Pumped storage power station construction management system based on large language model

Abstract

The invention belongs to the technical field of energy storage power station construction management, and discloses a pumped storage power station construction management system based on a large language model. A pumped storage power station big language model data database is generated through the model construction module, a big language model is trained, evaluated and finely adjusted, a big language model oriented to pumped storage power station construction management is generated, input questions are processed through the intelligent AI assistant module, and answers, drawings and schemes are generated based on the database. A model is adjusted through a model correction module, a generated scheme and the like are corrected, the generated scheme and the like are stored through a resource library module, an accurate version is iterated, access control over a user is achieved through an authorization management module, and functions of a web end are converted through a mobile end module. Question and answer of professional knowledge of the pumped storage power station can be achieved, the management difficulty of managers is reduced, and schemes and drawings are generated according to user requirements.

Description

Technical Field

[0001] The invention belongs to the technical field of construction management of energy storage power stations, and in particular relates to a construction management system of a pumped storage power station based on a large language model. Background Art

[0002] The Big Language Model is a language processing model based on deep learning. It has powerful computing power and integrates multiple modal information. It can process multiple different types of data at the same time, such as text, images, audio, etc., to achieve cross-modal understanding and generation. It has a wide range of applications in natural language processing, computer vision, speech recognition and other fields. It is often used in voice assistants, intelligent customer service, machine translation, image description generation and other tasks, providing people with more intelligent and convenient services and experiences.

[0003] In addition to the above advantages, large language models also have obvious defects, mainly including the following two points: (1) It is unable to give relatively accurate answers to related professional fields; (2) It is unable to generate drawings based on the characteristics of pumped storage power plants; (3) The results are highly dependent on the training corpus. If the training data is insufficient, the generated results will be inaccurate. At present, there are many large language models on the market. Such large language models can answer users' routine questions, but the response results related to pumped storage power plants are almost meaningless. Summary of the invention

[0004] The present invention aims to address the deficiencies in the prior art and provide a pumped-storage power station construction management system based on a large language model that integrates technologies such as professional data integration, data preprocessing, intelligent AI, a large language model and natural language processing, which can realize question and answer of professional knowledge of pumped-storage power stations, reduce the management difficulty of managers, and generate plans and drawings according to user needs.

[0005] In order to achieve the above technical objectives, the present invention adopts the following technical solutions:

[0006] A pumped storage power station construction management system based on a large language model, including a model construction module, an intelligent AI assistant module, a model correction module, a resource library module and an authorization management module;

[0007] The model building module is used to collect data and generate a large language model data database for pumped storage power stations. Based on the database, the Transformer neural network model architecture is selected, and Zero-shot Learning, Few-shot Learning and RLHF technologies are combined to generate a large language model for pumped storage power station construction management.

[0008] The intelligent AI assistant module is used to connect the model building module and process the input questions to obtain input sequences. It uses the self-attention mechanism to extract the correlation between the input sequences, matches the answers that meet the requirements based on the correlation, and generates drawings and plans based on the large language model data database of pumped storage power stations.

[0009] The model correction module is used to make corrections to the large language model for pumped storage power station construction management generated by the model construction module, and to make corrections to the answers, drawings and plans generated by the intelligent AI assistant module by industry experts;

[0010] The resource library module is used to store the answers, drawings and plans generated by the intelligent AI assistant module. The user communicates with the intelligent AI assistant module based on the content stored in the resource library. The system iterates the stored content to generate an accurate version according to the user's needs and stores the accurate version again.

[0011] The authorization management module is used to enter user attributes into the system through dialogue between the user and the intelligent AI assistant module, and to perform access control on the user based on the user attributes.

[0012] Preferably, the specific contents executed by the model building module include:

[0013] A1. Basic data support: Collect basic data of pumped storage power stations to form a large language model data database of pumped storage power stations;

[0014] A2. Data preprocessing: Data preprocessing is performed on the large language model database of pumped storage power stations. Database deduplication and semantic similarity calculation methods are used to clean the database, delete duplicate or irrelevant content, and format the preprocessed database.

[0015] A3. Model generation: Select the Transformer neural network model architecture, and use its ability to process sequence data and capture long-term dependencies to select the number of model parameters, hidden units, and hyperparameters. Combine Zero-shot Learning, Few-shot Learning, and RLHF techniques to generate a large language model.

[0016] A4. Model training: Use the formatted database to train the large language model. During the training process, predict the next word in the sentence based on the previous one or several words. Use the back propagation and stochastic gradient descent optimization algorithms to adjust the parameters of the large language model to generate a large language model for pumped storage power station construction management.

[0017] A5. Evaluation and Adjustment: Use perplexity and downstream task performance indicators to evaluate model performance and adjust model parameters.

[0018] Preferably, the specific contents executed by the intelligent AI assistant module include:

[0019] B1. Use the intelligent AI assistant module to connect the model construction module to obtain a large language model for pumped storage power station construction management, input the question into the intelligent AI assistant module, and the intelligent AI assistant module obtains the input content of the question to perform context understanding and extract the input content, label and split the text content, and semantically and structurally split the image and document content to obtain a labeled, semantically and structured input sequence;

[0020] B2. Execute the self-attention mechanism: obtain the input sequence in the execution content B1 and perform linear transformation, use the self-attention mechanism to obtain the attention weight distribution, and weight each element in the input sequence according to the distribution weight, extract the correlation between the input contents, and form each initial representation and the degree of representation correlation;

[0021] B3, processing the initial representation and the degree of representation association obtained in the execution content B2, reducing the judgment of the noun part of speech in the initial representation, improving the judgment of the verb part of speech, inferring the probability of the next word appearing, and obtaining an input sequence with a degree of representation association greater than a preset value, and matching the input sequence with a large language model data database of a pumped storage power station through a similarity matching method to generate an answer that meets the requirements;

[0022] B4, inputting the input sequence obtained in executing content B1 into the large language model for pumped storage power station construction management, and generating drawings using an image rendering method based on the large language model data database of the pumped storage power station;

[0023] B5. Input the input sequence obtained in the execution content B1 into the large language model for pumped storage power station construction management, obtain a solution based on the large language model data database of the pumped storage power station, and convert the solution into PDF or WORD format through the AIOffice method.

