Power generation industry data intelligent treatment method, device and equipment based on large model

Abstract

The invention relates to the technical field of natural language processing, and discloses a power generation industry data intelligent treatment method, device and equipment based on a large model, and the method comprises the steps: constructing a multi-source heterogeneous task data set covering structured and unstructured information, completing the fine tuning training of a plurality of industry sub-fields based on a language model, forming a large language model set with specific scene adaptability; constructing a multi-view semantic representation structure for actual input data, integrating modeling task intention, application scene and model adaptability, predicting an optimal target model and a Top-K candidate model, and generating a unified semantic embedding vector; and realizing accurate matching of the structured knowledge fragments through the graph neural network. The problems that an existing model is insufficient in semantic understanding, inflexible in model selection and inaccurate in knowledge calling in the data management process are solved, the requirements of diversified tasks for accuracy and specialty are met, and then management of data assets is facilitated.

Description

Technical Field

[0001] This application relates to the field of natural language processing technology, for example to a data intelligent governance method, device, or equipment for the power generation industry based on a large model. Background Technology

[0002] Large language models have demonstrated powerful capabilities in natural language processing, knowledge understanding, and content generation, providing new pathways for the intelligent transformation of multiple industries. In the power generation industry, facing business needs characterized by complex operational data, strong semantic logic, and dense technical terminology, building language model systems with intelligent question answering and data governance capabilities has become a research hotspot. However, existing large language models generally rely on open and general-purpose corpora for pre-training, making it difficult to accurately understand the terminology, scheduling logic, and contextual habits in power generation operations. This leads to problems such as misunderstandings, missing content, and unstable responses during industry implementation.

[0003] Meanwhile, the power generation industry has significantly higher requirements for the accuracy, interpretability, and controllability of model outputs than general scenarios, especially in key tasks such as dispatch instruction verification, equipment anomaly diagnosis, and policy response analysis, which pose greater challenges to the system's professional responsiveness and knowledge consistency. In practical applications, different types of power generation tasks have significant differences in response methods, knowledge depth, and language style. How to intelligently select and efficiently schedule among multiple pre-trained models has become a core bottleneck restricting the widespread deployment of intelligent governance systems in the industry.

[0004] Existing methods based on keyword matching and rule retrieval, general large language models directly applied to power generation industry scenarios, and semantic retrieval and knowledge enhancement methods based on vectorized embedding generally suffer from insufficient semantic understanding, inflexible model selection, and inaccurate knowledge retrieval. These methods are difficult to meet the accuracy and professionalism requirements of diverse tasks, and are also detrimental to the management of data assets.

[0005] It should be noted that the information disclosed in the background section above is only used to enhance the understanding of the background of this application. Summary of the Invention

[0006] To provide a basic understanding of some aspects of the disclosed embodiments, a brief summary is given below. This summary is not intended as a general commentary, nor is it intended to identify key / important components or describe the scope of protection of these embodiments, but rather as a prelude to the detailed description that follows.

[0007] The data intelligent governance method, device, equipment, and storage medium for the power generation industry based on large models provided in this disclosure, combined with key aspects such as semantic modeling, model adaptation, and knowledge retrieval, solves the problems of insufficient semantic understanding, inflexible model selection, and inaccurate knowledge retrieval in existing large models, thereby improving the system's responsiveness in complex business environments.

[0008] The data-driven intelligent governance method for the power generation industry based on a large model, as disclosed in this embodiment, includes:

[0009] We collected structured, semi-structured, and unstructured enterprise corpora, combined them with historical data diagnostic records and knowledge fragments, constructed a multi-source heterogeneous corpus dataset, and fine-tuned the training on data from different industry sub-domains based on a general pre-trained language model, forming multiple vertical large language models for the power generation industry with business adaptability.

[0010] For the power generation business data provided by users, a multi-perspective semantic representation structure is constructed. The optimal target large model, the Top-K candidate model set and their semantic embedding vectors are output through a supervised learning model for subsequent knowledge matching and data diagnostic content generation.

[0011] The graph structure semantic representation is constructed using key features extracted from power generation data. The output embedding vector, the identifier corresponding to the target large model, and the Top-K candidate model information are integrated. The graph neural network is used to achieve accurate matching of structured knowledge fragments and output knowledge diagnostic content for reference.

[0012] Based on the target large model, the raw power generation data and the matched knowledge text are jointly input, and data analysis and governance suggestions are generated by constructing prompt templates. At the same time, user feedback information is collected to optimize the model selector and knowledge matching module respectively, and user profiles are constructed to realize the dynamic updating and personalized optimization of the system.

[0013] This disclosure discloses a data-driven intelligent governance device for the power generation industry based on a large-scale model. The device includes:

[0014] The generation module is used to collect structured, semi-structured and unstructured enterprise corpora, combine historical data diagnostic records and knowledge fragments to construct a multi-source heterogeneous corpus dataset, and fine-tune and train it on data from different industry sub-domains based on a general pre-trained language model to form multiple vertical large language models for the power generation industry with business adaptability.

[0015] The data processing module is used to construct a multi-perspective semantic representation structure for the power generation business data provided by users. It outputs the optimal target large model, the Top-K candidate model set and their semantic embedding vectors through a supervised learning model, which are used for subsequent knowledge matching and data diagnostic content generation.

[0016] The output module is used to construct a graph structure semantic representation based on key features extracted from power generation data, and to integrate the output embedding vector, the identifier corresponding to the target large model, and the Top-K candidate model information. Through graph neural networks, it achieves accurate matching of structured knowledge fragments and outputs knowledge diagnostic content for reference.

[0017] The training module is used to input user questions and knowledge text together based on the target model and Top-K candidate model, and generate target content through prompts; at the same time, it collects user feedback information, optimizes the model selector and knowledge matching module respectively, builds user behavior profiles, and realizes dynamic system updates and personalized optimization.

[0018] The electronic device provided in this disclosure includes at least one processor;

[0019] and a memory communicatively connected to the at least one processor;

[0020] The memory stores instructions that can be executed by the at least one processor, which enables the at least one processor to execute the aforementioned data-intelligent governance method for the power generation industry based on a large model.

[0021] The storage medium provided in this embodiment stores program instructions, which, when executed, perform the above-described intelligent data governance method for the power generation industry based on a large model.

[0022] The data intelligent governance method, apparatus, equipment, and storage medium for the power generation industry based on a large model provided in this disclosure can achieve the following technical effects:

[0023] The data intelligence governance method for the power generation industry based on large models disclosed herein constructs a semantic modeling method that integrates task intent, scenario affiliation and model capability features. It constructs a unified high-dimensional semantic expression to simultaneously predict the optimal target model and Top-K candidate models, which significantly improves the accuracy and robustness of model selection and meets the needs of precise model scheduling for different types of power data and task intents.

