Intelligent response and reasoning method based on fusion of scheduling knowledge base and semantic model

Abstract

The invention discloses an intelligent response and reasoning method based on fusion of a scheduling knowledge base and a semantic model, and the method comprises the steps: constructing a data structure based on a graph, organizing the knowledge in the field of power scheduling into structured information, and storing the structured information in a graph database; combining domain knowledge with a general semantic model by adopting a deep learning technology; the system can apply the output of a BERT model to a knowledge base query and reasoning module, and in multiple rounds of dialogues, the system continuously updates the dialogue state by using LSTM according to historical dialogues and context information, and generates a response better meeting user requirements in combination with domain knowledge; and the system predicts and recommends the optimal operation decision, and optimizes the scheduling decision process through reinforcement learning. According to the method, technical terms and complex situations in the field of power dispatching can be more accurately understood and processed, and the accuracy and timeliness of decision making are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of power dispatching, and in particular to an intelligent response and reasoning method based on the fusion of a dispatching knowledge base and a semantic model. Background Technology

[0002] In existing technologies, power dispatching systems generally rely on traditional dispatching methods and often lack intelligent and automated support. Although some power dispatching systems have implemented data analysis and model reasoning, they often rely on static rules and lack dynamic decision support based on real-time data and historical experience. In addition, although some systems have document generation functions, most of them are merely static template filling, lacking flexibility and intelligent recommendation functions.

[0003] Chinese patent CN118261164A discloses a method and system for intelligent response in power dispatching based on a large language model. The method includes the following steps: constructing a large language model for power dispatching; selecting the ChatGLM 6B language model; constructing a power dispatching knowledge graph; power dispatchers inputting voice or text-based questions; the large language model calling the power dispatching knowledge graph to generate specific commands or answers; and feeding back the specific commands or answers to relevant personnel via a display screen and / or audio. This invention uses ChatGLM 6B as the large language model and incorporates a power dispatching knowledge graph into it, enabling intelligent questioning, answering, and timely responses in the field of power dispatching. This is beneficial for improving the work efficiency of power dispatchers and filling knowledge gaps. However, the construction and maintenance of the power dispatching knowledge graph in this patent may require a large amount of manual operation, and the graph data may be outdated or incomplete, thus affecting the accuracy of intelligent responses.

[0004] Chinese Patent CN119558488A discloses an optimization method and system based on a large-scale power dispatching model, belonging to the field of power dispatching technology. The method includes the following steps: acquiring historical power dispatching knowledge; dividing historical power dispatching knowledge into regional knowledge and global knowledge based on regional characteristics; constructing a first teacher model based on the differences between global knowledge and regional knowledge; constructing a student model based on the lowest fluctuation cost; acquiring regional dispatching correlation based on historical power dispatching knowledge; constructing a second teacher model based on regional dispatching correlation; associating the first teacher model, second teacher model, and student model based on knowledge distillation to construct a large-scale power dispatching model; and outputting an optimization strategy based on the large-scale power dispatching model and current power data. The beneficial effect of this application is that the power dispatching strategy output balances relevance and efficiency. However, while the knowledge distillation method used in this patent can improve model performance, the process may be complex, especially in the case of multi-model association, potentially increasing computational overhead and system complexity. Therefore, there is an urgent need for an intelligent response and reasoning method based on the fusion of a dispatching knowledge base and a semantic model. Summary of the Invention

[0005] Purpose of the invention: The purpose of this invention is to provide an intelligent response and reasoning method based on the fusion of a scheduling knowledge base and a semantic model.

[0006] Technical solution: The intelligent response and reasoning method based on the fusion of scheduling knowledge base and semantic model described in this invention is characterized by the following steps:

[0007] (1) Knowledge is automatically extracted from literature in the field of power dispatching through natural language processing technology, and a graph-based data structure is constructed to organize the knowledge in the field of power dispatching into structured information and store it in a graph database;

[0008] (2) The system fine-tunes the pre-trained BERT model so that it can understand the professional terminology and context in the field of power dispatching. It uses deep learning technology to combine domain knowledge with a general semantic model.

[0009] (3) After fine-tuning the domain semantic model, the system can apply the output of the BERT model to the knowledge base query and reasoning module. The module relies on the constructed scheduling knowledge graph and semantic model. When the user makes a query, it extracts relevant entities and relationships based on the input natural language and performs accurate knowledge retrieval or reasoning from the knowledge base.