[0024] Preferably, the specific steps of executing content B4 are:

[0025] B401, linearly transform the labeled, semantic, and structured input sequence in the execution content B1, input the linearly transformed input sequence into the model, use the image rendering method to obtain the required number of first set of engineering drawings, and label and encode the first set of engineering drawings, compare and analyze them with the input sequence, and obtain the first set of difference points;

[0026] B402, fusing the input sequence, the first set of engineering drawings that have been labeled and encoded, and the first set of difference points, and generating a second set of engineering drawings by using an image rendering method based on the fused input sequence, the first set of engineering drawings that have been labeled and encoded, and the first set of difference points, performing labeling and encoding on the second set of engineering drawings, and performing comparative analysis with the input sequence to obtain a second set of difference points;

[0027] B403, fusing the input sequence, the first set of engineering drawings processed by labeling and encoding, the second set of engineering drawings processed by labeling and encoding, and the second set of difference points, and generating a third set of engineering drawings by using an image rendering method based on the fused input sequence, the first set of engineering drawings processed by labeling and encoding, the second set of engineering drawings processed by labeling and encoding, and the second set of difference points;

[0028] B404. Combine the first set of engineering drawings, the second set of engineering drawings and the third set of engineering drawings to obtain a total set of engineering drawings, label and encode the total set of engineering drawings, compare and analyze it with the input sequence, select several engineering drawings whose difference points are less than a preset value, record them as a set of drawings that meet the requirements, and apply scale and coding content annotations to the drawings.

[0029] Preferably, the specific steps of executing content B5 are:

[0030] B501, linearly transform the labeled, semantic, and structured input sequence in the execution content B1, and generate a set of scheme outlines, label and semantically process each level of the scheme outline set, merge the labeled and semantically processed scheme outline set and input it into the model, and match the input sequence and the labeled and semantically processed scheme outline set with the large language model data database of the pumped storage power station to obtain the content corresponding to each level of the scheme outline;

[0031] B502. Label, semanticize and structure the contents corresponding to each level of the matched scheme outline, generate drawings using the image rendering method, generate multiple attributes for the drawings in combination with the input sequence content, and further match the drawing content to generate the drawing number, drawing association description and drawing notes;

[0032] B503. Use the blade template engine and WebKit / Blink rendering engine to arrange and combine the contents of each level of the matched scheme outline and the generated drawings to obtain the scheme, and convert the scheme into PDF or WORD format through the AIOffice method.

[0033] Preferably, the model correction module performs the following steps: integrating professional data and professional knowledge, making correction adjustments to the large language model for pumped storage power station construction management by adjusting the large language model and prompt words, and making correction adjustments to the generated answers, drawings and plans by industry experts using a restricted knowledge base, specifically:

[0034] C1. Adjust the large language model: Collect information on the pumped storage industry, generate a database for a specified field, train the large language model for pumped storage power station construction management based on the database for the specified field, adjust the performance of the model in NPL tasks in a specific field, and realize model self-correction;

[0035] C2. Adjust prompt words: further realize model self-correction through designing template technology;

[0036] C3. Restricted knowledge base: Analyze, clean, correct and convert the content of the large language model database of pumped storage power stations, eliminate invalid data, label valid data, and combine the precise answers of industry experts to perform deep sentence correction and corresponding accuracy correction, so as to realize industry expert correction of generated answers, drawings and plans.

[0037] Preferably, the specific steps of executing content C2 are:

[0038] C201. Design template: Input questions to the large language model for pumped storage power station construction management, extract prompt words from the input questions, assume that the extracted prompt word is x, and transform the prompt word x into X containing empty slots through the design template technology. Fill the empty slots of X with different filling values ​​multiple times to obtain a set X containing different filling values 集 ;

[0039] C202, Answer search: for set X 集 Through the big language model for pumped storage power station construction management, multiple answers are obtained by searching in the big language model database of pumped storage power station;

[0040] C203, answer correction: For the multiple answers obtained in step C202, the best answer is found through a deep traversal method to achieve model self-correction.

[0041] Preferably, the specific contents executed by the resource library module include:

[0042] D1. Set the question as the initial key point, retrieve the closest preset question and answer through the initial key point, use the retrieved closest preset question and answer as the context, splice it with the question, and generate the final key point;

[0043] D2. Label the final key point to obtain the label corresponding to the final key point. Based on the final key point and the label corresponding to the final key point, use the large language model combined with the rsync algorithm for asynchronous addressing;

[0044] D3. Use the out-of-order mode and ordered mode of Flink's asynchronous addressing IO technology to store the content required in the context as the generated results in the resource library.

[0045] Preferably, the specific contents executed by the authorization management module include:

[0046] E1. The user communicates with the intelligent AI assistant module in natural language. The intelligent AI assistant module obtains the text, image, and document in the context, performs labeling and attribute processing, extracts user attribute parameters, and creates the underlying user table based on the user attribute parameters.

[0047] E2. The user communicates with the intelligent AI assistant module in natural language. The system compares the content obtained by the intelligent AI assistant module with the underlying user table to obtain the difference content, and compares whether the underlying user table contains the user. If the user is included, the difference content is extended to the underlying user table to complete the entry and editing of the user table;

[0048] E2. Through natural language communication with the intelligent AI assistant module, the text, images, documents and other content in the context are obtained to complete user access control and authentication.

[0049] Preferably, the pumped storage power station construction management system based on the large language model also includes a mobile terminal module, and the model construction module, the intelligent AI assistant module, the model correction module, the resource library module and the authorization management module are all functions of the web terminal;

[0050] The mobile terminal module is used to convert the functions of the web terminal into the functions of the mobile terminal. The functions of the mobile terminal are combined with the intelligent AI assistant module to complete the question answering and resource acquisition of the resource library module.

[0051] Preferably, the specific contents executed by the mobile terminal module include:

[0052] The web and mobile terminals are linked through data communication technology, and the functions and data of the web terminal are converted. The intelligent AI assistant module cooperates with the functions of the mobile terminal to answer questions, generate solutions, and complete resource acquisition.