[0024] Furthermore, for the large number of structured procedures, operation and maintenance knowledge and scheduling logic in the power generation industry, a knowledge matching method combining keyword graph structure modeling and model label semantic fusion is proposed; graph neural networks are used to achieve accurate positioning and retrieval of structured knowledge fragments, thereby improving the accuracy and consistency of knowledge retrieval in tasks with complex professional terminology and strong semantic context dependencies.

[0025] Furthermore, a knowledge-enhanced content generation method based on prompt templates is proposed. This method inputs data semantics and structured knowledge fragments into the target model to guide the target model in generating task response content. At the same time, user feedback data is introduced to continuously optimize the model selection and knowledge matching process, thereby adapting to the requirements of accuracy and interpretability of response content in practical applications in the power generation industry. This solves the problems of insufficient semantic understanding, inflexible model selection, and inaccurate knowledge retrieval in existing solutions, meeting the accuracy and professionalism requirements of diverse tasks and facilitating the management of data assets.

[0026] The above general description and the description below are exemplary and illustrative only and are not intended to limit this application. Attached Figure Description

[0027] One or more embodiments are illustrated by way of example with reference to the accompanying drawings. These illustrations and drawings do not constitute a limitation on the embodiments. Elements having the same reference numerals in the drawings are shown as similar elements. The drawings are not to be scaled. And wherein:

[0028] Figure 1 This is a flowchart illustrating a data intelligence governance method for the power generation industry based on a large model, provided in an embodiment of this disclosure.

[0029] Figure 2 This is a flowchart of the overall technical route for a data-driven intelligent governance method for the power generation industry based on a large model, as provided in this embodiment of the disclosure.

[0030] Figure 3 This is a data-driven large model adaptive selection module network structure diagram provided in an embodiment of the present disclosure;

[0031] Figure 4 This is a network structure diagram of a knowledge fragment fine matching module that integrates semantics and candidate features, provided in an embodiment of this disclosure;

[0032] Figure 5 This is a schematic diagram showing the comparison of Micro-F1 Score evaluation results of different models in eight categories of electron emission fields provided by embodiments of this disclosure;

[0033] Figure 6 This is a schematic diagram showing the comparison of MRR evaluation results of different models in eight types of power generation fields provided by the embodiments of this disclosure;

[0034] Figure 7 This is a schematic diagram of the structure of a data intelligent governance device for the power generation industry based on a large model, provided in an embodiment of this disclosure;

[0035] Figure 8This is a schematic diagram of the structure of a data intelligent governance device for the power generation industry based on a large model, provided in an embodiment of this disclosure. Detailed Implementation

[0036] To provide a more detailed understanding of the features and technical content of the embodiments of this disclosure, the implementation of the embodiments of this disclosure will be described in detail below with reference to the accompanying drawings. The accompanying drawings are for illustrative purposes only and are not intended to limit the embodiments of this disclosure. In the following technical description, for ease of explanation, several details are used to provide a full understanding of the disclosed embodiments. However, one or more embodiments may still be implemented without these details. In other cases, well-known structures and devices may be simplified in their depiction to simplify the drawings.

[0037] The terms "first," "second," etc., used in the embodiments of this disclosure are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate for the embodiments of this disclosure described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion.

[0038] Unless otherwise stated, the term "multiple" means two or more.

[0039] In this embodiment of the disclosure, the character " / " indicates that the objects before and after it are in an "or" relationship. For example, A / B means: A or B.

[0040] The term "and / or" describes an association between objects, indicating that three relationships can exist. For example, A and / or B means: A or B, or A and B.

[0041] The term "correspondence" can refer to an association or binding relationship. The correspondence between A and B means that there is an association or binding relationship between A and B.

[0042] The following description, in conjunction with the accompanying drawings, illustrates the data intelligent governance method, apparatus, equipment, and storage medium for the power generation industry based on a large model, as provided in the embodiments of this disclosure.

[0043] Figure 1 This is a flowchart illustrating a data intelligence governance method for the power generation industry based on a large model, provided in an embodiment of this disclosure.

[0044] like Figure 1 As shown, data-driven intelligent governance methods for the power generation industry based on large-scale models may include:

[0045] S101 collects structured, semi-structured, and unstructured enterprise corpora, combines historical data diagnostic records and knowledge fragments to construct a multi-source heterogeneous corpus dataset, and fine-tunes the training on data from different industry sub-domains based on a general pre-trained language model to form multiple vertical large language models for the power generation industry with business adaptability.

[0046] S102, for the power generation business data provided by the user, construct a multi-view semantic representation structure, and output the optimal target large model, Top-K candidate model set and its semantic embedding vector through a supervised learning model, which are used for subsequent knowledge matching and data diagnostic content generation;

[0047] S103 constructs a graph structure semantic representation based on key features extracted from power generation data, and integrates the output embedding vector, the identifier corresponding to the target large model, and Top-K candidate model information. Through graph neural networks, it achieves accurate matching of structured knowledge fragments and outputs knowledge diagnostic content for reference.

[0048] S104, based on the target large model, inputs raw power generation data and matched knowledge text together, and generates data analysis and governance suggestions by constructing prompt templates; at the same time, it collects user feedback information to optimize the model selector and knowledge matching module respectively, and constructs user profiles to realize the dynamic updating and personalized optimization of the system.

[0049] In some embodiments, the aforementioned collected structured, semi-structured, and unstructured enterprise corpora, combined with historical data diagnostic records and knowledge fragments, construct a multi-source heterogeneous corpus dataset. Based on a general pre-trained language model, fine-tuning training is performed on data from different industry sub-domains to form multiple vertical large-scale language models for the power generation industry with business adaptability, including:

[0050] Construct a multi-source corpus dataset covering multiple core business modules of the power generation industry, wherein the multi-source corpus dataset includes equipment operation management, dispatch instruction execution, safety and compliance supervision, technical support services, and energy regulation and operation and maintenance management;

[0051] Based on the initial text data Text n Define or filter business requirement text data Q n The corresponding answer text A was labeled by professionals. n ;

[0052] Based on the constructed business requirement text data Q n Based on this, experts were organized to label the corresponding optimal fit large language model and Top-K candidate model labels S. n This forms structured input-output pairs;

[0053] The basic model ChatGLM2-6B was chosen as the starting point for pre-training for subsequent fine-tuning in the vertical domain.

[0054] Based on the constructed multi-source heterogeneous corpus dataset and the selected base model ChatGLM2-6B, the models for each electronic domain were fine-tuned and optimized to form multiple vertical large language models with domain adaptability.