[0010] (4) In multi-turn dialogues, the system will continuously update the dialogue state using LSTM based on historical dialogues and contextual information, and combine domain knowledge to generate responses that better meet user needs.

[0011] (5) Based on historical dispatch data and real-time power system information, the system predicts and recommends the best operational decisions and optimizes the dispatch decision-making process through reinforcement learning.

[0012] Further, step (1) includes first preprocessing documents, standards, operating procedures, historical records and real-time data in the field of power dispatching, which usually includes word segmentation and removal of stop words. The system automatically identifies key entities in the documents through named entity recognition technology. Then, the logical relationships between these entities are automatically generated using relation extraction technology. Finally, by constructing a knowledge graph, these entities and their relationships are stored in a graph database in a graphical structure.

[0013] Furthermore, step (2) includes BERT pre-training and model fine-tuning.

[0014] Furthermore, the pre-training of BERT includes MLM and NSP; in the MLM task, the model randomly masks some words in the input text and is required to predict these masked words, thereby learning the contextual relationships between words; the NSP task requires the model to determine whether two given sentences are adjacent, thereby learning the relationships between sentences.

[0015] Furthermore, the model fine-tuning includes:

[0016] Prepare domain corpus and labeled data: The documents, operating procedures and historical records of power dispatch constitute the domain corpus, and the labeled data includes annotations of domain terms and relationships;

[0017] Task definition and model input: During fine-tuning, BERT's input includes not only word vectors but also task-specific labels;

[0018] The fine-tuning model training process involves inputting labeled data from the power dispatching domain into the BERT model and training it on a specified task. During training, the goal is to minimize the model's loss function. During fine-tuning, the goal is to enable the BERT model to accurately predict tasks. In optimization, gradient descent is used to adjust the model parameters, with the optimization objective being to minimize the loss function. In BERT fine-tuning, the loss function is typically defined as:

[0019]

[0020] Wherein, P(y i |x i ;θ) represents the model's response to the input word x i Predict label y i The probability is given by θ, where θ is the model parameter; the parameters of the BERT model are adjusted using the Adam optimizer, and the parameter update formula is:

[0021]

[0022] Where, m t and v t Let represent the first-order moment estimate and the second-order moment estimate of the gradient, respectively, and η be the learning rate. It is a small constant.

[0023] Further, step (3) includes:

[0024] The system uses the BERT model to parse user queries, transform them into structured information, and extract keywords and entities. Based on this, the system uses a graph query engine to search for relevant entities and their relationships in the knowledge base.

[0025] In complex scenarios, graph attention networks are used to model the relationships between power plants:

[0026]

[0027] Where: h i ' is the new state of node i, α ij is the attention weight of node i to its neighbor node j, and W is the learned transformation matrix;

[0028] During the reasoning process, the reasoning model is adjusted and the system decision is optimized by calculating the loss of the reasoning task;

[0029]

[0030] Where θ is the parameter of the inference model, P(z) i |x i ,y i ;θ) represents the entity information and rule reasoning results in a given knowledge base.

[0031] Furthermore, step (4) includes: the system continuously updates the dialogue state using LSTM based on historical dialogue and contextual information, and generates a response that better meets user needs by combining domain knowledge. The LSTM state update formula is as follows:

[0032]

[0033] Where: x t h is the current input. t−1 This is the memory state from the previous step.

[0034] Furthermore, step (5) includes collaborative filtering algorithms, reinforcement learning optimization decision-making, and intelligent decision support.

[0035] Furthermore, the collaborative filtering algorithm includes:

[0036]

[0037] Where, r ui This represents user u's rating of item i. Let be the average rating of user u, and sim(u,v) be the similarity between user u and user v.

[0038] Furthermore, the reinforcement learning optimization decision objective is to maximize the cumulative reward, and its formula is:

[0039]

[0040] Among them, R t For the cumulative reward starting from time step t, r k The immediate reward at time step k is γ, which is a discount factor to control the impact of long-term rewards.