[0053] Compared with the prior art, the beneficial effects of the present invention are:

[0054] (1) The present invention integrates technologies such as data integration, data preprocessing, intelligent AI, large language model and natural language processing to ensure the intelligent response capability of the intelligent AI assistant module;

[0055] (2) The present invention uses the massive data of pumped-storage power stations as basic data support, which can realize questions and answers on professional knowledge of pumped-storage power stations, suggestions for improvement of construction content during the construction period, etc., greatly reducing the management difficulty of managers;

[0056] (3) The present invention uses the authorization management module to enable managers to enter user information and roles into the system simply by communicating with the intelligent AI assistant module, provide authorization management services to users and applications, implement user access control mechanisms, replace the traditional web-based personnel management process, ensure that relevant personnel can quickly get started, and simplify system development and maintenance;

[0057] (4) The present invention uses natural language processing (NPL) to implement a dialogue system, combined with a large language model and machine learning technology, so that data can be quickly transmitted in both directions, greatly improving the real-time and convenience of users when receiving and processing messages or tasks;

[0058] (5) The present invention establishes an intelligent management system for pumped-storage power stations based on the "Regulations for the Preparation of Feasibility Study Reports for Hydropower Projects" (NB / T 11013-2022) and in combination with actual project use, providing the ability to handle professional issues for personnel at all stages and positions during the construction period, laying the foundation for achieving the construction requirements of a "digital power plant";

[0059] (6) The present invention introduces online collaborative technology, combines the large language model and the large language model data database of the pumped storage power station, and can provide multi-source supplements for professional issues responded by the model to enrich the model capabilities, playing a significant role in improving the standardization and standardization of response capabilities;

[0060] (7) The present invention solves the problems of difficulty in answering professional questions and backlog of repetitive work during the construction of pumped-storage power stations, and realizes intelligent response and automated learning accumulation. It can classify, sort out and learn various construction drawings of previous projects. Users only need to input the requirements related to the construction of pumped-storage power stations to generate plans and drawings. If the results are inaccurate, more parameters can be input. The model can be further adjusted on the current basis through the resource library module to obtain a drawing that better meets the requirements. Technical personnel can simply improve the drawing and put it into use, reducing the huge amount of drawing. The series of results formed can provide reference value for the digital and intelligent applications of projects under construction or already built. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 This is an architecture diagram of a pumped storage power station construction management system based on a large language model of the present invention;

[0062] Figure 2It is a technical roadmap for the construction management system of a pumped-storage power station based on a large language model according to an embodiment of the present invention. DETAILED DESCRIPTION

[0063] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0064] Example 1

[0065] Combination Figure 1 and Figure 2 As shown, an embodiment of the present invention provides a pumped storage power station construction management system based on a large language model, including a model construction module, an intelligent AI assistant module, a model correction module, a resource library module and an authorization management module;

[0066] The model building module is used to collect data and generate a large language model data database for pumped storage power stations. Based on the database, the Transformer neural network model architecture is selected, and Zero-shot Learning, Few-shot Learning and RLHF technologies are combined to generate a large language model for pumped storage power station construction management.

[0067] Specifically, the model construction module provides basic support for the system. By uploading and entering various professional data such as project information, geographic information, personnel scale, equipment data, drawing information, industry standards and specifications, document information, the role of pumped-storage power stations, energy consumption, project information, installed capacity, etc., various professionals integrate a large amount of data collected on the Internet based on various pumped-storage projects, including power station construction archives, technical requirements, personnel composition, construction progress, books, articles, websites, etc., to generate a large language model data database for pumped-storage power stations. The database covers various data information related to pumped-storage power stations, and trains the large language model based on the database. The model can automatically parse the relevant data for machine learning. After learning is completed, it can provide professional ability support in subsequent dialogues to form a large language model for pumped-storage power station construction management.

[0068] The intelligent AI assistant module is used to connect the model building module and process the input questions to obtain input sequences. It uses the self-attention mechanism to extract the correlation between the input sequences, matches the answers that meet the requirements based on the correlation, and generates drawings and plans based on the large language model data database of pumped storage power stations.

[0069] Specifically, the intelligent AI assistant module can provide users with professional answers to questions, handle user consultation questions, and process sequence data according to the context. It uses a self-attention mechanism to establish the correlation between the elements in the input sequence, and captures the semantic and grammatical relationship between words by encoding and decoding the words or tokens in the input sequence, thereby generating coherent and professional answers, including completing pumped storage problem answers, drawing generation, and solution generation.

[0070] The model correction module is used to make corrections to the large language model for pumped storage power station construction management generated by the model construction module, and to make corrections to the answers, drawings and plans generated by the intelligent AI assistant module by industry experts;

[0071] Specifically, the model correction module can make corrections to the model and the response results of the questions. After the user asks a relevant question, the answer given by the system may be different from the actual situation. Relevant professionals can correct the corresponding questions and determine the standard degree of each answer. The system learns, summarizes and analyzes multiple answers. When encountering similar questions in the future, the system can provide users with relatively correct response content based on the correction results.

[0072] The resource library module is used to store the answers, drawings and plans generated by the intelligent AI assistant module. The user communicates with the intelligent AI assistant module based on the content stored in the resource library. The system iterates the stored content to generate an accurate version according to the user's needs and stores the accurate version again.

[0073] Specifically, the resource library module can uniformly save the answers, drawings, plans, source codes and other contents generated by the system. Users can conduct in-depth dialogues with the intelligent AI assistant module based on the contents stored in the resource library. The system iteratively generates more accurate and compliant versions according to the in-depth requirements of users, and stores and annotates the generation basis of each version, so that users can uniformly manage the response results and review the process, thereby improving the convenience of operation and the traceability of response results.

[0074] The authorization management module is used to enter user attributes into the system through dialogue between the user and the intelligent AI assistant module, and to control user access based on user attributes;

[0075] Specifically, the authorization management module enables managers to enter user information and roles into the system simply by communicating with the intelligent AI assistant module, provide authorization management services to users and applications, implement user access control mechanisms, replace the traditional web-based personnel management process, and simplify system development and maintenance;

[0076] Furthermore, the pumped storage power station construction management system based on the large language model also includes a mobile terminal module;

[0077] The model building module, intelligent AI assistant module, model correction module, resource library module and authorization management module are all web-based functions;

[0078] The mobile terminal module is used to convert the functions of the web terminal into the functions of the mobile terminal. The functions of the mobile terminal are combined with the intelligent AI assistant module to complete the question answering and resource acquisition of the resource library module;

[0079] Specifically, mobile devices such as mobile phones and pads can be used to perform inquiries, answers, and solution designs based on the underlying model data, achieving the goal of multi-terminal integration, improving the efficiency and convenience of user use, and realizing a closed loop from the foundation to the entire process and life cycle of the solution;

[0080] The pumped-storage power station construction management system based on the large language model can answer questions related to the construction of pumped-storage power stations and generate solutions, such as: being able to respond to the construction process of pumped-storage power stations, generate construction plans for ventilation tunnels of pumped-storage power stations, answer precautions for flood discharge tunnel construction, etc.

