Large model training method and system based on electric power intelligent engineering

Abstract

The invention discloses a large model training method based on electric power intelligent engineering. The method comprises the following steps: acquiring sample data; the sample data comprises first text data and picture data; performing format conversion on the picture data to obtain second text data converted from the picture data; constructing an initial electric power knowledge graph based on a preset model framework; performing association mapping on the first text data and the second text data with entities and relationships in the initial electric power knowledge graph to obtain a fused electric power knowledge graph; and performing fine tuning training on the pre-trained large model based on the fused electric power knowledge graph to obtain a trained target large model. According to the method, the large model can be optimized towards a specific task and an application scene in the field of electric power intelligent engineering on the basis of the ability of general language understanding and the like, so that the trained target large model better meets the actual demand of the electric power industry, and the output result is more targeted and professional; and the application effect in electric power related services is effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent power engineering, and particularly to a large model training method and system based on intelligent power engineering. Background Art

[0002] With the continuous expansion of the complexity and scale of power systems, traditional engineering design and operation and maintenance methods are facing more and more challenges. Intelligent power engineering aims to improve the efficiency, reliability, and sustainability of power systems by introducing advanced information technology and artificial intelligence technology. In the field of bidding for intelligent power engineering, the application of large models has gradually emerged, aiming to provide intelligent assistance for the bidding process, such as assisting the tenderer in formulating tender documents, evaluating tender proposals, answering bidding-related questions, etc.

[0003] However, for the existing large models applied to the bidding of intelligent power engineering, since power bidding involves many complex documents, such as tender documents, bid documents, contract templates, and technical specifications, and power bidding problems have strong professionalism and logic, the model needs to be able to deeply understand the interrelationships among power engineering technical details, business terms, and laws and regulations. However, the existing architectures do not fully consider these characteristics, resulting in the model showing insufficient understanding ability and limited reasoning ability when dealing with complex bidding scenarios, such as multi-round negotiations and comprehensive bid evaluations. As a result, it is difficult for the model to establish a systematic knowledge system, affecting its accurate understanding of problems and the accuracy of answers. Summary of the Invention

[0004] The purpose of the present invention is to solve the problems in the prior art, and a large model training method and system based on intelligent power engineering are proposed.

[0005] To achieve the above purpose, the present invention adopts the following technical solutions:

[0006] A large model training method based on intelligent power engineering includes the following steps:

[0007] Obtain sample data; the sample data includes first text data and picture data;

[0008] Convert the format of the picture data to obtain second text data converted from the picture data;

[0009] Construct an initial power knowledge graph based on a preset model framework;

[0010] Perform association mapping on the first text data and the second text data with the entities and relationships in the initial power knowledge graph respectively to obtain a fused power knowledge graph;

[0011] Fine-tune and train the pre-trained large model based on the fused power knowledge graph to obtain the trained target large model.

[0012] According to a large model training method based on power intelligent engineering provided by the present invention, the first text data includes historical bidding documents, tender documents, winning bid result announcements, bid evaluation reports, industry standard update documents, policy and regulation documents, as well as relevant news and expert interpretation materials.

[0013] According to a large model training method based on power intelligent engineering provided by the present invention, the format conversion of the picture data to obtain the second text data converted from the picture data includes:

[0014] Preprocess the picture data to obtain the target picture data after preprocessing;

[0015] Extract and transform features based on the target picture data to obtain the transformed picture feature vector;

[0016] Extract and transform semantics based on the picture feature vector to obtain the second text data.

[0017] According to a large model training method based on power intelligent engineering provided by the present invention, the construction of the initial power knowledge graph based on the preset model framework includes:

[0018] Obtain the preset model framework and the ontology of the power knowledge graph; the ontology includes entity types, relationship types, and attributes;

[0019] Analyze and extract the first text data and the second text data to obtain the entities, relationships, and attributes in the first text data and the second text data;

[0020] Import the entities, relationships, and attributes in the first text data and the second text data into the preset model framework according to the ontology of the power knowledge graph to obtain the initial power knowledge graph.