[0041] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages: By combining advanced technologies such as natural language processing, knowledge graphs, deep learning and reinforcement learning, the present invention intelligently constructs a knowledge base and graph in the field of power dispatching, improving the efficiency of knowledge organization and retrieval, and enabling real-time dynamic support for decision-making; through an optimized domain semantic model, the system can more accurately understand and process professional terms and complex situations in the field of power dispatching, significantly improving the accuracy and timeliness of decision-making. Attached Figure Description

[0042] Figure 1 This is a flowchart of the intelligent response and reasoning method based on the fusion of scheduling knowledge base and semantic model described in this invention. Detailed Implementation

[0043] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0044] This invention proposes an intelligent decision support method for power dispatching based on natural language processing, knowledge graphs, deep learning, and reinforcement learning. This method automatically constructs a dispatching knowledge base and graph, optimizes the domain semantic model, and combines intelligent recommendation and decision support to help dispatchers improve decision-making efficiency. Simultaneously, utilizing natural language generation technology, the system can automatically generate standard-compliant power dispatching reports, significantly improving the intelligence level and work efficiency of power dispatching. The implementation steps of this invention are as follows: Figure 1 As shown:

[0045] (1) Construction of scheduling knowledge base and graph

[0046] This section automatically extracts knowledge from literature in the field of power dispatching using natural language processing technology and constructs a graph-based data structure. In this way, knowledge in the field of power dispatching is organized into structured information and stored in a graph database, thereby providing effective support for subsequent queries and reasoning.

[0047] First, a large amount of documents, standards, operating procedures, historical records, and real-time data in the power dispatching field need to be preprocessed. This typically includes steps such as word segmentation and stop word removal to ensure the text structure is clear and provides effective input data for subsequent entity recognition. The system automatically identifies key entities in the documents using Named Entity Recognition (NER) technology. Then, it automatically generates logical relationships between these entities using relation extraction technology. Finally, by constructing a knowledge graph, these entities and their relationships are stored in a graph database in a graphical structure, enabling efficient retrieval of various types of knowledge through the graph query engine.

[0048] (2) Domain semantic model training and optimization

[0049] In this section, the system fine-tunes a pre-trained BERT model to enable it to understand specialized terminology and context within the power dispatching domain. Deep learning techniques are employed to combine domain knowledge with a general semantic model, enhancing the system's understanding of dispatching issues. This part optimizes the semantic model through domain adaptation techniques, achieving more accurate language understanding and processing.

[0050] (2.1) Pre-training of BERT

[0051] BERT's pre-training process includes two main tasks: Masked Language Model (MLM) and Next Sentence Prediction (NSP). In the MLM task, the model randomly masks some words in the input text and is required to predict these masked words, thereby learning the contextual relationships between words. The NSP task requires the model to determine whether two given sentences are adjacent, thus learning the relationships between sentences.

[0052] The context representation formula for BERT is as follows:

[0053]

[0054] Here, H represents the contextual representation of the output, which contains semantic information about the input vocabulary.

[0055] After pre-training on these two tasks, the BERT model has a basic understanding of language, but it still needs to be fine-tuned on data in specific domains to adapt to specific tasks.

[0056] (2.2) Model fine-tuning

[0057] Documents in the power dispatching domain contain a large number of specific terms. While BERT, after pre-training, can understand general language structures, its ability to process these specific terms may be insufficient. Therefore, the fine-tuning process uses labeled data from the power dispatching domain to adjust the model's parameters, enabling it to better understand and process these domain-specific terms and grammatical structures.

[0058] The fine-tuning process includes the following steps:

[0059] Prepare domain corpus and labeled data: The domain corpus consists of documents, operating procedures, historical records, etc. of power dispatching, and the labeled data includes annotations of domain terms and relationships.

[0060] Task definition and model input: During fine-tuning, BERT's input includes not only word vectors but also task-specific labels.

[0061] Fine-tuning the model training process: By inputting labeled data from the power dispatching domain into the BERT model and training it on a specified task, the model's parameters are adjusted to better adapt to the specific task. During training, the goal is to minimize the model's loss function, enabling the model to effectively identify key entities and relationships in the dispatching domain.

[0062] During fine-tuning, the goal is to enable the BERT model to accurately predict tasks. To achieve this, the model parameters are adjusted by optimizing the loss function. During optimization, we use the gradient descent algorithm to adjust the model parameters, with the optimization objective being to minimize the loss function. In BERT fine-tuning, the loss function is typically defined as:

[0063]

[0064] Where P(yi|xi;θ) represents the model's response to the input word x. i Predict label y i The probability is θ, and θ is the parameter of the model.

[0065] The parameters of the BERT model are adjusted using the Adam optimizer. During optimization, the formula for parameter updates is:

[0066]

[0067] Where, m t and v t Let represent the first-order moment estimate and the second-order moment estimate of the gradient, respectively, and η be the learning rate. It is a small constant.