[0081] Example 2

[0082] like Figure 1 and Figure 2 As shown, a pumped storage power station construction management system based on a large language model includes a model construction module, an intelligent AI assistant module, a model correction module, a resource library module and an authorization management module;

[0083] The model building module is used to collect data and generate a large language model data database for pumped storage power stations. Based on the database, the Transformer neural network model architecture is selected, and Zero-shot Learning, Few-shot Learning and RLHF technologies are combined to generate a large language model for pumped storage power station construction management.

[0084] The intelligent AI assistant module is used to connect the model building module and process the input questions to obtain input sequences. It uses the self-attention mechanism to extract the correlation between the input sequences, matches the answers that meet the requirements based on the correlation, and generates drawings and plans based on the large language model data database of pumped storage power stations.

[0085] The model correction module is used to make corrections to the large language model for pumped storage power station construction management generated by the model construction module, and to make corrections to the answers, drawings and plans generated by the intelligent AI assistant module by industry experts;

[0086] The resource library module is used to store the answers, drawings and plans generated by the intelligent AI assistant module. The user communicates with the intelligent AI assistant module based on the content stored in the resource library. The system iterates the stored content to generate an accurate version according to the user's needs and stores the accurate version again.

[0087] The authorization management module is used to enter user attributes into the system through dialogue between the user and the intelligent AI assistant module, and to control user access based on user attributes;

[0088] Furthermore, the pumped storage power station construction management system based on the large language model also includes a mobile terminal module, and the model construction module, intelligent AI assistant module, model correction module, resource library module and authorization management module are all web-side functions;

[0089] The mobile terminal module is used to convert the functions of the web terminal into the functions of the mobile terminal. The functions of the mobile terminal are combined with the intelligent AI assistant module to complete the question answering and resource acquisition of the resource library module;

[0090] On this basis, in this embodiment, the model building module performs the following:

[0091] Data preprocessing is performed on the pumped-storage power station big language model data database, and the big language model for the construction and management of pumped-storage power stations is trained based on the database. The parameters of the model are adjusted, the performance of the model is evaluated and the parameters are adjusted to generate a big language model for the construction and management of pumped-storage power stations.

[0092] The specific steps are as follows:

[0093] A1. Basic data support: Collect basic data of pumped storage power stations to form a large language model data database of pumped storage power stations;

[0094] A2. Data preprocessing: Data preprocessing is performed on the large language model database of pumped storage power stations. Database deduplication and semantic similarity calculation methods are used to clean the database, delete duplicate or irrelevant content, and format the preprocessed database.

[0095] A3. Model generation: Select the Transformer neural network model architecture, determine the size of the model based on its ability to efficiently process sequence data and capture long-term dependencies, select the number of model parameters, hidden units, and hyperparameters such as learning rate and batch size, and combine Zero-shot Learning, Few-shot Learning, and RLHF (Reinforcement Learning based on human feedback) techniques to generate a large language model (Fuyao AI);

[0096] A4. Model training: Use the formatted database to train the large language model. During the training process, predict the next word in the sentence based on the previous one or several words. Use back propagation and stochastic gradient descent optimization algorithms to adjust the parameters (weights and biases) of the large language model to generate a large language model for pumped storage power station construction management.

[0097] A5. Evaluation and Adjustment: Use perplexity (a measure of how well the model predicts the data) and downstream task performance indicators to evaluate model performance, adjust the model parameters (specifically performance, accuracy, etc.), and perform adjustments to improve specific aspects of model performance.

[0098] Example 3

[0099] like Figure 1 and Figure 2 As shown, a pumped storage power station construction management system based on a large language model includes a model construction module, an intelligent AI assistant module, a model correction module, a resource library module and an authorization management module;

[0100] The model building module is used to collect data and generate a large language model data database for pumped storage power stations. Based on the database, the Transformer neural network model architecture is selected, and Zero-shot Learning, Few-shot Learning and RLHF technologies are combined to generate a large language model for pumped storage power station construction management.

[0101] The intelligent AI assistant module is used to connect the model building module and process the input questions to obtain input sequences. It uses the self-attention mechanism to extract the correlation between the input sequences, matches the answers that meet the requirements based on the correlation, and generates drawings and plans based on the large language model data database of pumped storage power stations.

[0102] The model correction module is used to make corrections to the large language model for pumped storage power station construction management generated by the model construction module, and to make corrections to the answers, drawings and plans generated by the intelligent AI assistant module by industry experts;

[0103] The resource library module is used to store the answers, drawings and plans generated by the intelligent AI assistant module. The user communicates with the intelligent AI assistant module based on the content stored in the resource library. The system iterates the stored content to generate an accurate version according to the user's needs and stores the accurate version again.

[0104] The authorization management module is used to enter user attributes into the system through dialogue between the user and the intelligent AI assistant module, and to control user access based on user attributes;

[0105] Furthermore, the pumped storage power station construction management system based on the large language model also includes a mobile terminal module, and the model construction module, intelligent AI assistant module, model correction module, resource library module and authorization management module are all web-side functions;

[0106] The mobile terminal module is used to convert the functions of the web terminal into the functions of the mobile terminal. The functions of the mobile terminal are combined with the intelligent AI assistant module to complete the question answering and resource acquisition of the resource library module;

[0107] On this basis, in this embodiment, the contents executed by the intelligent AI assistant module include:

[0108] Use the intelligent AI assistant module to connect to the large language model for pumped storage power station construction management obtained in step A4, perform visual interaction, generate a user interface, and input the question into the intelligent AI assistant module. After obtaining the input question, the intelligent AI assistant module performs context understanding and content labeling, semantic and structural processing, executes the self-attention mechanism, and generates answers, drawings and plans based on the large language model data database of the pumped storage power station;