[0021] According to a large model training method based on power intelligent engineering provided by the present invention, the association mapping of the first text data and the second text data with the entities and relationships in the initial power knowledge graph to obtain the fused power knowledge graph includes:

[0022] Perform entity recognition based on the first text data and the second text data to obtain the power domain-related entities in the first text data and the second text data;

[0023] Obtain the relationships between the power domain-related entities in the first text data and the second text data;

[0024] Match and map the relationships between the relevant entities in the power domain and the corresponding entities and relationships in the initial power knowledge graph to obtain a fused power knowledge graph.

[0025] According to a large model training method based on power intelligent engineering provided by the present invention, fine-tuning and training a pre-trained large model based on the fused power knowledge graph to obtain a trained target large model, including:

[0026] Based on the fused power knowledge graph, obtain target training data; the target training data includes input features and their corresponding true labels;

[0027] Input the target training data into the pre-trained large model to obtain the predicted value output by the pre-trained large model;

[0028] Based on the loss function of the pre-trained large model, obtain a loss value according to the predicted value;

[0029] Based on the loss value, adjust the optimization objective function of the pre-trained model to obtain a trained target large model.

[0030] According to a large model training method based on power intelligent engineering provided by the present invention, the formula of the loss function is:

[0031] L(θ) = βL ta (θ) + (1 - β)L en (θ);

[0032]

[0033] Among them, θ represents the parameters of the model, β represents the trade-off coefficient; M represents the number of samples; C represents the number of categories of bidding task types; y i,c represents the true label of the i-th sample in the task type c; represents the probability that the model predicts that the i-th sample belongs to the task type c; N represents the number of entity pairs involved in the sample; K represents the number of entity pairs in each sample; e j,k represents the true matching situation of the k-th entity pair in the j-th sample; represents the matching situation of the k-th entity pair predicted by the model in the j-th sample; represents the difference function, which is used to measure the difference between the true matching situation and the predicted matching situation.

[0034] A large model training system based on power intelligent engineering, including:

[0035] An acquisition unit: used to acquire sample data; the sample data includes first text data and picture data;

[0036] A conversion unit: used for performing format conversion on the image data to obtain second text data converted from the image data;

[0037] Construction unit: used to construct the initial power knowledge graph based on the preset model framework;

[0038] A mapping fusion unit: used for associating and mapping the first text data and the second text data with the entities and relationships in the initial electric power knowledge graph, respectively, to obtain a fused electric power knowledge graph;

[0039] Training unit: used to fine-tune the pre-trained large model based on the fused power knowledge graph to obtain the trained target large model.

[0040] The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the above-mentioned large model training method based on smart power engineering when executing the program.

[0041] The present invention also provides a non-transitory computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of the above-mentioned large model training method based on electric power smart engineering are implemented.

[0042] Compared with the prior art, the present invention has the following advantages:

[0043] The present invention provides a large model training method and system based on electric power smart engineering. Sample data is first obtained and the graph data is formatted to obtain first text data and second text data covering all sample data. An initial electric power knowledge graph is then constructed based on a specific model framework. The first text data and the converted second text data are respectively associated and mapped with entities and relationships in the initial electric power knowledge graph to form a fused knowledge graph. Finally, the fused knowledge graph is used to fine-tune the pre-trained large model to obtain a trained target large model, which is convenient for the large model to mine deep semantic information and logical relationships and improve its cognitive ability for electric power business scenarios. The large model can be optimized towards specific tasks and application scenarios in the field of electric power smart engineering based on its existing capabilities such as general language understanding, so that the trained target large model is more in line with the actual needs of the electric power industry, and the output results are more targeted and professional, which effectively improves the application effect in electric power-related businesses and solves the problems of the difficulty of the model in accurately understanding problems and low answer accuracy in the prior art. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] To more clearly illustrate the technical solutions in the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0045] Figure 1 Schematic flowchart of the large model training method based on the power intelligent project provided by the embodiment of the present invention;