[0055] In some embodiments, the above-mentioned multi-view semantic representation structure is constructed based on the power generation business data provided by the user. A supervised learning model outputs the optimal target large model, a Top-K candidate model set, and their semantic embedding vectors for subsequent knowledge matching and data diagnostic content generation, including:

[0056] To address three sub-objectives—data type identification, business scenario attribution judgment, and model capability matching—a multi-perspective semantic representation structure adapted to the needs of complex tasks is constructed.

[0057] A multi-perspective semantic fusion mechanism is designed to deeply fuse the feature vectors obtained from the paths corresponding to the three types of sub-objectives, in order to construct a high-order problem semantic expression with multiple semantic dimensions of task understanding, scene attribution and capability adaptation.

[0058] In the fusion semantic embedding F Emb Based on this, a two-branch model prediction structure is designed, which includes a prediction branch for the Top-K candidate model set and a prediction branch for the best target model.

[0059] Based on the constructed optimally adapted large language model labels and the corresponding dataset, a joint multi-task supervised training strategy is adopted to optimize the two sub-tasks of Top-K prediction and target model classification.

[0060] In some embodiments, the key features extracted from power generation data are used to construct a graph structure semantic representation, and the output embedding vector, the identifier corresponding to the target large model, and Top-K candidate model information are fused together. A graph neural network is then used to achieve accurate matching of structured knowledge fragments, outputting knowledge diagnostic content for reference, including:

[0061] Employing a composite rule based on TF-IDF and part-of-speech tagging, from business data text Q n Extract the weighted keywords that meet the preset conditions and construct a keyword set: Kw1, Kw2, ..., Kw M Among them, Kw m Indicates user question Q n The m-th keyword in the text, where m∈[1,M]; the keyword is converted into corresponding word vectors Kf1,Kf2,...,Kf using a pre-trained Word2Vec model. M ;

[0062] The design uses a graph-based data construction method, treating each keyword as a node and its feature vector as a node feature. Cosine similarity is used to calculate the pairwise similarity between node features. When the similarity exceeds a preset threshold, the edge is considered. th At this point, an edge is constructed between the two nodes, and the similarity calculation value is used as the feature of the edge. The final constructed graph data is Graph(V,E), where V represents the set of all nodes, and the node feature corresponding to the i-th node is n. i E represents the set of all edges, and the edge between the i-th node and the j-th node has the characteristic e. ij ;

[0063] The constructed graph data Graph(V,E) is input into a four-layer stacked graph Transformer encoder, each layer including a multi-head graph attention mechanism, a feedforward network, and a global average pooling operation;

[0064] The target model label output by the trained large model adaptive selector. and the predicted Top-K candidate set The concatenation is then performed, followed by input into a two-layer fully connected ReLU network for high-order feature extraction, resulting in the scheduling label semantic features F. label ;

[0065] The main channel output feature F graph With auxiliary channel feature F label F Emb Channel splicing is performed to form a fusion feature F Con ; Fuse features F Con The higher-order representations are extracted from two sets of MLP-ReLU layers, and a Dropout layer is added to prevent overfitting. Finally, a sigmoid activation function is used for multi-label classification, and the confidence prediction vector of the knowledge fragment index label is output. It also calls the corresponding structured knowledge matching text based on the prediction results of the knowledge fragment index tags.

[0066] Binary Cross-Entropy is used as the multi-label supervised loss function to optimize the model parameters. The training process is combined with an early stopping mechanism and the Adam optimizer to ensure that the model achieves the best balance between generalization performance and matching accuracy. Finally, the model can locate the most relevant knowledge fragments in the structured knowledge base for the input question, thereby improving the relevance and contextual consistency of the downstream analysis results.

[0067] In some embodiments, based on the target large model, the raw power generation data and matched knowledge text are jointly input, and data analysis and governance suggestions are generated by constructing prompt templates; simultaneously, user feedback information is collected to optimize the model selector and knowledge matching module respectively, and user profiles are constructed to achieve dynamic updates and personalized optimization of the system, including:

[0068] The optimal target large model is selected. To generate the model, text matching data from the knowledge base is introduced. As a knowledge enhancement prompt, it compensates for the knowledge omissions and fuzzy expressions in the model;

[0069] Based on users' subjective ratings of the model scheduling accuracy, Ev n A model selector tuning path based on user feedback ratings is constructed. After receiving the system output, users evaluate the accuracy of model selection based on a rating mechanism that includes scheduling rationality, semantic understanding, and business style suitability, and obtain an evaluation score Ev. n ;

[0070] Based on user feedback text content, a semantically driven graph construction strategy optimization mechanism is established. If the target tag appears in the user feedback text content, the system identifies relevant entities and relationships based on the target tag, thus preventing deviations in the construction of structured knowledge graphs.

[0071] For each user, we maintain historical data types, model call preferences, and knowledge matching habits and behavioral characteristics, and realize the semantic mapping between user behavior patterns and model strategies by constructing a semantic similarity matching matrix.

[0072] In some embodiments, the above-described construction of a multi-source corpus dataset covering multiple core business modules of the power generation industry includes equipment operation management, dispatch instruction execution, safety and compliance supervision, technical support services, and energy regulation and operation and maintenance management, including:

[0073] Initial business text data was collected from multiple information sources, including power plant management systems, monitoring and maintenance platforms, safety manuals, technical guidelines, expert response systems, and duty work order databases, forming raw corpora from multiple vertical domains: Text1, Text2, ..., Text N , where Text n Let n represent the initial text data of the nth domain, where n∈[1,N];

[0074] Text n Divided into multiple sub-segments and This represents a sub-segment of text containing T characters; each segment The first 1 to t words are used as input The last t+1 to T words are output. Then, corresponding operations are performed on all sub-segments of data to ultimately form a domain autoregressive dataset for all vertical domains;

[0075] Based on the initial text data Text n Define or filter business requirement text data Q n This was used as input to a domain-specific question-and-answer dataset; and the corresponding answer text A was labeled by professionals. n This is used as the output of a domain question-answering dataset to create domain question-answering datasets for all vertical domains.

[0076] In some embodiments, the above-mentioned three sub-objectives—data type identification, business scenario attribution judgment, and model capability matching—are used to construct a multi-perspective semantic representation structure that adapts to the needs of complex tasks, including:

[0077] Using a pre-trained bert-base-chinese model to process business data text data Q n Encode the sequence of context-dependent word vectors F. base ;

[0078] The data vector F is processed through a five-layer cross-attention mechanism. glo With scene category vector Layer-by-layer alignment modeling is used to capture semantic relevance, and the output of each layer is processed by the ReLU activation function to enhance the semantic nonlinearity expression capability.