[0109] The specific steps are as follows:

[0110] B1. Use the intelligent AI assistant module to connect the model construction module to obtain a large language model for pumped storage power station construction management, input the question into the intelligent AI assistant module, and the intelligent AI assistant module obtains the input content of the question to perform context understanding and extract the input content, label and split the text content, and semantically and structurally split the image and document content to obtain a labeled, semantically and structured input sequence;

[0111] B2. Execute the self-attention mechanism: obtain the input sequence in the execution content B1 and perform linear transformation, use the self-attention mechanism to obtain the attention weight distribution, and weight each element in the input sequence according to the distribution weight, extract the correlation between the input contents, and form each initial representation and the degree of representation correlation;

[0112] B3. Generate an answer based on the large language model database of the pumped storage power station: process each initial representation and the degree of association of the representations in step B2, reduce the judgment of the noun part of speech in the initial representation, improve the judgment of the verb part of speech, infer the probability of the next word appearing, and obtain an input sequence with a degree of association of the representation greater than a preset value, and match the input sequence with the large language model database of the pumped storage power station through a similarity matching method to generate a solution that meets the requirements;

[0113] B4, inputting the input sequence obtained in executing content B1 into the large language model for pumped storage power station construction management, and generating drawings using an image rendering method based on the large language model data database of the pumped storage power station;

[0114] B5. Input the input sequence obtained in executing content B1 into the large language model for pumped storage power station construction management, obtain a solution based on the large language model data database of the pumped storage power station, and convert the solution into PDF or WORD format through the AIOffice method;

[0115] Specifically, the specific steps for executing content B4 are:

[0116] B401, linearly transform the labeled, semantic, and structured input sequence in the execution content B1, input the linearly transformed input sequence into the model, use the image rendering method (Object-order rendering) to obtain the required number of first set of engineering drawings, and label and encode the first set of engineering drawings, and compare and analyze them with the input sequence. Specifically, the depth-first traversal and breadth-first traversal techniques can be used for comparison and analysis to obtain the first set of difference points;

[0117] B402, fusing the input sequence, the first set of engineering drawings that have been labeled and encoded, and the first set of difference points, and generating a second set of engineering drawings based on the fused input sequence, the first set of engineering drawings that have been labeled and encoded, and the first set of difference points using an image rendering method, and performing labeling and encoding on the second set of engineering drawings, and performing comparative analysis with the input sequence to obtain a second set of difference points;

[0118] B403, fusing the input sequence, the first set of engineering drawings processed by labeling and encoding, the second set of engineering drawings processed by labeling and encoding, and the second set of difference points, and generating a third set of engineering drawings by using an image rendering method based on the fused input sequence, the first set of engineering drawings processed by labeling and encoding, the second set of engineering drawings processed by labeling and encoding, and the second set of difference points;

[0119] B404. The first set of engineering drawings, the second set of engineering drawings and the third set of engineering drawings are combined to obtain a total set of engineering drawings, the total set of engineering drawings is labeled and coded, and compared with the input sequence for analysis, and several engineering drawings with differences less than a preset value are selected as the drawing sets that meet the requirements, and scales and coding content are added to the drawings;

[0120] The specific steps to implement content B5 are:

[0121] B501. Perform linear transformation on the labeled, semantic, and structured input sequence in the execution content B1, and generate a set of scheme outlines. Specifically, the set of scheme outlines can be generated using the thinking chain technology, and each level of the scheme outline set can be labeled and semantically processed. The labeled and semantically processed scheme outline set is merged and input into the model. Based on the input sequence and the labeled and semantically processed scheme outline set, a natural language processing method is used to match with the large language model data database of the pumped storage power station to obtain the content corresponding to each level of the scheme outline;

[0122] B502. Label, semanticize and structure the contents corresponding to each level of the matched scheme outline, generate drawings using image rendering methods, and generate multiple attributes (such as width, height, length, etc.) for the drawings in combination with the input sequence content. Specifically, use text segmentation technology and semantic technology in combination with the input sequence content to generate multiple attributes for the drawings, and further match the drawing content to generate drawing numbers, drawing association instructions, drawing notes, etc.;

[0123] B503. Use the blade template engine and WebKit / Blink rendering engine to arrange and combine the contents of each level of the matched scheme outline and the generated drawings to obtain the scheme, and convert the scheme into PDF or WORD format through the AIOffice method.

[0124] Example 4

[0125] like Figure 1 and Figure 2 As shown, a pumped storage power station construction management system based on a large language model includes a model construction module, an intelligent AI assistant module, a model correction module, a resource library module and an authorization management module;

[0126] The model building module is used to collect data and generate a large language model data database for pumped storage power stations. Based on the database, the Transformer neural network model architecture is selected, and Zero-shot Learning, Few-shot Learning and RLHF technologies are combined to generate a large language model for pumped storage power station construction management.

[0127] The intelligent AI assistant module is used to connect the model building module and process the input questions to obtain input sequences. It uses the self-attention mechanism to extract the correlation between the input sequences, matches the answers that meet the requirements based on the correlation, and generates drawings and plans based on the large language model data database of pumped storage power stations.

[0128] The model correction module is used to make corrections to the large language model for pumped storage power station construction management generated by the model construction module, and to make corrections to the answers, drawings and plans generated by the intelligent AI assistant module by industry experts;

[0129] The resource library module is used to store the answers, drawings and plans generated by the intelligent AI assistant module. The user communicates with the intelligent AI assistant module based on the content stored in the resource library. The system iterates the stored content to generate an accurate version according to the user's needs and stores the accurate version again.