[0046] Figure 2 Schematic structural diagram of the large model training system based on the power intelligent project provided by the embodiment of the present invention;

[0047] Figure 3 Schematic structural diagram of the electronic device proposed by the present invention. Detailed implementation manners

[0048] To make the objectives, technical solutions and advantages of the present invention clearer, the following will clearly and completely describe the technical solutions in the present invention with reference to the drawings in the present invention. Obviously, the described embodiments are some but not all of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art without creative efforts based on the embodiments in the present invention belong to the scope of protection of the present invention.

[0049] The following combines Figures 1 to 3 to describe a large model training method and system based on the power intelligent project of the present invention.

[0050] Figure 1 is a schematic flowchart of a large model training method based on the power intelligent project provided by the present invention. As Figure 1 shown, the method includes:

[0051] Step 101, obtaining sample data; the sample data includes first text data and picture data.

[0052] Specifically, the first text data includes historical bidding documents, tender documents, winning bid result announcements, bid evaluation reports, industry standard update documents, policy and regulation documents, as well as relevant news and expert interpretation materials. When obtaining the first text data, relevant power bidding data can be collected through multiple channels such as the official power bidding platform, the enterprise internal bidding database, and the industry association document library.

[0053] The picture data includes pictures of the power engineering site, design drawings of power equipment, power system architecture diagrams, geographical information images of power projects, etc. These picture information can be obtained through the database or uploaded for acquisition.

[0054] By acquiring sample data containing first text data and image data, different forms of information sources are integrated. Text data can convey detailed text information such as descriptions, specifications, and processes related to power engineering, while image data can intuitively display actual scenes such as power equipment and construction sites. The two complement each other, allowing the large model to learn more comprehensive and rich power knowledge and enhance its ability to understand various complex situations in the field of power smart engineering.

[0055] Step 102, convert the format of the image data to obtain the second text data converted from the image data. By constructing the initial power knowledge graph and associating and mapping the text data (the first text data and the converted second text data) with it to form a fused power knowledge graph, the scattered power knowledge is structured and integrated. This makes the relationship between knowledge clearer, making it easier for the large model to mine deep semantic information and logical relationships from it, and improving the cognitive ability of power business scenarios.

[0056] Specifically, they include:

[0057] The image data is preprocessed to obtain preprocessed target image data.

[0058] Image data preprocessing includes operations such as image cropping, scaling, and grayscale processing. The purpose of image cropping is to crop out irrelevant edge parts of the image and highlight key information. The cropped image is then scaled to the same size according to subsequent processing requirements to improve processing efficiency. Grayscale-processed images can reduce the amount of data and simplify the subsequent feature extraction process.

[0059] Feature extraction and transformation are performed based on the target image data to obtain the transformed image feature vector.

[0060] When extracting features, a computer vision algorithm, such as a convolutional neural network, can be used to extract features of the image in the target image data, wherein the convolutional neural network can automatically learn feature information such as texture, shape, color, etc. in the image and convert it into a feature vector of the image. For example, in one embodiment, for an image of power equipment, features such as the shape of key parts of the equipment and the direction of connecting lines can be extracted.

[0061] Semantic extraction and conversion are performed based on the image feature vector to obtain second text data.

[0062] When performing semantic extraction and conversion, feature description algorithms such as scale-invariant feature transformation or accelerated robust features are used to further convert the extracted feature vector into a text description with semantic information, thereby converting the image characteristic vector into second text data.

[0063] Step 103: construct an initial power knowledge graph based on a preset model framework.

[0064] Specifically, it includes:

[0065] Step 1031, obtain a preset model framework and the ontology of the power knowledge graph; the ontology includes entity types, relationship types, and attributes.