[0079] Using the chinese-roberta-wwm-ext model to process business data text data Q n Perform sentence-level modeling to obtain the initial semantic embedding vector F glo Construct a set of business scenario category vectors Let n represent the category vector in the nth scene domain, where n∈[1,N];

[0080] The data vector F is processed through a five-layer cross-attention mechanism. glo With scene category vector Layer-by-layer alignment modeling is used to capture semantic relevance. The output of each layer is processed by the ReLU activation function to enhance the semantic nonlinearity representation capability; the final output scene feature vector F Sap Characterize the semantic attribution features of data within the business structure;

[0081] The pre-trained model Qwen1.5-0.5B, which has instruction understanding capabilities, was used to process the data text Q. n Embedding is performed to extract semantic representation vector F with operation-oriented features. taskPreserve the original semantic information of the data at the structural and logical intent level;

[0082] The task semantic representation vector F task The data is input into a five-layer stacked Transformer encoder structure, which uses a multi-head attention mechanism to model different combinations of data elements and their preferences for different models. Each Transformer output is followed by a ReLU activation function to ultimately obtain the fitting feature vector F representing the data-model matching. Mat This is used in the subsequent semantic fusion process.

[0083] Figure 2 This is a flowchart illustrating the overall technical roadmap for a data-driven intelligent governance method for the power generation industry based on a large model, as provided in this disclosure. Figure 3 This is a data-driven large model adaptive selection module network structure diagram provided in an embodiment of the present disclosure. Figure 4 This is a network structure diagram of a knowledge fragment fine matching module that integrates semantics and candidate features, provided in an embodiment of this disclosure. Figure 5 This disclosure provides a comparison of Micro-F1Score evaluation results for different models in eight electron-generating domains. Figure 6 This disclosure provides a comparison of MRR evaluation results for different models in eight types of power generation fields, combined with... Figures 2 to 6 ,right Figure 1 The paper further describes a data-driven intelligent governance method for the power generation industry based on a large model.

[0084] Specifically, Figure 1 The data-driven intelligent governance method for the power generation industry based on large-scale models can further include:

[0085] S1, Construction of Multi-Source Power Generation Industry Dataset and Training of Vertical Domain Large Model: Collect structured, semi-structured and unstructured power generation industry business data, combine historical data diagnostic records and knowledge fragments to construct a multi-source heterogeneous corpus dataset; and based on a general pre-trained language model, fine-tune the training on data from different industry sub-domains to form multiple vertical large language models for the power generation industry with business adaptability.

[0086] S2, Data-driven large model adaptive selection module construction: For the power generation business data provided by users, construct a multi-view semantic representation structure, and output the optimal target large model, Top-K candidate model set and its semantic embedding vector through supervised learning model, which are used for subsequent knowledge matching and data diagnostic content generation;

[0087] S3, the knowledge fragment fine matching module that integrates semantics and candidate features: constructs a graph structure semantic representation based on key features extracted from power generation data, and integrates the semantic vector output by S2, target model identifier and Top-K candidate model information, and achieves accurate matching of structured knowledge fragments through graph neural network, outputting knowledge diagnostic content that can be referenced;

[0088] S4, the construction of a knowledge-enhanced large model content generation and optimization module: Based on the target large model selected in S2, the raw power generation data and the knowledge text matched in S3 are jointly input, and data analysis and governance suggestions are generated by constructing prompt templates; at the same time, user feedback information is collected to optimize the model selector and knowledge matching module respectively, and user profiles are constructed to realize the dynamic updating and personalized optimization of the system.

[0089] In some embodiments, the process of constructing a multi-source power generation industry dataset and training a large vertical domain model in S1 includes:

[0090] To achieve semantic understanding and governance analysis capabilities for various business data scenarios in the power generation industry, this invention first constructs a multi-source heterogeneous power generation data corpus and then conducts fine-tuning training of a large-scale model for vertical domains based on a general pre-trained language model. Simultaneously, to support subsequent adaptive selection of large models and precise matching of knowledge fragments, the system synchronously constructs a model scheduling training dataset and a knowledge label matching sample set. Specifically, the following steps are included:

[0091] S1-1, Construction of Multi-Source Power Generation Industry Dataset: To realize the ability of the large language model in this invention to understand multi-task scenarios in power generation business, a multi-source corpus dataset covering multiple core business modules of the power generation industry is first constructed. Data collection covers, but is not limited to, the following sub-domains: equipment operation management, dispatch instruction execution, safety and compliance supervision, technical support services, energy regulation and operation and maintenance management, etc. The following data collection process is executed for each sub-domain:

[0092] (1) Initial business text data was collected from multiple information sources, including power plant management systems, monitoring and maintenance platforms, safety manuals, technical guidelines, expert response systems, and duty work order databases, forming raw corpus sets Text1, Text2, ..., Text1 from multiple vertical domains. N Text n Let n represent the initial text data of the nth domain, where n∈[1,N];

[0093] (2) Based on the original text data n Construct a domain autoregressive dataset: Text n Divided into multiple sub-segments and This represents a sub-segment of text containing T characters; each segment The first 1 to t words are used as input. The next t+1 to T words are output. This operation is then performed on all sub-segments of data, ultimately forming a domain autoregressive dataset for all vertical domains;

[0094] (3) Based on the initial text data n Create a domain-specific question-and-answer dataset; specifically, based on the initial text data Text... n Define or filter business requirement text data Q n This was used as input to a domain-specific question-and-answer dataset; and the corresponding answer text A was labeled by professionals. n This is used as the output of the domain question-answering dataset; domain question-answering datasets are then created for all vertical domains based on this method.

[0095] S1-2, Large Model Adaptive Selection Dataset Construction: To support the large model scheduling mechanism in module S2 for different power generation data tasks, a multi-label classification training dataset needs to be constructed, covering various model adaptation scenarios. Specifically, the business requirement text data Q constructed in S1-1... n Based on this, experts were organized to label the corresponding optimal fit large language model (target model ID). n ) and Top-K candidate model labels S n This forms structured input-output pairs. The training set covers multiple subdomains of power generation operations, thereby improving the accuracy and response stability of large-scale model scheduling.

[0096] S1-3, Knowledge Fragment Fine Matching Dataset Construction: To train the knowledge matcher in module S3, a dataset capable of associating power generation business semantics with structured knowledge items is constructed; specifically, the power generation business demand text data Q generated in S1-1 is used. n Compared with the large model labels (including IDs) determined in S1-2 n and S n Based on this, further label K is used to annotate knowledge fragments that are strongly related to the semantics of the data. n This forms a sample set for matching business data and knowledge fragments.

