An entity recognition model and data through system coupling a large model and a NER layer

Abstract

This invention relates to an entity recognition model and data integration system that couples a large-scale model with a Network Entity Recognition (NER) layer. The system includes a backend integrated entity recognition model for recognizing named entities and a database for storing knowledge graphs of upstream and downstream relationships between entities. The application program of the data integration system communicates with the frontend and backend via a data interface. The entity recognition model includes a large-scale model module that receives main and distribution network data from the power system, recognizes the text within the main and distribution network data, replaces it with entity tags, and outputs a tag sequence. The NER layer identifies entities within the power system based on the correlation between entity tags and related entities. The data interface and knowledge graph ensure data exchange and sharing between entities, offering advantages such as high data integration efficiency and high data utilization efficiency. The entity recognition model boasts high data processing accuracy and high entity recognition accuracy, and can be widely applied in the fields of power system main and distribution network data management and application.

Description

Technical Field

[0001] This invention belongs to the field of intelligent power system and data fusion, and specifically relates to an entity recognition model and data connection system that couples a large model with a NER layer. Background Technology

[0002] With the construction of new power systems, information from various dispatchable objects such as distributed power sources is widely integrated into dispatch automation systems. The voltage levels of power systems now comprehensively cover 220kV to 220V, and the number of distribution network equipment and measurement points is growing exponentially. However, at the current stage, with the access of various data sources, the matching of measurement information and the verification of the topology connections of main and distribution network equipment still rely heavily on manpower, placing a heavy burden on grassroots work in terms of basic data governance.

[0003] Therefore, there is an urgent need to explore deeper relationships beyond simple ID matching in order to improve the completeness of data matching and the accuracy of topological connections. Summary of the Invention

[0004] To address the problems existing in the background technology, this invention provides an entity recognition model and data integration system that couples a large model with a NER layer, which has the advantages of high data processing accuracy, high entity recognition accuracy, and high data integration efficiency.

[0005] To address the aforementioned technical issues, this invention employs an entity recognition model that couples a large model with a NER layer. The large model module receives main and distribution network data from the power system, identifies the text within the main and distribution network data, replaces it with entity tags, and outputs a tag sequence. The NER layer identifies entities within the power system based on the correlation between entity tags and related entities.

[0006] Furthermore, the NER layer includes: based on The function converts the relevance between entity labels and related entities into a probability distribution and outputs the related entity with the highest probability. The function is: , in, Indicates the number of related entities. This represents the mapping degree between the entity label and the i-th related entity.

[0007] Furthermore, the entity recognition model includes an optimization and fine-tuning strategy: it measures the difference between the recognized entity and the real entity based on the cross-entropy loss function, and then fine-tunes the parameters within the NER layer based on this difference. The cross-entropy loss function is: , in, The predicted proportion of each type of identified entity. The actual proportion of various real entities, The number of entity categories. The predicted proportion of the i-th type of entity. This represents the actual proportion of the i-th type of entity.

[0008] Furthermore, the unstructured text information in the main distribution network data includes scheduling logs, operation and maintenance records, inspection reports, and equipment nameplate information, while the structured data information in the main distribution network data includes SCADA measurement data, equipment ledger data, and power grid topology connection data.

[0009] Furthermore, the preprocessing of the main distribution network data includes: removing duplicate, inconsistent, or incomplete information from the text information; filtering outliers in the data information; filling in missing values ​​in the data information; and standardizing the format and dimensions of the data information.

[0010] An entity data integration system includes a backend integrated entity recognition model for identifying named entities and a database for storing knowledge graphs of upstream and downstream relationships between entities. The application program of the data integration system communicates with the frontend and backend through a data interface.

[0011] Furthermore, the entity data integration system uses the Django web application framework.

[0012] Furthermore, entity information of the identified entities, location relationship information between entities, connection relationship information between entities, and attribute information of entities are obtained from the main distribution network data to construct a knowledge graph.

[0013] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention constructs an entity recognition model by coupling a pre-trained large language model (large model module) with a NER layer. By combining the text recognition capabilities of the large language model and the entity recognition capabilities of the NER layer, it can more accurately identify entities in the main distribution network data.

[0014] By constructing a knowledge graph within the data integration system, the hierarchical relationships between different entities within the power system can be more comprehensively displayed, improving the efficiency of integrating and managing entities and corresponding data within the main and distribution network.