[0066] Here, the preset model framework can adopt the GLM3-6B large model architecture, or graph database frameworks such as Neo4j or JanusGraph. These frameworks provide powerful graph data storage, query, and management functions, and can efficiently build and process large-scale knowledge graphs. For example, in an embodiment, power bidding projects, power enterprises, power equipment, power technologies, personnel, etc. are defined as entity types; the tenderer, bidder, equipment included in the project, technologies used by the equipment, personnel participating in the project, etc. are defined as relationship types; the attributes of power bidding projects include project number, project name, tender amount, tender time, etc.; the attributes of power equipment include equipment model, equipment power, production date, etc.

[0067] Step 1032, analyze and extract the first text data and the second text data to obtain entities, relationships, and attributes in the first text data and the second text data.

[0068] For example, in an embodiment, the name of the bidding project is extracted from the power bidding documents as an entity, the tenderer and bidder information as relationships and related attributes, and the equipment parameters as the attributes of the equipment entity, etc.

[0069] Step 1033, import the entities, relationships, and attributes in the first text data and the second text data into the preset model framework according to the ontology of the power knowledge graph to obtain an initial power knowledge graph. By continuously importing and integrating data, gradually enrich and improve the structure and content of the knowledge graph, and form a semantic network containing a large amount of knowledge in the power field.

[0070] Step 104, perform association mapping on the first text data and the second text data respectively with the entities and relationships in the initial power knowledge graph to obtain a fused power knowledge graph; by constructing an initial power knowledge graph and performing association mapping with the text data (the first text data and the converted second text data) to form a fused power knowledge graph, the scattered power knowledge is structurally integrated. This makes the relationships between knowledge clearer and easier for the large model to mine deep semantic information and logical relationships from it, and improves the cognitive ability of the power business scenario.

[0071] Specifically, it includes:

[0072] Step 1041, perform entity recognition based on the first text data and the second text data to obtain power domain-related entities in the first text data and the second text data.

[0073] For the first text data, text cleaning is first performed to remove noise information therein, such as punctuation marks, special symbols, and duplicate text paragraphs, etc. For the second text data, since it is converted from a picture, text error correction is first performed on it to avoid problems that may occur due to errors in picture recognition in the converted text. Mainly using a language model and a vocabulary library in the power field, spelling check and semantic correction are performed on the converted text.

[0074] When performing entity recognition, named entity recognition algorithms in natural language processing technology are used. For example, in power bidding and tendering texts, entities such as project names, enterprise names, and equipment names are recognized, and in text data converted from pictures, entities such as equipment component names and geographical landmark names are recognized. Entities can also be recognized by constructing an entity model, where the entity recognition model is obtained based on sample entities and their corresponding entity recognition label results. The model includes common entity types such as power equipment names, power engineering terms, and concepts related to bidding and tendering.

[0075] Step 1042: Obtain the relationships between power field-related entities in the first text data and the second text data.

[0076] When obtaining the relationships between relevant entities, a relationship extraction model can be constructed, adopting a machine learning method based on feature vectors or a deep learning method. Among them, when adopting a machine learning method such as support vector machine, first extract the feature vectors of the text, and then use the power field text data with entity relationships marked for training to enable the model to learn the mapping relationship between different features and entity relationships. When adopting a deep learning method, such as a relationship extraction model of a convolutional neural network or a recurrent neural network, use the convolutional layer or the recurrent layer to extract local features or sequence features of the text, and then perform relationship classification through the fully connected layer.

[0077] After the relationship extraction model is constructed, the first text data and the second text data containing power field-related entities are input into the trained relationship extraction model to obtain the relationships between entities output by the relationship extraction model.

[0078] Step 1043: Based on the relationships between power field-related entities and the corresponding entities and relationships in the initial power knowledge graph, perform matching and mapping to obtain a fused power knowledge graph.