[0130] The authorization management module is used to enter user attributes into the system through dialogue between the user and the intelligent AI assistant module, and to control user access based on user attributes;

[0131] Furthermore, the pumped storage power station construction management system based on the large language model also includes a mobile terminal module, and the model construction module, intelligent AI assistant module, model correction module, resource library module and authorization management module are all web-side functions;

[0132] The mobile terminal module is used to convert the functions of the web terminal into the functions of the mobile terminal. The functions of the mobile terminal are combined with the intelligent AI assistant module to complete the question answering and resource acquisition of the resource library module;

[0133] On this basis, in this embodiment, the model correction module performs the following operations:

[0134] Integrate professional data and professional knowledge, use the adjustment of the large language model and the adjustment of prompt words to correct the large language model for pumped storage power station construction management, and use the restricted knowledge base to correct the generated answers, drawings and plans by industry experts, so as to further enhance the professional capabilities of the system for pumped storage power stations;

[0135] The specific steps are as follows:

[0136] C1. Adjust the large language model: Collect information on the pumped storage industry, generate a database for a specified field, train the large language model for pumped storage power station construction management based on the database for the specified field, adjust the performance of the model in NPL tasks in a specific field, and realize model self-correction;

[0137] Specifically, NPL tasks in specific fields include sentiment analysis, entity recognition, text, image, speech, video content classification, dialogue generation, etc.

[0138] C2. Adjust prompt words: further realize model self-correction through designing template technology;

[0139] C3. Restricted knowledge base: Analyze, clean, convert and correct the content of the large language model database of pumped storage power stations, eliminate invalid data, and label valid data. Industry experts mark the accuracy of the generated answers and weight the generated answers to obtain the accurate answers of industry experts. Combined with the accurate answers of industry experts, deep sentence correction and corresponding accuracy correction are performed to achieve industry expert correction of the generated solutions, drawings, and plans, and realize the return of clean data.

[0140] Furthermore, the specific steps for executing content C2 are:

[0141] C201. Design template: Input questions to the large language model for pumped storage power station construction management, extract prompt words from the input questions, assume that the extracted prompt word is x, and transform the prompt word x into X containing empty slots through the design template technology. Fill the empty slots of X with different filling values ​​multiple times to obtain a set X containing different filling values 集 ;

[0142] C202, Answer search: for set X 集 Through the big language model for pumped storage power station construction management, multiple answers are obtained by searching in the big language model database of pumped storage power station;

[0143] C203, answer correction: For the multiple answers obtained in step C202, the best answer is found through a deep traversal method to achieve model self-correction.

[0144] Example 5

[0145] like Figure 1 and Figure 2 As shown, a pumped storage power station construction management system based on a large language model includes a model construction module, an intelligent AI assistant module, a model correction module, a resource library module and an authorization management module;

[0146] The model building module is used to collect data and generate a large language model data database for pumped storage power stations. Based on the database, the Transformer neural network model architecture is selected, and Zero-shot Learning, Few-shot Learning and RLHF technologies are combined to generate a large language model for pumped storage power station construction management.

[0147] The intelligent AI assistant module is used to connect the model building module and process the input questions to obtain input sequences. It uses the self-attention mechanism to extract the correlation between the input sequences, matches the answers that meet the requirements based on the correlation, and generates drawings and plans based on the large language model data database of pumped storage power stations.

[0148] The model correction module is used to make corrections to the large language model for pumped storage power station construction management generated by the model construction module, and to make corrections to the answers, drawings and plans generated by the intelligent AI assistant module by industry experts;

[0149] The resource library module is used to store the answers, drawings and plans generated by the intelligent AI assistant module. The user communicates with the intelligent AI assistant module based on the content stored in the resource library. The system iterates the stored content to generate an accurate version according to the user's needs and stores the accurate version again.

[0150] The authorization management module is used to enter user attributes into the system through dialogue between the user and the intelligent AI assistant module, and to control user access based on user attributes;

[0151] Furthermore, the pumped storage power station construction management system based on the large language model also includes a mobile terminal module, and the model construction module, intelligent AI assistant module, model correction module, resource library module and authorization management module are all web-side functions;

[0152] The mobile terminal module is used to convert the functions of the web terminal into the functions of the mobile terminal. The functions of the mobile terminal are combined with the intelligent AI assistant module to complete the question answering and resource acquisition of the resource library module;

[0153] On this basis, in this embodiment, the content executed by the resource library module is to connect the large language model based on the intelligent AI assistant module, capture key points by obtaining the content in the context, asynchronously address in the database, organize the relevant content into various results and store them in the resource library, and further review the resource library response content, files, drawings, etc., and put forward in-depth and detailed requirements for the system to obtain more accurate results, which are as follows:

[0154] D1. Set the question as the initial key point, retrieve the closest preset question and answer through the initial key point, use the retrieved closest preset question and answer as the context, splice it with the question, and generate the final key point;

[0155] D2. Label the final key point to obtain the label corresponding to the final key point. Based on the final key point and the label corresponding to the final key point, use the large language model combined with the rsync algorithm for asynchronous addressing;

[0156] D3. Using the unordered mode and ordered mode of Flink's asynchronous addressing IO technology, the content required in the context is stored in the resource library as the generated results. Users can preview and further refine the generated results.

[0157] Example 6

[0158] like Figure 1 and Figure 2 As shown, a pumped storage power station construction management system based on a large language model includes a model construction module, an intelligent AI assistant module, a model correction module, a resource library module and an authorization management module;

[0159] The model building module is used to collect data and generate a large language model data database for pumped storage power stations. Based on the database, the Transformer neural network model architecture is selected, and Zero-shot Learning, Few-shot Learning and RLHF technologies are combined to generate a large language model for pumped storage power station construction management.

[0160] The intelligent AI assistant module is used to connect the model building module and process the input questions to obtain input sequences. It uses the self-attention mechanism to extract the correlation between the input sequences, matches the answers that meet the requirements based on the correlation, and generates drawings and plans based on the large language model data database of pumped storage power stations.

[0161] The model correction module is used to make corrections to the large language model for pumped storage power station construction management generated by the model construction module, and to make corrections to the answers, drawings and plans generated by the intelligent AI assistant module by industry experts;

[0162] The resource library module is used to store the answers, drawings and plans generated by the intelligent AI assistant module. The user communicates with the intelligent AI assistant module based on the content stored in the resource library. The system iterates the stored content to generate an accurate version according to the user's needs and stores the accurate version again.