[0015] By setting up data interfaces within the data integration system, integrated main and distribution network data can be seamlessly transmitted, shared, and exchanged between different systems, platforms, or components, thereby improving the utilization efficiency and application flexibility of main and distribution network data. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating an entity recognition model and data integration system that couples a large model with a NER layer in an embodiment of the present invention. Figure 2 This is a schematic diagram illustrating the process of preprocessing main and distribution network data in an embodiment of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0018] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0019] The present invention will be further described below with reference to specific embodiments, but these are not intended to limit the scope of the invention.

[0020] An entity recognition model that couples a large model with a NER layer, such as Figure 1 and Figure 2 As shown, this includes acquiring and preprocessing the main and distribution network data of the power system. Unstructured text information in the main and distribution network data includes dispatch logs, operation and maintenance records, inspection reports, and equipment nameplate information. Structured data in the main and distribution network data includes SCADA measurement data, equipment ledger data, and power grid topology connection data.

[0021] Preprocessing of main and distribution network data includes: removing duplicate, inconsistent, or incomplete information from textual data; filtering outliers; filling in missing values; and standardizing the format and units of the data. Specifically, removing duplicate, inconsistent, or incomplete information from textual data ensures the contextual consistency and accuracy of the description. The removal process is expressed by the following formula: , In the formula, For the cleaned dataset, For the original dataset, For each text record in the text information, As a measure of text quality, This is the threshold for text quality.

[0022] Filtering outliers in data includes identifying and filtering noise and outliers to improve the reliability and usability of the data. Common filtering methods include mean filtering and median filtering. Median filtering is effective for various noise types; here, we use median filtering, and the formula is as follows: , In the formula, For the first Data points, For window size, This is a median-taking operation.

[0023] Imputing missing values ​​in data includes: for missing values ​​in the data, mean imputation is used to ensure data integrity and continuity. The mean imputation formula is as follows: , In the formula, For the first The result of filling in missing values, For the dataset after removing missing values, This is for the operation of taking the average.

[0024] Standardizing data formats includes converting data information into a standardized format. (4) In the formula, For the first The transformed data, For the first One set of raw data, This is a conversion function.

[0025] Standardizing data information involves: standardizing data information to give it a unified dimension and scope, facilitating comparison and analysis between different data sources. Data standardization can be expressed by the following formula: (5) In the formula, For the first A standardized set of data, For the first One set of raw data, The mean of the data. denoted as the standard deviation of the data.

[0026] The large model module receives preprocessed main and distribution network data, identifies the text within the main and distribution network data, replaces it with entity labels, and outputs a label sequence. The large model module is a pre-trained large language model. Specifically, based on the specific task of named entity recognition in the power system's main and distribution network data, a pre-trained large language model, such as GPT, BERT, or RoBERTa, is selected. This large language model has undergone unsupervised learning on large-scale text data, learning rich language representations, enabling it to accurately identify text within the main and distribution network data, and more accurately convert the text into entity labels and output the corresponding label sequence.

[0027] Large language models are typically stored in specific model repositories, such as Hugging Face's model repository, as weight matrices and configuration files. When a large model module loads a large language model, it simply uses the appropriate library (such as Hugging Face's Transformers library) to load these parameters into the large model module.

[0028] The NER layer identifies entities within the power system based on the degree of correlation between entity tags and related entities. One embodiment of this application describes an NER layer consisting of a fully connected layer or a CRF (Conditional Random Field) layer, with entities including power plants, substations, transmission lines, etc.

[0029] The NER layer includes based on The function converts the relevance between entity labels and related entities into a probability distribution and outputs the related entity with the highest probability. The function is: , in, Indicates the number of related entities. This represents the mapping degree between the entity label and the i-th related entity.

[0030] The large model module adds the defined NER layer to the large language model through the API (Application Programming Interface), realizing the coupling of data between the large language model and the NER layer, forming an end-to-end entity recognition model.

[0031] To improve the accuracy of entity recognition models, the parameters within the NER layer of the entity recognition model are fine-tuned using training set data. During fine-tuning, the cross-entropy loss function is used to measure the difference between the entities recognized by the model and the real entities. The formula for the cross-entropy loss function is shown below: , in, The predicted proportion of each type of identified entity. The actual proportion of various real entities, The number of entity categories. The predicted proportion of the i-th type of entity. This represents the actual proportion of the i-th type of entity.

[0032] After fine-tuning the entity recognition model, validation set data is used to evaluate the model's performance. Standard NER evaluation metrics—precision, recall, and F1 score—are typically used to evaluate the model's performance in recognizing various entity types. The accuracy formula is shown below: , in, The number of samples that the model correctly predicts as positive. The number of samples that the model incorrectly predicted as positive.