[0079] First, match the entities related to the power field identified from the first text data and the second text data with the existing entities in the initial power knowledge graph. A matching method based on string similarity can be used, such as the edit distance algorithm, cosine similarity algorithm, etc., to calculate the similarity between entity names. For example, for the identified entity of high-voltage switchgear cabinet, there may be an entity of high-voltage switch cabinet in the initial knowledge graph. By calculating the similarity of their names, it is judged whether they are the same type of entity. If the similarity exceeds a certain threshold, it is considered a match, and the attributes of the newly identified entity (such as the device model, production date, etc. extracted from the text) are updated to the corresponding entity in the knowledge graph.

[0080] For the matching and mapping of entity relationships, the similarity between the relationships extracted from the text data and the relationships in the initial knowledge graph is also calculated. For example, the relationship of device - installed in - project location extracted from the text has a high semantic similarity with the relationship of device - deployment location - project area in the knowledge graph and can be matched. The new relationship instance is added between the corresponding entity nodes according to the structure specification of the knowledge graph to improve the relationship network of the knowledge graph.

[0081] In the process of entity and relationship matching and mapping, some new entities or new relationships not included in the initial knowledge graph may be found. For new entities, according to the ontology definition of the knowledge graph, determine their types, attributes, etc., and add them to the corresponding positions in the knowledge graph. For example, if a new type of power energy storage device entity is identified from the text data, it is added to the knowledge graph as a new type of power device entity, and the technical parameters, manufacturer, etc. of the device obtained from the text are supplemented as attributes. For new relationships, create new relationship type nodes (if it is a completely new relationship type) or add new relationship instances under the existing relationship types in the knowledge graph.

[0082] Step 105, fine-tune and train the pre-trained large model based on the fused power knowledge graph to obtain the trained target large model. Using the fused power knowledge graph to fine-tune and train the pre-trained large model can enable the large model to be optimized towards specific tasks and application scenarios in the field of power intelligent engineering on the basis of its existing general language understanding and other capabilities, making the trained target large model more in line with the actual needs of the power industry, the output results more targeted and professional, and effectively improving the application effect in power-related services.

[0083] Specifically, it includes:

[0084] Based on the integrated power knowledge graph, obtain the target training data; the target training data includes input features and their corresponding true labels; from the integrated power knowledge graph, first determine the key nodes and relationships related to the power intelligent engineering task as the basis for data extraction. For the input features, use knowledge graph embedding techniques, such as algorithms like TransE and TransH, to map the entities and relationships in the knowledge graph to a low-dimensional vector space. For the determination of the true labels, it is based on the actual business results. For example, for the winning bid prediction task, if the bidding enterprise finally wins the bid, the corresponding true label is 1, and if it does not win the bid, it is 0.

[0085] Input the target training data into the pre-trained large model to obtain the predicted values output by the pre-trained large model; the pre-trained large model is based on the open-source GLM3-6B large model and loads its pre-trained weights. Input the input feature vectors into the pre-trained large model, and through the calculations of each layer of the model, from the input layer, after processing through layers such as the multi-head attention mechanism and the feed-forward neural network, finally obtain the predicted values at the output layer.

[0086] Based on the loss function of the pre-trained large model, obtain the loss value according to the predicted values.

[0087] The formula for the loss function is:

[0088] L(θ)=βL ta (θ)+(1-β)L en (θ);

[0089]

[0091] Among them, θ represents the parameters of the model, β represents the trade-off coefficient, 0 < β < 1, which is used to adjust the weights of the task type accuracy loss and the entity matching accuracy loss in the total loss; M represents the number of samples; C represents the number of categories of the bidding task type; y i,c represents the true label of the i-th sample on the task type c (1 if the sample belongs to this task type, otherwise 0); represents the probability that the model predicts the i-th sample belongs to the task type c; N represents the number of entity pairs involved in the sample; K represents the number of entity pairs in each sample; e j,k represents the true matching situation of the k-th entity pair in the j-th sample (matching is 1, not matching is 0); represents the matching situation of the k-th entity pair predicted by the model in the j-th sample; represents the difference function, which is used to measure the difference between the true matching situation and the predicted matching situation. When it is the case, it means there is no difference (i.e., the prediction is correct). When it is the case, Indicates the existence of a difference (i.e., a prediction error).