[0068] (3) Knowledge base query and reasoning

[0069] After fine-tuning the domain semantic model, the system can apply the output of the BERT model to the knowledge base query and reasoning module. This module relies on the constructed scheduling knowledge graph and semantic model to extract relevant entities and relationships based on the input natural language when a user submits a query, and then performs accurate knowledge retrieval or reasoning from the knowledge base.

[0070] The system uses the BERT model to parse user queries, transforming them into structured information and extracting keywords and entities. Based on this, the system uses a graph query engine to search for relevant entities and their relationships in the knowledge base.

[0071] In complex scenarios, graph attention networks can be used to model the relationships between power plants.

[0072]

[0073] Where: hi′ is the new state of node i, α ij Let W be the attention weight of node i to its neighbor node j, and W be the learned transformation matrix.

[0074] During the reasoning process, by calculating the loss of the reasoning task, the reasoning model can be further adjusted to optimize system decision-making.

[0075]

[0076] Where θ is the parameter of the reasoning model, and P(zi∣xi,yi;θ) represents the entity information and rule reasoning result in the given knowledge base.

[0077] (4) Multi-turn dialogue and context tracking

[0078] In multi-turn dialogues, the system will continuously update the dialogue state using LSTM based on historical dialogues and contextual information, and combine domain knowledge to generate responses that better meet user needs.

[0079] The LSTM state update formula is as follows:

[0080]

[0081] Where: x t h is the current input. t−1 This is the memory state from the previous step.

[0082] (5) Intelligent recommendation and decision support

[0083] This section provides intelligent recommendations and decision support for dispatchers through collaborative filtering and reinforcement learning algorithms. Based on historical dispatch data and real-time power system information, the system can predict and recommend optimal operational decisions. Through reinforcement learning, the system optimizes the dispatch decision-making process, improves the accuracy and timeliness of power dispatch, and helps dispatchers make more efficient decisions.

[0084] (5.1) Collaborative Filtering Algorithm

[0085] By analyzing the historical decision-making behavior of schedulers and the behavior of other schedulers, the system predicts decision-making solutions for similar scenarios. When a user queries a scheduling problem, the system will make recommendations based on the scheduling records of similar historical problems. The formula is:

[0086]

[0087] Where, r ui This represents user u's rating of item i. Let be the average rating of user u, and sim(u,v) be the similarity between user u and user v.

[0088] (5.2) Enhance learning to optimize decision-making

[0089] Reinforcement learning helps the system optimize its decisions by learning through trial and error in the simulated scheduling process. The power dispatching problem is usually a Markov decision process, in which the system state is real-time data such as grid load and generator output, the actions are the dispatcher's decisions, and the reward is the dispatching result.

[0090] The core objective of reinforcement learning is to maximize cumulative reward, and its formula is:

[0091]

[0092] Among them, R t For the cumulative reward starting from time step t, r k The immediate reward at time step k is γ, which is a discount factor to control the impact of long-term rewards.

[0093] (5.3) Intelligent decision support

[0094] Intelligent decision support systems recommend optimal dispatch strategies to dispatchers by analyzing power system data in real time and combining it with historical dispatching experience. The recommendation system can provide: real-time data-driven decision suggestions; similar scenario recommendations based on historical data; real-time load optimization; and grid stability prediction.

Claims

1. An intelligent response and reasoning method based on the fusion of a scheduling knowledge base and a semantic model, characterized in that, Includes the following steps: (1) Knowledge is automatically extracted from literature in the field of power dispatching through natural language processing technology, and a graph-based data structure is constructed to organize the knowledge in the field of power dispatching into structured information and store it in a graph database; (2) The system fine-tunes the pre-trained BERT model so that it can understand the professional terminology and context in the field of power dispatching. It uses deep learning technology to combine domain knowledge with a general semantic model. (3) After fine-tuning the domain semantic model, the system can apply the output of the BERT model to the knowledge base query and reasoning module. The module relies on the constructed scheduling knowledge graph and semantic model. When the user makes a query, it extracts relevant entities and relationships based on the input natural language and performs accurate knowledge retrieval or reasoning from the knowledge base. (4) In multi-turn dialogues, the system will continuously update the dialogue state using LSTM based on historical dialogues and contextual information, and generate responses that better meet user needs by combining domain knowledge. (5) Based on historical dispatch data and real-time power system information, the system predicts and recommends the best operational decisions and optimizes the dispatch decision-making process through reinforcement learning.