[0163] The authorization management module is used to enter user attributes into the system through dialogue between the user and the intelligent AI assistant module, and to control user access based on user attributes;

[0164] Furthermore, the pumped storage power station construction management system based on the large language model also includes a mobile terminal module, and the model construction module, intelligent AI assistant module, model correction module, resource library module and authorization management module are all web-side functions;

[0165] The mobile terminal module is used to convert the functions of the web terminal into the functions of the mobile terminal. The functions of the mobile terminal are combined with the intelligent AI assistant module to complete the question answering and resource acquisition of the resource library module;

[0166] On this basis, in this embodiment, the authorization management module performs the following:

[0167] Through the human-computer dialogue of the system, the intelligent AI assistant module is used to replace the traditional web page to achieve user authorization and access control;

[0168] E1. The user communicates with the intelligent AI assistant module in natural language. The intelligent AI assistant module obtains the text, image, and document in the context, performs labeling and attribute processing, extracts user attribute parameters, and creates the underlying user table based on the user attribute parameters.

[0169] E2. The user communicates with the intelligent AI assistant module in natural language. The system compares the content obtained by the intelligent AI assistant module with the underlying user table to obtain the difference content, and compares whether the underlying user table contains the user. If the user is included, the difference content is extended to the underlying user table to complete the entry and editing of the user table;

[0170] E2, through natural language communication with the intelligent AI assistant module, obtain the text, images, documents and other content in the context to complete user access control and authentication;

[0171] Specifically, if "Zhang San 12345678901" is input, the large language model system automatically splits the content into: "Zhang San" and "12345678901", and searches the user table to see if the user Zhang San exists, and obtains the set of all users who are Zhang San; compares the attributes of all users in the set to determine whether "12345678901" is a unique identifier among multiple attributes of a certain user Zhang San. If the content is not determined to be a unique identifier, the intelligent AI assistant module is used to obtain more content from the user and perform addressing judgment again. If the unique identifier can be determined, the user behavior is analyzed (such as login IP, login location, login device, etc.);

[0172] If the user behavior analysis is normal, the user is allowed to log in and use the large language model system normally. If the user behavior analysis is abnormal, the core fields are determined from the user's multiple attributes (there are fields in Zhang San's attributes that can prove his identity, such as ID number and set verification code), and obtained from the user through the intelligent AI assistant module. The system can be used normally only after verification to ensure user data security.

[0173] Furthermore, the specific contents executed by the mobile terminal module include:

[0174] The web and mobile terminals are linked through data communication technology, and the functions and data of the web terminal are converted. The intelligent AI assistant module cooperates with the functions of the mobile terminal to answer questions, generate solutions, and complete resource acquisition.

[0175] The above description is only an embodiment of the present invention and is not intended to limit the present invention. Any modification, equivalent replacement and improvement made within the application scope of the present invention shall be included in the protection scope of the present invention.

Claims

1. A pumped storage power station construction management system based on a large language model, characterized in that: It includes model construction module, intelligent AI assistant module, model correction module, resource library module and authorization management module; The model building module is used to collect data and generate a large language model data database for pumped storage power stations. Based on the database, the Transformer neural network model architecture is selected, and Zero-shot Learning, Few-shot Learning and RLHF technologies are combined to generate a large language model for pumped storage power station construction management. The intelligent AI assistant module is used to connect the model building module and process the input questions to obtain input sequences. It uses the self-attention mechanism to extract the correlation between the input sequences, matches the answers that meet the requirements based on the correlation, and generates drawings and plans based on the large language model data database of pumped storage power stations. The model correction module is used to make corrections to the large language model for pumped storage power station construction management generated by the model construction module, and to make corrections to the answers, drawings and plans generated by the intelligent AI assistant module by industry experts; The resource library module is used to store the answers, drawings and plans generated by the intelligent AI assistant module. The user communicates with the intelligent AI assistant module based on the content stored in the resource library. The system iterates the stored content to generate an accurate version according to the user's needs and stores the accurate version again. The authorization management module is used to enter user attributes into the system through dialogue between the user and the intelligent AI assistant module, and to perform access control on the user based on the user attributes.

2. The pumped storage power station construction management system based on a large language model according to claim 1 is characterized in that: The specific contents of the model building module include: A1. Basic data support: Collect basic data of pumped storage power stations to form a large language model data database of pumped storage power stations; A2. Data preprocessing: Data preprocessing is performed on the large language model database of pumped storage power stations. Database deduplication and semantic similarity calculation methods are used to clean the database, delete duplicate or irrelevant content, and format the preprocessed database. A3. Model generation: Select the Transformer neural network model architecture, and select the number of parameters, hidden units, and hyperparameters of the model based on its ability to process sequence data and capture long-term dependencies. Combine Zero-shot Learning, Few-shot Learning, and RLHF technical methods to generate a large language model. A4. Model training: Use the formatted database to train the large language model. During the training process, predict the next word in the sentence based on the previous one or several words. Use the back propagation and stochastic gradient descent optimization algorithms to adjust the parameters of the large language model to generate a large language model for pumped storage power station construction management. A5. Evaluation and Adjustment: Use perplexity and downstream task performance indicators to evaluate model performance and adjust model parameters.

3. The pumped storage power station construction management system based on a large language model according to claim 1 is characterized in that: The specific contents executed by the intelligent AI assistant module include: B1. Use the intelligent AI assistant module to connect the model construction module to obtain a large language model for pumped storage power station construction management, input the question into the intelligent AI assistant module, and the intelligent AI assistant module obtains the input content of the question to perform context understanding and extract the input content, label and split the text content, and semantically and structurally split the image and document content to obtain a labeled, semantically and structured input sequence; B2. Execute the self-attention mechanism: obtain the input sequence in the execution content B1 and perform linear transformation, use the self-attention mechanism to obtain the attention weight distribution, and weight each element in the input sequence according to the distribution weight, extract the correlation between the input contents, and form each initial representation and the degree of representation correlation; B3, processing the initial representation and the degree of representation association obtained in the execution content B2, reducing the judgment of the noun part of speech in the initial representation, improving the judgment of the verb part of speech, inferring the probability of the next word appearing, and obtaining an input sequence with a degree of representation association greater than a preset value, and matching the input sequence with a large language model data database of a pumped storage power station through a similarity matching method to generate an answer that meets the requirements; B4, inputting the input sequence obtained in executing content B1 into the large language model for pumped storage power station construction management, and generating drawings using an image rendering method based on the large language model data database of the pumped storage power station; B5. Input the input sequence obtained in the execution content B1 into the large language model for pumped storage power station construction management, obtain a solution based on the large language model data database of the pumped storage power station, and convert the solution into PDF or WORD format through the AIOffice method.