[0092] Furthermore, for L ta (θ), based on the multi-class cross-entropy loss function, in the bidding task, different task types (such as engineering bidding, material procurement bidding, service bidding, etc.) need to be accurately classified. The multi-class cross-entropy loss function can measure the difference between the probability distribution of the task types predicted by the model and the true label distribution. By minimizing this loss, the model can improve the accuracy of judging the bidding task types. For L en (θ) is a loss function designed for entity matching accuracy. In bidding data, there are many entity information (such as the tenderer, bidder, project name, equipment name, etc.). The accuracy of entity matching is crucial for understanding the bidding content. This loss function calculates the average proportion of entity pair matching errors in each sample. By minimizing this loss, the model can better learn the relationships and matching rules between entities and improve the accuracy of entity matching.

[0093] Adjust the parameters of the optimization objective function of the pre-trained model based on the loss value to obtain the trained target large model.

[0094] The formula for the optimization objective function is:

[0095]

[0096] Among them, P represents the number of samples, represents the predicted value of the model for the l-th sample, t represents the preset threshold, γ represents the trade-off coefficient, represents the loss function, and τ represents the model parameters.

[0097] Specifically, in the power bidding winning bid prediction task, if the threshold t = 0.5 represents the critical probability of winning the bid or not, when the model predicts the winning bid probability u of a bidding enterprise l close to 0.5, it indicates that the model's prediction is in a relatively reasonable range and does not require significant adjustment. represents considering the gradient of the loss function with respect to the model parameters. Even if the predicted value is close to the threshold, but if the gradient of the loss function is large, it means that the model may be unstable or there is room for further optimization under the current parameters. For example, when the gradient of the loss function is large, it means that a small change in the model parameters may cause a large change in the loss value, which may be due to the model not converging to a better local minimum or there being overfitting and other problems. By adding this term, when the predicted value is close to the threshold and the gradient of the loss function is large, the model parameters will also be adjusted to optimize the model performance.

[0098] Furthermore, during the training process, first calculate the predicted values of the model for all samples Then calculate the value of the optimization objective function J(τ). If J(τ) meets the preset conditions (such as being less than a preset value), it is considered that the model has reached a good state and there is no need to optimize the objective function. Otherwise, calculate the gradient of the model parameters according to the optimization objective function, and use optimization algorithms such as gradient descent to update the parameters. For example: where η represents the learning rate. By continuously iteratively updating the parameters, the predicted values of the model gradually approach the threshold and the loss function is also within a reasonable range, so as to obtain the trained target large model, enabling it to give more stable and more compliant prediction results in power intelligent engineering related tasks.

[0099] The present invention relates to the field of power intelligent engineering, and proposes a large model training method based on power intelligent engineering. In the present invention, the proposed large model training method based on power intelligent engineering first obtains sample data, performs format conversion on the graph data to obtain the first text data and the second text data covering all sample data, then constructs an initial power knowledge graph based on a specific model framework, and then maps the first text data and the converted second text data to the entities and relationships in the initial power knowledge graph respectively, finally forming a fused knowledge graph. Finally, use the fused knowledge graph to fine-tune the pre-trained large model to obtain the trained target large model, which is convenient for the large model to mine deep semantic information and logical relationships from it, improve the cognitive ability of power business scenarios, enable the large model to optimize towards specific tasks and application scenarios in the field of power intelligent engineering on the basis of the existing general language understanding and other capabilities, make the trained target large model more in line with the actual needs of the power industry, the output results are more targeted and professional, and effectively improve the application effect in power related services, solving the problems of difficult accurate understanding of the model and low answer accuracy in the prior art.