[0097] Therefore, the input for each sample includes business requirement text data Q. n Target model identifier ID n Top-K candidate model S n The output includes the index label K of the matched structured knowledge fragment. n Through this construction process, a high-quality knowledge matching dataset is formed for multiple types of knowledge bases, which is used to support the knowledge accuracy and task relevance of the content generation process.

[0098] S1-4, Selection of pre-trained large model: In order to support the understanding and governance analysis of multi-business data in the power generation industry, a suitable pre-trained basic model is selected. Taking into account indicators such as parameter size, semantic modeling ability, and inference efficiency, ChatGLM2-6B is selected as the basic model and is used as the pre-trained basic model for fine-tuning training of various vertical business models.

[0099] S1-5, Fine-tuning Training of Large-Scale Models in Vertical Domains: Based on the dataset constructed in S1-1 and the pre-trained large-scale models in S1-4, fine-tuning and optimization are performed on the models for each electronics domain to improve their task adaptability in specific business scenarios. The training process includes the following two steps:

[0100] (1) Autoregressive Training Stage: Based on the basic language modeling structure and model parameters of the original pre-trained ChatGLM2-6B model, the autoregressive datasets of each vertical domain in S1-1 are used as training datasets. Through continued training, the transfer learning of language patterns in each domain is completed, enabling the model to master common expressions and business logic structures in the power industry, and obtaining the autoregressive training large model parameters AM1, AM2,..., AM for N vertical domains. N ;

[0101] (2) Task adaptation fine-tuning stage: Based on autoregressive training, large model parameters AM1, AM2,..., AM N Furthermore, using domain-specific question-answering datasets as training samples, supervised learning is employed to fine-tune the model, enabling it to respond to specific task scenarios. Ultimately, N large-scale vertical models QAM1, QAM2, ..., QAM are obtained, each with business customization capabilities. N ;

[0102] The trained model parameters QAM1, QAM2, ..., QAM are obtained through the above two-stage fine-tuning process. N These models were then applied to the ChatGLM2-6B model, ultimately generating large language models Model1, Model2, ..., Model1, each with corresponding task adaptability for each vertical domain. N .

[0103] In some embodiments, the construction process of the data-driven large model adaptive selection module in S2 includes:

[0104] To achieve intelligent scheduling and dynamic adaptation of large language models for diverse business data in the power generation industry, this invention constructs a large model adaptive selection module based on multi-perspective semantic modeling and deep ensemble reasoning. This module automatically predicts the optimal target large model and the Top-K candidate model set, and outputs semantic embedding vectors as the semantic basis for subsequent knowledge matching in the S3 module. This module includes multi-perspective semantic embedding, semantic fusion processing, model prediction, and semantic embedding; its overall structure is as follows: Figure 2 As shown. Specifically, it includes the following steps:

[0105] S2-1, Multi-perspective semantic embedding: For a comprehensive understanding of power generation business text data Q n To address the expressive features across different semantic dimensions, this module designs three semantic embedding paths, each targeting a sub-objective: data type identification, business scenario attribution judgment, and model capability matching. This constructs a multi-perspective semantic representation structure adaptable to complex task requirements, specifically including:

[0106] (1) Data Intent Identification Path: This path aims to identify the structural attributes and task intent of power generation data, such as "operation monitoring," "energy consumption analysis," and "fault diagnosis." This path uses a pre-trained bert-base-chinese model to analyze the business data text data Q. n Encode the sequence of context-dependent word vectors F. base ;

[0107] Subsequently, since gated recurrent neural networks (GRUs) have strong memory capabilities and are suitable for capturing long-term dependency features of data, and bidirectional design can enhance the model's sensitivity to the association between preceding and following words, and can simultaneously model causal structures and modification dependencies, this approach chooses to use F... base The input is fed into a five-layer stacked bidirectional GRU for contextual semantic modeling, and the output of each layer of bidirectional GRU is followed by a ReLU activation function to enhance semantic discriminability; the final output is the intent feature vector F. Irp It can effectively represent the task attributes and operational objectives of the data;

[0108] (2) Scene Awareness Path: This path aims to identify the business module to which the power generation business data belongs, such as "operation shift", "equipment maintenance", "energy consumption monitoring", "dispatch approval", etc., so as to provide contextual basis for subsequent domain model matching; firstly, the chinese-roberta-wwm-ext model is used to analyze the business data text data Q. n Perform sentence-level modeling to obtain the initial semantic embedding vector F glo Subsequently, a set of business scenario category vectors are constructed. (e.g., equipment, scheduling, security, compliance, etc.), which represents the category vector in the i-th scenario domain, and n∈[1,N];

[0109] Subsequently, the data vector F is processed through a five-layer cross-attention mechanism. glo With scene category vector Layer-by-layer alignment modeling is used to capture semantic relevance. The output of each layer is processed by the ReLU activation function to enhance the semantic nonlinearity representation capability; the final output scene feature vector F Sap It can characterize the semantic attribution of data within the business structure.

[0110] Model adaptation path: This path is used to model the power generation business data Q. n The capability adaptation relationship between the model and the large language model helps determine which model should be used to analyze and process the current data. First, a pre-trained model Qwen1.5-0.5B with instruction understanding capabilities is used to process the data text Q. n Embedding is performed to extract semantic representation vector F with operation-oriented features. task Preserve the original semantic information of the data at the structural and logical intent level;

[0111] Subsequently, the semantic representation vector F of the task task The data is input into a five-layer stacked Transformer encoder structure, which uses a multi-head attention mechanism to model different combinations of data elements and their preferences for different models. Each Transformer output is followed by a ReLU activation function to ultimately obtain the fitting feature vector F representing the data-model matching. Mat This is used in the subsequent semantic fusion process;

[0112] S2-2, Multi-view semantic fusion: Design a multi-view semantic fusion mechanism to process the feature vectors (intent features F) obtained from the three sub-paths in S2-1. Irp Scene features F Sap Adaptation feature F Mat Deep integration is carried out to construct a high-order problem semantic expression with multiple semantic dimensions such as task understanding, scenario attribution and capability adaptation;

[0113] Specifically, the output vectors of the three paths mentioned above are concatenated to form the initial fusion representation F. Mix Subsequently, semantically selective integration is performed based on a gating fusion mechanism, which involves introducing three learnable gating parameters g. Irp g Sap g Mat This achieves dynamic trade-offs and suppression of semantic information from different paths, and the fused representation is further normalized through a normalization layer to obtain the final high-dimensional semantic embedding vector F. Emb ;

[0114] S2-3, Model Prediction and Semantic Embedding: To achieve accurate scheduling of large language models, this module integrates semantic embedding... Emb Based on this, a two-branch model prediction structure is designed; firstly, a prediction branch for the Top-K candidate model set; secondly, a prediction branch for the optimal target model; specifically, F Emb First, the embedding vector is fed into a fully connected layer for dimensionality compression. Then, the embedding vector is fed into two branches in parallel:

[0115] (1) Top-K candidate prediction branch: Using a multi-label binary classification structure, the Sigmoid activation function is used to predict the independent fit probability between the current business data and each candidate model, and the output is the confidence distribution vector of the candidate large model.

