2 Rule Matching Process Input representation: The input entity (X) and rule knowledge graph (G) in the operation ticket are transformed into vector representations, and node feature matrix (H) and adjacency matrix (A) are generated through graph embedding technology; Graph Convolution Operation: Update node features through Graph Convolutional Network (GCN) to capture higher-order relations: [H^{(l+1)} = \sigma(AH{(l)}W{(l)})] where (H^{(l)}) is the node feature of layer (l), (W^{(l)}) is the learnable weight matrix, and (\sigma) is the activation function; Rule matching score calculation: The degree of matching between the input entity and the rule is calculated using the matching score function: [\text{Score}(X, G) = \text{softmax}(H^{(L)} \cdot W_{\text{match}})] where (H^{(L)}) is the node feature of the last layer, and (W_{\text{match}}) is the matching weight matrix.

[0023] 2.3 Rule Matching Optimization Attention mechanism: Introduce a Transformer-based attention mechanism to enhance the context awareness of rule matching; Multi-task learning: Simultaneously optimize rule matching and rule priority prediction tasks to improve the model's generalization ability.

[0024] 3. Enhance multi-agent learning and collaborative decision-making 3.1 Design of Intelligent Agents The system is designed with multiple intelligent agents, each responsible for different sub-tasks: The rule extraction agent is responsible for extracting rules from historical operation tickets. Rule-optimizing agent: dynamically adjusts the rule knowledge graph; Operation ticket generation agent: Generates operation tickets based on a rule knowledge graph; Workflow generation agent: Transforms operation tickets into modular workflows; Module optimization agent: Self-optimizes workflow modules; Global optimization agent: Performs global optimization of the entire workflow.

[0025] 3.2 Introduction to Reinforcement Learning Cooperative decision-making among intelligent agents is achieved through reinforcement learning, specifically including: State representation: The state of each agent is represented as (s_i^t), and the global state is (S^t = {s_1^t,s_2^t, ..., s_n^t}). Policy function: The policy of each agent is (\pi_i ), which is optimized through deep reinforcement learning (such as DQN or PPO); Reward function: The global reward function ( R(S^t, a^t) ) combines the accuracy, security and efficiency of operation ticket generation.

[0026] The collaborative optimization objective is: [J(\pi) = \mathbb{E}\left[\sum_{t=0}^{T} \gamma^t \cdot R(S^t, a^t)\right]] where ( \gamma ) is the discount factor.

[0027] 3.3 Information Sharing Mechanism Agents share information through an attention mechanism, specifically: [C^t = \text{Attention}(Q^t, K^t, V^t)] where (Q^t, K^t, V^t) are the query, key, and value vectors, respectively.

[0028] 4. Distributed Workflow 4.1 Workflow Generation After an operation ticket is generated, the system transforms it into a modular workflow, with each module corresponding to a specific operation step or task. The workflow is stored in the form of code blocks or prompts and supports dynamic updates.

[0029] 4.2 Module Optimization Each module supports self-testing and optimization. The optimization goal is: [W^{t+1} = \text{Optimize}(W^t,\text{Distributed}(D^t))] where (W^t) is the current state of the module, and (\text{Distributed}) represents the distributed optimization process.

[0030] 4.3 Global Optimization Global optimization is achieved through reinforcement learning, with the goal of minimizing workflow execution time and error rate.

[0031] 5. Personalized Recommendation and Privacy Protection Module 5.1 Personalized Recommendations By capturing user behaviors (such as editing actions and feedback signals), the system constructs a user preference model (U): [U= {u_{\text{style}}, u_{\text{focus}}, u_{\text{format}}, u_{\text{depth}}}]. Based on reinforcement learning, the system optimizes the generation strategy, with the objective function being: [J(\theta) = \mathbb{E}{X, R, C, U}[r(y{\theta}, y_{\text{user}})]]. 5.2 Privacy Protection The system adopts a multi-layered privacy protection architecture: Local processing: Sensitive data should be processed locally first; Federated learning: Training models through federated learning avoids data leakage; Differential privacy: Adding noise to data to protect individual characteristics.