[0100] Figure 2 is a schematic structural diagram of a look-ahead speed planning device based on a continuous trajectory provided by the present invention, as Figure 2 shown. The device includes: an acquisition unit 10 for acquiring sample data; the sample data includes the first text data and picture data; a conversion unit 20 for performing format conversion on the picture data to obtain the second text data converted from the picture data; a construction unit 30 for constructing an initial power knowledge graph based on a preset model framework; a mapping and fusion unit 40 for respectively associating and mapping the first text data and the second text data with the entities and relationships in the initial power knowledge graph to obtain a fused power knowledge graph; a training unit 50 for fine-tuning the pre-trained large model based on the fused power knowledge graph to obtain the trained target large model.

[0101] Figure 3 is a schematic structural diagram of the electronic device provided by the present invention. As Figure 3 shown, the electronic device may include: a processor 310, a communications interface 320, a memory 330, and a communication bus 340. Among them, the processor 33, the communication interface 320, and the memory 330 communicate with each other through the communication bus 340. The processor 33 may call the logic instructions in the memory 330 to execute a large model training method based on power intelligent engineering. The method includes: obtaining sample data; the sample data includes first text data and picture data; performing format conversion on the picture data to obtain second text data converted from the picture data; constructing an initial power knowledge graph based on a preset model framework; performing association mapping on the first text data and the second text data with the entities and relationships in the initial power knowledge graph respectively to obtain a fused power knowledge graph; and performing fine-tuning training on a pre-trained large model based on the fused power knowledge graph to obtain a trained target large model.

[0102] In addition, when the logic instructions in the above-mentioned memory 330 can be implemented in the form of software function units and sold or used as an independent product, they can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of this technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods in various embodiments of the present invention. The foregoing storage medium includes: various media such as a USB flash drive, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk, or an optical disc that can store program codes.

[0103] On the other hand, the present invention also provides a computer program product. The computer program product includes a computer program stored on a non-transitory computer-readable storage medium. The computer program includes program instructions. When the program instructions are executed by a computer, the computer is capable of executing a large model training method based on electric power intelligent engineering provided by each of the above methods. The method includes: obtaining sample data; the sample data includes first text data and picture data; performing format conversion on the picture data to obtain second text data converted from the picture data; constructing an initial electric power knowledge graph based on a preset model framework; performing association mapping on the first text data and the second text data respectively with entities and relationships in the initial electric power knowledge graph to obtain a fused electric power knowledge graph; and performing fine-tuning training on a pre-trained large model based on the fused electric power knowledge graph to obtain a trained target large model.

[0104] In another aspect, the present invention also provides a non-transitory computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it is implemented to execute a large model training method based on electric power intelligent engineering provided by each of the above. The method includes: obtaining sample data; the sample data includes first text data and picture data; performing format conversion on the picture data to obtain second text data converted from the picture data; constructing an initial electric power knowledge graph based on a preset model framework; performing association mapping on the first text data and the second text data respectively with entities and relationships in the initial electric power knowledge graph to obtain a fused electric power knowledge graph; and performing fine-tuning training on a pre-trained large model based on the fused electric power knowledge graph to obtain a trained target large model.

[0105] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place, or may be distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. A person of ordinary skill in the art can understand and implement it without creative efforts.

[0106] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a necessary general hardware platform, and of course, can also be implemented by hardware. Based on such an understanding, the essence of the above technical solution, or the part that contributes to the prior art, can be embodied in the form of a software product. The computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disc, etc., and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods of each embodiment or some parts of the embodiments.

[0107] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention should cover within the protection scope of the present invention any equivalent replacement or change made according to the technical solution and inventive concept of the present invention.

Claims

1. A large model training method based on power intelligent engineering, characterized in that Including the following steps: Obtain sample data; the sample data includes first text data and picture data; Perform format conversion on the picture data to obtain second text data converted from the picture data; Construct an initial power knowledge graph based on a preset model framework; Perform association mapping on the first text data and the second text data respectively with the entities and relationships in the initial power knowledge graph to obtain a fused power knowledge graph; Fine-tune and train a pre-trained large model based on the fused power knowledge graph to obtain a trained target large model.