[0116] (2) Target Model Identification Branch: The Softmax activation function is used to normalize and classify all candidate models, and the best target model label is output. This branch ensures that the target model with the highest fit is selected from the Top-K result set.

[0117] In addition, to balance scheduling accuracy and semantic interpretability, the system will use F Emb As the final question semantic embedding vector output, it is provided to the S3 module for deep semantic matching with structured fragments in the knowledge base;

[0118] S2-4, Model Training: Based on the large model built in S1-2, the dataset is adaptively selected. This module adopts a joint multi-task supervised training strategy, and optimizes the two sub-tasks of Top-K prediction and target model classification at the same time.

[0119] Specifically, the Binary Cross-Entropy (BCE) loss is used to measure the Top-K multi-label prediction error, and the Categorical Cross-Entropy (CCE) loss is used to measure the single-label classification error of the target model. In addition, the Adam optimizer is used to update parameters during the optimization process, and the maximum number of iterations and early stopping strategy are set. Finally, a trained large model adaptive selector is obtained.

[0120] In some embodiments, the construction process of the knowledge fragment fine matching module that integrates semantics and candidate features in S3 includes:

[0121] To achieve accurate retrieval of structured knowledge fragments within the semantic context of complex power generation business data, this invention designs a knowledge fragment precision matching module that integrates multi-source semantic representation and model scheduling features. This module uses the main semantic channel driven by business data as its core channel, combining a large model scheduling label channel and a semantic embedding vector channel to construct a multi-modal feature fusion mechanism. Ultimately, it achieves the prediction of high-quality knowledge fragment index labels, providing accurate support for structured knowledge retrieval and intelligent analysis conclusion generation. The overall structure of this module is as follows: Figure 4 As shown, the specific steps include:

[0122] S3-1, Business Keyword Graph Structure Modeling: This module serves as a model for business data text Q in the power generation industry. n As the main input channel, it aims to capture the semantic relationships and structural dependencies between keywords within it; firstly, it employs a composite rule based on TF-IDF and part-of-speech tagging (POS) to extract data from business data text Q. n Extract high-weight keywords and construct a keyword set: Kw1, Kw2, ..., Kw M Kw m The text Q represents business data. n The m-th keyword in the text, where m∈[1,M]; subsequently, the keyword is converted into corresponding word vectors Kf1,Kf2,...,Kf using a pre-trained Word2Vec model. M ;

[0123] Based on this, a graph-structured data construction method is designed, where each keyword is treated as a node, and the word vector of each keyword is used as a node feature. The similarity between node features is calculated pairwise using cosine similarity. When the similarity exceeds a preset threshold, the edge is selected. th At this point, an edge is constructed between the two nodes, and the similarity calculation value is used as the feature of the edge. The final constructed graph data is Graph(V,E), where V represents the set of all nodes, and the node feature corresponding to the i-th node is n. i E represents the set of all edges, and the edge between the i-th node and the j-th node has the characteristic e. ij ;

[0124] Next, the constructed graph data Graph(V,E) is passed through a four-layer stacked graph Transformer encoder, each layer including a multi-head graph attention mechanism, a feedforward network, and a global average pooling operation. This structure can capture high-order semantic relationships between keywords in business data, and finally outputs the main channel graph semantic feature vector F. graph ;

[0125] S3-2, Construction of Auxiliary Semantic Channels: To enhance the main channel's ability to model the semantics of power generation data, the system introduces two auxiliary semantic channels:

[0126] (1) Large model selection scheduling channel: The target model labels output by the large model adaptive selector trained by S2 are selected. and the predicted Top-K candidate set The concatenation is then performed, followed by input into a two-layer fully connected ReLU network for high-order feature extraction, resulting in the scheduling label semantic features F. label ;

[0127] (2) Problem semantic embedding vector channel: directly adopt the multi-view semantic fusion vector F output by the S2 module. Emb It represents the semantic expression characteristics of global business data;

[0128] S3-3, Multi-channel fusion and index label prediction: The main channel output feature F... graph With auxiliary channel feature F label F Emb Channel splicing is performed to form a fusion feature F Con The fused features are input into two sets of MLP (Multi-Layer Perceptron)-ReLU layers to extract higher-order representations, followed by a Dropout layer to prevent overfitting, and finally, a sigmoid activation function is used for multi-label classification, outputting a confidence prediction vector for the knowledge fragment index label. It also calls the corresponding structured knowledge matching text based on the prediction results of the knowledge fragment index tags. To support subsequent diagnostic analysis tasks;

[0129] S3-4, Model Training and Optimization: During the training phase, based on the knowledge fragment matching dataset constructed in S1-3, Binary Cross-Entropy (BCE) is used as the multi-label supervised loss function to optimize the parameters of the S3 model. The training process incorporates an early stopping mechanism and the Adam optimizer to ensure the model achieves the best balance between generalization performance and matching accuracy. The final trained model can quickly locate the most semantically relevant knowledge fragments in a structured knowledge base for any type of power generation business data, thereby significantly improving the relevance and contextual consistency of downstream analysis results.

[0130] In some embodiments, the construction process of the knowledge-enhanced large model content generation and optimization module in S4 includes:

[0131] To achieve accurate response generation of power generation business data by a large language model driven by structured knowledge, and to implement a feedback-based continuous optimization mechanism, this invention constructs a dynamic scheduling and content generation module that combines a knowledge-enhanced generation mechanism with a feedback optimization path. This module addresses the semantic guidance and quality improvement goals under complex governance tasks in the power generation industry, integrating model optimization mechanisms and personalized behavior modeling mechanisms to enhance the accuracy, business relevance, and system adaptability of content generation. Specifically, it includes the following steps:

[0132] S4-1, Construction of a Knowledge-Enhanced Large-Scale Model Content Generation Module: To improve the accuracy and knowledge support capabilities of the generated results in tasks such as governance recommendations, anomaly detection, and trend analysis, this module proposes a knowledge-enhanced large-scale language model content generation mechanism. This mechanism uses the optimal target large-scale model selected in S2. To generate the model, text matching data from the knowledge base generated by S3 is introduced. As a knowledge enhancement prompt, it can compensate for any knowledge gaps or misunderstandings that may exist in the basic model;

[0133] Specifically, the system first processes the original power generation business data text Q. n Knowledge fragment text generated by the S3 module A structured assembly process is performed to construct an enhanced Prompt template that includes problem context and knowledge supplementation. This template serves as the input to the large model to guide the generation process, enabling the model to have clear knowledge basis and semantic control capabilities during content output. The final generated response results are used to assist in completing governance tasks such as operational data diagnosis, strategy suggestions, and risk warnings.