[0033] The recall rates are shown below: , in, This represents the number of positive class samples that the model failed to predict correctly.

[0034] The F1 scores are shown below: , The F1 score is a comprehensive indicator for evaluating model performance, taking into account both accuracy and recall. Its value ranges from 0 to 1, with a higher value indicating better model performance.

[0035] After multiple fine-tuning and performance evaluations, the model with the best performance was selected as the final entity recognition model.

[0036] An entity data integration system includes a backend integrated entity recognition model for identifying named entities and a database for storing knowledge graphs of upstream and downstream relationships between entities. The application of the data integration system communicates with the frontend and backend through data interfaces.

[0037] The construction of the knowledge graph includes: first, identifying the entities contained in the main distribution network data based on an entity recognition model; then, extracting the attribute information of these entities from the main distribution network data and classifying and organizing the attributes of different entities to form standardized data; next, obtaining the positional and connection relationships between entities to further define the associations between entities and establish hierarchical relationships between entities to represent parent-child or upstream-downstream relationships. Then, using the Protégé ontology building tool, an ontology model of the knowledge graph is built. Based on the standardized data of entities and application requirements, entity categories are determined. Based on the ontology model and standardized data, specific entities, attributes, and relationships are populated, the ontology model is instantiated, and the knowledge graph of the power system's main distribution network is constructed.

[0038] The specific process of establishing the data integration system is as follows: Django, an open-source web application framework, is selected as the front-end page development framework. The trained entity recognition model is integrated into the framework, and the constructed knowledge graph is stored in the database of the Django project.

[0039] Create and edit HTML template files in a Django project to design and customize the front-end pages of the data integration system.

[0040] Deploy a Django application on the server, design the data interface of the Django application, integrate and transform the received entity recognition model feedback data, database feedback data, and front-end page feedback data to realize a complete power system main and distribution network data interconnection platform.

[0041] The above are merely preferred embodiments of the present invention and are not intended to limit the implementation methods and protection scope of the present invention. Those skilled in the art should recognize that any equivalent substitutions and obvious changes made based on the content of this specification should be included within the protection scope of the present invention.

Claims

1. An entity recognition model that couples a large model with a NER layer, characterized in that, The large model module receives the main and distribution network data of the power system, identifies the text in the main and distribution network data, replaces it with entity labels, and outputs a label sequence. The NER layer identifies entities in the power system based on the degree of correlation between entity labels and related entities.

2. The power system entity recognition model coupled with a large language model and a NER layer according to claim 1, characterized in that, The NER layer includes: based on The function converts the relevance between entity labels and related entities into a probability distribution and outputs the related entity with the highest probability. The function is: ， in, Indicates the number of related entities. This represents the mapping degree between the entity label and the i-th related entity.

3. The entity recognition model coupled with a large model and a NER layer according to claim 1, characterized in that, The entity recognition model includes an optimization and fine-tuning strategy: it measures the difference between the recognized entity and the real entity based on the cross-entropy loss function, and then fine-tunes the parameters within the NER layer based on this difference. The cross-entropy loss function is: ， in, The predicted proportion of each type of identified entity. The actual proportion of various real entities, The number of entity categories. The predicted proportion of the i-th type of entity. This represents the actual proportion of the i-th type of entity.

4. The entity recognition model coupled with a large model and a NER layer according to claim 1, characterized in that, The unstructured text information in the main distribution network data includes dispatch logs, operation and maintenance records, inspection reports, and equipment nameplate information. The structured data information in the main distribution network data includes SCADA measurement data, equipment ledger data, and power grid topology connection data.

5. The entity recognition model coupled with a large model and a NER layer according to claim 4, characterized in that, Preprocessing the main distribution network data includes: removing duplicate, inconsistent, or incomplete information from the text information; Filter outliers in the data, fill in missing values ​​in the data, and standardize the format and dimensions of the data.

6. An entity data integration system, comprising an entity recognition model coupled with a large model and a NER layer according to any one of claims 1-5, characterized in that, The system includes a backend-integrated entity recognition model for identifying named entities and a database for storing knowledge graphs of upstream and downstream relationships between entities. The application of the data connectivity system communicates with the frontend and backend through data interfaces.

7. The entity data interconnection system according to claim 6, characterized in that, The entity data integration system uses the Django web application framework.

8. The entity data interconnection system according to claim 6, characterized in that, Constructing the knowledge graph includes: obtaining entity information of the identified entities, location relationship information between entities, connection relationship information between entities, and attribute information of entities from the main distribution network data to construct the knowledge graph.

Images