2. The large model training method based on the power intelligent project according to claim 1, wherein, The first text data includes historical tender documents, bid documents, winning bid result announcements, bid evaluation reports, industry standard update documents, policy and regulation documents, as well as relevant news and expert interpretation materials.

3. The large model training method based on the electric power intelligent project according to claim 1, wherein, The performing format conversion on the picture data to obtain second text data converted from the picture data includes: Perform preprocessing on the picture data to obtain preprocessed target picture data; Extract and transform features based on the target picture data to obtain transformed picture feature vectors; Perform semantic extraction and transformation based on the picture feature vectors to obtain second text data.

4. The large model training method based on the power intelligent project according to claim 3, characterized in that, The constructing an initial power knowledge graph based on a preset model framework includes: Obtain a preset model framework and the ontology of the power knowledge graph; the ontology includes entity types, relationship types, and attributes; Analyze and extract the first text data and the second text data to obtain entities, relationships, and attributes in the first text data and the second text data; Import the entities, relationships, and attributes in the first text data and the second text data into the preset model framework according to the ontology of the power knowledge graph to obtain an initial power knowledge graph.

5. The large model training method based on the intelligent power engineering according to claim 1, wherein, The performing association mapping on the first text data and the second text data respectively with the entities and relationships in the initial power knowledge graph to obtain a fused power knowledge graph includes: Perform entity recognition based on the first text data and the second text data to obtain power domain-related entities in the first text data and the second text data; Obtain the relationships between the power domain-related entities in the first text data and the second text data; Perform matching and mapping based on the relationships between the power domain-related entities and the corresponding entities and relationships in the initial power knowledge graph to obtain a fused power knowledge graph.

6. The large model training method based on the intelligent power engineering according to claim 1, wherein, The fine-tuning and training a pre-trained large model based on the fused power knowledge graph to obtain a trained target large model includes: Based on the fused power knowledge graph, obtain target training data; the target training data includes input features and their corresponding true labels; Input the target training data into the pre-trained large model to obtain predicted values output by the pre-trained large model; Obtain a loss value based on the loss function of the pre-trained large model according to the predicted values; Adjust the optimization objective function of the pre-trained model based on the loss value to obtain a trained target large model.

7. The large model training method based on the power intelligent project according to claim 1, characterized in that, The formula of the loss function is: L(θ) = βL ta (θ) + (1 - β)L en (θ); Among them, θ represents the parameters of the model, and β represents the trade-off coefficient; M represents the number of samples; C represents the number of categories of the bidding task type; y i,c represents the true label of the i-th sample on the task type c; represents the probability that the model predicts the i-th sample belongs to the task type c; N represents the number of entity pairs involved in the samples; K represents the number of entity pairs in each sample; e j,k represents the true matching situation of the k-th entity pair in the j-th sample; represents the matching situation of the k-th entity pair in the j-th sample predicted by the model; represents the difference function, which is used to measure the difference between the true matching situation and the predicted matching situation.

8. A large model training system based on the power intelligent project, characterized in that, Including: an acquisition unit: used to obtain sample data; the sample data includes first text data and picture data; Conversion unit: used to perform format conversion on the picture data to obtain second text data obtained by converting the picture data; Construction unit: used to construct an initial power knowledge graph based on a preset model framework; Mapping and fusion unit: used to perform associated mapping on the first text data and the second text data with the entities and relationships in the initial power knowledge graph respectively to obtain a fused power knowledge graph; Training unit: used to fine-tune and train a pre-trained large model based on the fused power knowledge graph to obtain a trained target large model.

9. An electronic device, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method for training a large model based on a power intelligent project according to any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the method for training a large model based on a power intelligent project according to any one of claims 1 to 7.

Images