[0032] The differential privacy formula is: [f_{\text{DP}}(X) = f(X) + \mathcal{N}(0, \sigma^2)] Through the above-mentioned detailed solutions, this invention has achieved comprehensive innovation in dynamic updating of rule knowledge graphs, collaborative decision-making of intelligent agents, workflow optimization, personalized recommendation and privacy protection, significantly improving the intelligence level and practical application value of the power grid operation ticket generation system.

[0033] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A power grid operation order generation system based on multi-agent collaborative work, characterized in that: Comprise: A rule knowledge graph for storing the rules of the power grid operation order and their relationships; A graph neural network for matching rules in the rule knowledge graph; A plurality of agents, each responsible for different sub-tasks, collaborative decision-making between agents through reinforcement learning, and information sharing through attention mechanism, ultimately generating an operation order; a distributed workflow that receives the operation order generated by multiple agents and converts it into a modular workflow, each module corresponding to a specific operation step or task, the workflow is stored in the form of code blocks or prompts, supporting dynamic updates.

2. The multi-agent collaboration based power grid work order generation system of claim 1, wherein: The rule knowledge graph is constructed by the following steps: Step one, extract rules from historical operation orders to construct an initial knowledge graph, the specific way is: first, perform word segmentation, entity recognition and semantic analysis on historical operation orders, then extract key entities through named entity recognition technology, then identify the logical relationship between entities using relationship extraction technology, finally store the extracted entities and relationships in the form of a graph structure to obtain an initial graph; Step two, extract rules from new operation orders and dynamically update the knowledge graph, the dynamic update method is: first, consistency detection is performed on the new rules and the existing rules in the graph to avoid rule conflicts, then for new rules, a deep learning model is used to generate semantic representations of the rules, and the vector similarity is calculated to integrate them into the existing graph, finally the graph embedding technology is used to optimize the representation of the graph to ensure the availability and efficiency of the rule knowledge graph; Step three, dynamically expand the rule graph, the specific way is: first, discover potential rules through unsupervised learning methods, and finally dynamically adjust the priority according to the frequency and importance of the rules in historical data.

3. The multi-agent collaboration based power grid operation order generation system according to claim 1 or 2, characterized in that: The specific steps of the graph neural network for rule matching are as follows: Step 1, convert the input entities (X) in the operation order and the rule knowledge graph (G) into vector representations, generate node feature matrix (H) and adjacency matrix (A) through graph embedding technology; Step 2, update the node features through the graph convolution network to capture high-order relationships; Step 3, calculate the matching degree of input entities and rules through the matching score function; Wherein, in the above matching process, the attention mechanism based on Transformer is also introduced to enhance the context awareness ability of rule matching; at the same time, the rule matching and rule priority prediction tasks are optimized to improve the generalization ability of the model.

4. The multi-agent collaboration based power grid operation order generation system according to claim 1 or 2, characterized in that: The plurality of agents includes: Rule extraction agent: responsible for extracting rules from historical operation orders; Rule optimization agent: for dynamically adjusting the rule knowledge graph; Operation order generation agent: for generating operation orders according to the rule knowledge graph; Workflow generation agent: for converting operation orders into modular workflows; Module optimization agent: for self-optimizing workflow modules; Global optimization agent: for global optimization of the entire workflow.

5. The multi-agent collaboration based power grid operation order generation system according to claim 1 or 2, characterized in that: The collaborative decision-making of the plurality of agents is as follows: Step 4, the state of each agent is represented as ( s_i^t ), and the global state is ( S^t = {s_1^t, s_2^t,..., s_n^t} ); Step 5, the strategy of each agent is obtained as ( \pi_i ), which is optimized through deep reinforcement learning and a global reward function; Wherein, the global reward function ( R(S^t, a^t) ) combines the accuracy, safety and efficiency of the operation ticket generation.

6. The multi-agent collaboration based power grid work order generation system of claim 1 or 2, wherein: The distributed workflow adopts dynamic updating through module optimization and global optimization.

7. The multi-agent collaboration based power grid operation order generation system according to claim 1 or 2, characterized in that: It also includes a personalized recommendation and privacy protection module, which captures user behavior to build a user preference model, and through local processing of sensitive data first and training the model through federated learning, it avoids data leakage and adds noise to the data to protect individual features.

Images