2. The intelligent response and reasoning method based on the fusion of scheduling knowledge base and semantic model according to claim 1, characterized in that, Step (1) includes first preprocessing documents, standards, operating procedures, historical records and real-time data in the field of power dispatching, which usually includes word segmentation and removal of stop words. The system automatically identifies key entities in the documents through named entity recognition technology. Then, the logical relationships between these entities are automatically generated using relation extraction technology. Finally, by constructing a knowledge graph, these entities and their relationships are stored in a graph database in a graphical structure.

3. The intelligent response and reasoning method based on the fusion of scheduling knowledge base and semantic model according to claim 1, characterized in that, Step (2) includes BERT pre-training and model fine-tuning.

4. The intelligent response and reasoning method based on the fusion of scheduling knowledge base and semantic model according to claim 3, characterized in that, The pre-training of BERT includes MLM and NSP. In the MLM task, the model randomly masks some words in the input text and is required to predict these masked words, thereby learning the contextual relationships between words. The NSP task requires the model to determine whether two given sentences are adjacent, thereby learning the relationships between sentences.

5. The intelligent response and reasoning method based on the fusion of scheduling knowledge base and semantic model according to claim 3, characterized in that, The model fine-tuning includes: Prepare domain corpus and labeled data: The documents, operating procedures and historical records of power dispatch constitute the domain corpus, and the labeled data includes annotations of domain terms and relationships; Task definition and model input: During fine-tuning, BERT's input includes not only word vectors but also task-specific labels; The fine-tuning model training process involves inputting labeled data from the power dispatching domain into the BERT model and training it on a specified task. During training, the goal is to minimize the model's loss function. During fine-tuning, the goal is to enable the BERT model to accurately predict tasks. In optimization, gradient descent is used to adjust the model parameters, with the optimization objective being to minimize the loss function. In BERT fine-tuning, the loss function is typically defined as: ， Wherein, P(y i |x i ;θ) represents the model's response to the input word x i Predict label y i The probability is given by θ, where θ is the model parameter; the parameters of the BERT model are adjusted using the Adam optimizer, and the parameter update formula is: ， Where, m t and v t Let represent the first-order moment estimate and the second-order moment estimate of the gradient, respectively, and η be the learning rate. It is a small constant.

6. The intelligent response and reasoning method based on the fusion of scheduling knowledge base and semantic model according to claim 1, characterized in that, Step (3) includes: The system uses the BERT model to parse user queries, transform them into structured information, and extract keywords and entities. Based on this, the system uses a graph query engine to search for relevant entities and their relationships in the knowledge base. In complex scenarios, graph attention networks are used to model the relationships between power plants: ， Where: h i ' is the new state of node i, α ij is the attention weight of node i to its neighbor node j, and W is the learned transformation matrix; During the reasoning process, the reasoning model is adjusted and the system decision is optimized by calculating the loss of the reasoning task; ， Where θ is the parameter of the inference model, P(z) i |x i ,y i ;θ) represents the entity information and rule reasoning results in a given knowledge base.

7. The intelligent response and reasoning method based on the fusion of scheduling knowledge base and semantic model according to claim 1, characterized in that, Step (4) includes: the system continuously updates the dialogue state using LSTM based on historical dialogue and context information, and generates a response that better meets user needs by combining domain knowledge. The LSTM state update formula is as follows: ， Where: x t h is the current input. t−1 This is the memory state from the previous step.

8. The intelligent response and reasoning method based on the fusion of scheduling knowledge base and semantic model according to claim 1, characterized in that, Step (5) includes collaborative filtering algorithm, reinforcement learning optimization decision and intelligent decision support.

9. The intelligent response and reasoning method based on the fusion of scheduling knowledge base and semantic model according to claim 8, characterized in that, The collaborative filtering algorithm includes: ， Where, r ui This represents user u's rating of item i. Let be the average rating of user u, and sim(u,v) be the similarity between user u and user v.

10. The intelligent response and reasoning method based on the fusion of scheduling knowledge base and semantic model according to claim 8, characterized in that, The optimization decision objective of the reinforcement learning is to maximize the cumulative reward, and its formula is: ， Among them, R t For the cumulative reward starting from time step t, r k The immediate reward at time step k is γ, which is a discount factor to control the impact of long-term rewards.

Images