[0134] S4-2, User Feedback-Based Model Selector (S2) Optimization Mechanism: Based on the user's subjective rating information Ev regarding the accuracy of model scheduling. n This invention constructs a model selector tuning path based on user feedback ratings. After receiving the system output, users evaluate the accuracy of model selection based on a rating mechanism that considers scheduling rationality, semantic understanding, and business style suitability, and obtain an evaluation score Ev. n ;

[0135] The scoring signals are then mapped to supervised training samples and updated and stored in the feedback training pool. The large model selection network model in S2 is fine-tuned periodically using a reinforcement supervision strategy. Secondly, the prediction weights of the Top-K candidate set and the final target model are corrected based on the scoring deviation, thereby improving the adaptability of the model selector under different task preferences.

[0136] S4-3, Optimization mechanism for knowledge matching model (S3) based on user feedback: In order to improve the fit and support strength of knowledge fragment matching results to power generation business data, this module establishes a semantic-driven graph construction strategy optimization mechanism for user feedback text content; if the user feedback content contains tags such as "incomplete knowledge", "weak analysis" or "lack of evidence for results", the system will automatically identify such text information and trigger the graph structure update mechanism for the S3 module.

[0137] Specifically, the system dynamically adjusts the edge construction threshold of the graph based on the prompts and correction suggestions extracted from the feedback. th This allows for adjustments to the semantic density and structural precision of the problem graph data representation; and simultaneously, feedback problem samples are added to the graph structure optimization training set for fine-tuning the graph semantic encoder, enabling it to generate index labels that are closer to the requirements in subsequent problem processing.

[0138] S4-4, Personalized Behavior Modeling and Dynamic System Evolution: To build an intelligent governance system with continuous memory and behavioral adaptability, this module establishes a user behavior modeling mechanism to record user interaction behaviors and feedback information at different stages to generate dynamic user profiles. Specifically, the system maintains behavioral characteristics such as historical data types, model invocation preferences, and knowledge matching habits for each user; and by constructing a semantic similarity matching matrix, it realizes the semantic mapping between user behavior patterns and model strategies, which is used for scheduling prior modeling and personalized strategy invocation in future tasks.

[0139] In addition, the system supports a scheduling priority adjustment mechanism based on users' historical performance, which enables high-satisfaction models to be called first in similar tasks in the future, thereby achieving intelligent evolution and behavior enhancement at the system level.

[0140] Regarding experimental results and analysis, to systematically verify the effectiveness and leading edge of this invention in multi-source data governance tasks in the power generation industry in terms of model selection accuracy, knowledge fragment matching ability, and generated content quality, the following three types of experimental evaluation tasks were designed:

[0141] Accuracy Evaluation of Large-Scale Model Selection in Power Generation Industry Data Scenarios: To verify the effectiveness of the data semantic modeling and multi-model capability matching mechanism in the S2 module, a multi-scenario test set was constructed in eight typical business sub-domains of the power generation industry (operation monitoring, safety management, dispatching instructions, equipment maintenance, regulatory compliance, energy analysis, environmental management, and integrated dispatching). The Micro-F1 Score was selected as the evaluation metric for model selection and compared with the classic BERT-MLC multi-label classifier and the TF-IDF+SVM method. Experimental results are as follows: Figure 5 As shown;

[0142] according to Figure 5The results show that the method of this invention significantly outperforms the comparative methods in all test domains, with Micro-F1 scores consistently ranging from 0.80 to 0.87, and an average score of 0.84. The method performed best in the scenarios of "integrated scheduling," "safety management," and "equipment maintenance," achieving high scores of 0.87, 0.85, and 0.84, respectively. In contrast, BERT-MLC only achieved scores of 0.78, 0.76, and 0.74 in the same domains, while TF-IDF+SVM had a generally low distribution, with scores below 0.70 in several domains. These results fully demonstrate that the present invention introduces a multi-perspective semantic modeling mechanism oriented towards business data in the S2 module, and combines it with joint inference using large model functional preference labels, significantly improving the accuracy and robustness of model scheduling in power generation scenarios.

[0143] Evaluation of the precise matching capability of structured knowledge fragments: Based on the precise matching mechanism of knowledge fragments in the S3 module, this invention further verifies its advantages in structured knowledge localization and semantic retrieval accuracy; with Mean Reciprocal Rank (MRR) as the core indicator, two comparison algorithms, S-BERT+FAISS and BM25, are set up to cover 8 power generation business sub-scenarios consistent with the S2 module experiment.

[0144] according to Figure 6 Experimental results show that the MRR score of this invention remains above 0.78 in all domains, exceeding 0.84 in several domains, with scores reaching 0.86 and 0.85 in "Integrated Scheduling" and "Energy Analysis," respectively. In contrast, S-BERT+FAISS generally falls within the range of 0.71–0.77, while BM25 exhibits greater fluctuations, with scores below 0.70 in several domains. This performance gap stems from the triple vector mechanism in the S3 module, which integrates semantic power generation data, model structure labels, and indexed knowledge fragments. This mechanism constructs a more accurate semantic matching representation, thereby more accurately locating matching content in a large-scale knowledge base and enhancing the knowledge support and interpretability of the system response.

[0145] User subjective rating analysis of large model generation results: In order to evaluate the actual effect of the knowledge enhancement generation mechanism (S4 module) proposed in this invention in the power generation data governance task, multiple rounds of simulation task experiments were organized. Real business data was used to trigger multi-model calls to generate equipment diagnosis suggestions, scheduling compliance instructions, energy consumption anomaly analysis and other content. Users were invited to rate the results from four dimensions (out of ten): accuracy, content completeness, semantic style consistency and usability of generated content.

[0146] Table 1 User Subjective Rating Table for Content Generation Quality of Large Models

[0147]

[0148] As shown in Table 1, this invention achieved scores above 9.0 across all scoring dimensions, with an average score of 9.25, significantly higher than ChatGLM2+Prompt generation (average score 8.45) and Baichuan-13B retrieval-filled generation (average score 8.03). It particularly excels in semantic style consistency (9.4) and content completeness (9.1), demonstrating the powerful adaptability of this method in multi-round context control and cross-module knowledge fusion. These results further corroborate the technical innovation and engineering practicality of this invention from a user experience perspective. Its knowledge enhancement mechanism and model scheduling strategy have significant advantages in practical application scenarios, effectively improving the service quality and task completion rate of large models in professional contexts.