4. The pumped storage power station construction management system based on a large language model according to claim 3 is characterized in that: The specific steps to implement content B4 are: B401, linearly transform the labeled, semantic, and structured input sequence in the execution content B1, input the linearly transformed input sequence into the model, use the image rendering method to obtain the required number of first set of engineering drawings, and label and encode the first set of engineering drawings, compare and analyze them with the input sequence, and obtain the first set of difference points; B402, fusing the input sequence, the first set of engineering drawings that have been labeled and encoded, and the first set of difference points, and generating a second set of engineering drawings by using an image rendering method based on the fused input sequence, the first set of engineering drawings that have been labeled and encoded, and the first set of difference points, performing labeling and encoding on the second set of engineering drawings, and performing comparative analysis with the input sequence to obtain a second set of difference points; B403, fusing the input sequence, the first set of engineering drawings processed by labeling and encoding, the second set of engineering drawings processed by labeling and encoding, and the second set of difference points, and generating a third set of engineering drawings by using an image rendering method based on the fused input sequence, the first set of engineering drawings processed by labeling and encoding, the second set of engineering drawings processed by labeling and encoding, and the second set of difference points; B404. Combine the first set of engineering drawings, the second set of engineering drawings and the third set of engineering drawings to obtain a total set of engineering drawings, label and encode the total set of engineering drawings, compare and analyze it with the input sequence, select several engineering drawings whose difference points are less than a preset value, record them as a set of drawings that meet the requirements, and apply scale and coding content annotations to the drawings.

5. The pumped storage power station construction management system based on a large language model according to claim 3 is characterized in that: The specific steps to implement content B5 are: B501, linearly transform the labeled, semantic, and structured input sequence in the execution content B1, and generate a set of scheme outlines, label and semantically process each level of the scheme outline set, merge the labeled and semantically processed scheme outline set and input it into the model, and match the input sequence and the labeled and semantically processed scheme outline set with the large language model data database of the pumped storage power station to obtain the content corresponding to each level of the scheme outline; B502. Label, semanticize and structure the contents corresponding to each level of the matched scheme outline, generate drawings using the image rendering method, generate multiple attributes for the drawings in combination with the input sequence content, and further match the drawing content to generate the drawing number, drawing association description and drawing notes; B503. Use the blade template engine and WebKit / Blink rendering engine to arrange and combine the contents of each level of the matched scheme outline and the generated drawings to obtain the scheme, and convert the scheme into PDF or WORD format through the AIOffice method.

6. The pumped storage power station construction management system based on a large language model according to claim 1, characterized in that: The model correction module performs the following tasks: integrating professional data and expertise, making corrections to the large language model for pumped storage power station construction management by adjusting the large language model and prompt words, and making corrections to the generated answers, drawings and plans by industry experts using a restricted knowledge base. Specifically: C1. Adjust the large language model: Collect information on the pumped storage industry, generate a database for a specified field, train the large language model for pumped storage power station construction management based on the database for the specified field, adjust the performance of the model in NPL tasks in a specific field, and realize model self-correction; C2. Adjust prompt words: further realize model self-correction through designing template technology; C3. Restricted knowledge base: Analyze, clean, correct and convert the content of the large language model database of pumped storage power stations, eliminate invalid data, label valid data, and combine the precise answers of industry experts to perform deep sentence correction and corresponding accuracy correction, so as to realize industry expert correction of generated answers, drawings and plans.

7. The pumped storage power station construction management system based on a large language model according to claim 6 is characterized in that: The specific steps to execute content C2 are: C201. Design template: Input questions to the large language model for pumped storage power station construction management, extract prompt words from the input questions, assume that the extracted prompt word is x, and transform the prompt word x into X containing empty slots through the design template technology. Fill the empty slots of X with different filling values ​​multiple times to obtain a set X containing different filling values 集 ; C202, Answer search: for set X 集 Through the big language model for pumped storage power station construction management, multiple answers are obtained by searching in the big language model database of pumped storage power station; C203, answer correction: For the multiple answers obtained in step C202, the best answer is found through a deep traversal method to achieve model self-correction.

8. The pumped storage power station construction management system based on a large language model according to claim 1, characterized in that: The specific contents of the resource library module include: D1. Set the question as the initial key point, retrieve the closest preset question and answer through the initial key point, use the retrieved closest preset question and answer as the context, splice it with the question, and generate the final key point; D2. Label the final key point to obtain the label corresponding to the final key point. Based on the final key point and the label corresponding to the final key point, use the large language model combined with the rsync algorithm for asynchronous addressing; D3. Use the out-of-order mode and ordered mode of Flink's asynchronous addressing IO technology to store the content required in the context as the generated results in the resource library.

9. The pumped storage power station construction management system based on a large language model according to claim 1, characterized in that: The specific contents of the authorization management module include: E1. The user communicates with the intelligent AI assistant module in natural language. The intelligent AI assistant module obtains the text, image, and document in the context, performs labeling and attribute processing, extracts user attribute parameters, and creates the underlying user table based on the user attribute parameters. E2. The user communicates with the intelligent AI assistant module in natural language. The system compares the content obtained by the intelligent AI assistant module with the underlying user table to obtain the difference content, and compares whether the underlying user table contains the user. If the user is included, the difference content is extended to the underlying user table to complete the entry and editing of the user table; E2. Through natural language communication with the intelligent AI assistant module, the text, images, documents and other content in the context are obtained to complete user access control and authentication.

10. The pumped storage power station construction management system based on a large language model according to claim 1, characterized in that: The pumped storage power station construction management system based on the large language model also includes a mobile terminal module. The model construction module, intelligent AI assistant module, model correction module, resource library module and authorization management module are all web-side functions; The mobile terminal module is used to convert the functions of the web terminal into the functions of the mobile terminal. The functions of the mobile terminal are combined with the intelligent AI assistant module to complete the question answering and resource acquisition of the resource library module.

11. The pumped storage power station construction management system based on a large language model according to claim 10, characterized in that: The specific contents of the mobile terminal module include: The web and mobile terminals are linked through data communication technology, and the functions and data of the web terminal are converted. The intelligent AI assistant module cooperates with the functions of the mobile terminal to answer questions, generate solutions, and complete resource acquisition.

Images