[0149] The data intelligent governance method for the power generation industry based on a large model provided in this disclosure first constructs a multi-source heterogeneous task dataset covering both structured and unstructured information, and performs fine-tuning training on multiple industry sub-domains based on a general pre-trained language model to form a large language model set with specific scenario adaptability. Then, a multi-perspective semantic representation structure is constructed for the actual input data, comprehensively modeling task intent, application scenarios, and model adaptability to predict the optimal target model and Top-K candidate models, and generate a unified semantic embedding vector. Based on this, keyword graph modeling and model scheduling labels are combined to achieve accurate matching of structured knowledge fragments through graph neural networks. Finally, the semantics of the original data and the matching knowledge are fused to guide the content generation process, and a feedback mechanism is introduced to dynamically optimize the model selection and knowledge retrieval strategies, achieving continuous system evolution and improved task adaptability.

[0150] Compared with the prior art, the present invention has the following innovative features:

[0151] (1) Multidimensional semantic modeling-driven large model adaptive selection mechanism: A semantic modeling method that integrates task intent, scene affiliation and model capability features is constructed. A unified high-dimensional semantic expression is constructed to simultaneously predict the optimal target model and Top-K candidate models, which significantly improves the accuracy and robustness of model selection and meets the needs of precise model scheduling for different types of power data and task intents. It can perform data analysis and governance for heating, new energy, gas turbine and other directions, and then accurately generate data assets.

[0152] (2) Knowledge Fragment Matching Module that Integrates Graph Structure and Semantic Tags: For a large number of structured procedures, operation and maintenance knowledge and scheduling logic in the power generation industry, a knowledge matching method that combines keyword graph structure modeling and model tag semantic fusion is proposed; the graph neural network is used to realize the accurate positioning and calling of structured knowledge fragments, thereby improving the accuracy and consistency of knowledge retrieval in tasks with complex professional terms and strong semantic context dependencies.

[0153] (3) Generation optimization mechanism based on knowledge enhancement and user feedback: A knowledge enhancement content generation method based on prompt templates is constructed, which inputs data semantics and structured knowledge fragments into the target model to guide the target model to generate task response content; at the same time, user feedback data is introduced to continuously optimize the model selection and knowledge matching process, so as to adapt to the requirements of accuracy and knowledge interpretability of response content in the actual application of the power generation industry.

[0154] and Figure 1 Corresponding to the data-intelligent governance method for the power generation industry based on large models, this disclosure also provides a data-intelligent governance device for the power generation industry based on large models, such as... Figure 7 As shown, the device may specifically include:

[0155] The generation module 701 is used to collect structured, semi-structured and unstructured enterprise corpora, combine historical data diagnostic records and knowledge fragments to construct a multi-source heterogeneous corpus dataset, and fine-tune and train it on data from different industry sub-domains based on a general pre-trained language model to form multiple vertical large language models for the power generation industry with business adaptability.

[0156] Data processing module 702 is used to construct a multi-perspective semantic representation structure for power generation business data provided by users, and output the optimal target large model, Top-K candidate model set and its semantic embedding vector through a supervised learning model for subsequent knowledge matching and data diagnostic content generation.

[0157] The output module 703 is used to construct a graph structure semantic representation based on key features extracted from power generation data, and to fuse the output embedding vector, the identifier corresponding to the target large model and the Top-K candidate model information. Through the graph neural network, it achieves accurate matching of structured knowledge fragments and outputs knowledge diagnostic content for reference.

[0158] The training module 704 is used to input user questions and knowledge text together based on the target model and the Top-K candidate model, and generate target content through prompts; at the same time, it collects user feedback information, optimizes the model selector and knowledge matching module respectively, builds user behavior profiles, and realizes dynamic system updates and personalized optimization.

[0159] In some embodiments, the aforementioned collected structured, semi-structured, and unstructured enterprise corpora, combined with historical data diagnostic records and knowledge fragments, construct a multi-source heterogeneous corpus dataset. Based on a general pre-trained language model, fine-tuning training is performed on data from different industry sub-domains to form multiple vertical large-scale language models for the power generation industry with business adaptability, including:

[0160] Construct a multi-source corpus dataset covering multiple core business modules of the power generation industry, wherein the multi-source corpus dataset includes equipment operation management, dispatch instruction execution, safety and compliance supervision, technical support services, and energy regulation and operation and maintenance management;

[0161] Based on the initial text data Text n Define or filter business requirement text data Q n The corresponding answer text A was labeled by professionals. n ;

[0162] Based on the constructed business requirement text data Q n Based on this, experts were organized to label the corresponding optimal fit large language model and Top-K candidate model labels S. n This forms structured input-output pairs;

[0163] The basic model ChatGLM2-6B was chosen as the starting point for pre-training for subsequent fine-tuning in the vertical domain.

[0164] Based on the constructed multi-source heterogeneous corpus dataset and the selected base model ChatGLM2-6B, the models for each electronic domain were fine-tuned and optimized to form multiple vertical large language models with domain adaptability.

[0165] In some embodiments, the above-mentioned multi-view semantic representation structure is constructed based on the power generation business data provided by the user. A supervised learning model outputs the optimal target large model, a Top-K candidate model set, and their semantic embedding vectors for subsequent knowledge matching and data diagnostic content generation, including:

[0166] To address three sub-objectives—data type identification, business scenario attribution judgment, and model capability matching—a multi-perspective semantic representation structure adapted to the needs of complex tasks is constructed.

[0167] A multi-perspective semantic fusion mechanism is designed to deeply fuse the feature vectors obtained from the paths corresponding to the three types of sub-objectives, in order to construct a high-order problem semantic expression with multiple semantic dimensions of task understanding, scene attribution and capability adaptation.

[0168] In the fusion semantic embedding F Emb Based on this, a two-branch model prediction structure is designed, which includes a prediction branch for the Top-K candidate model set and a prediction branch for the best target model.

[0169] Based on the constructed optimally adapted large language model labels and the corresponding dataset, a joint multi-task supervised training strategy is adopted to optimize the two sub-tasks of Top-K prediction and target model classification.

[0170] In some embodiments, the key features extracted from power generation data are used to construct a graph structure semantic representation, and the output embedding vector, the identifier corresponding to the target large model, and Top-K candidate model information are fused together. A graph neural network is then used to achieve accurate matching of structured knowledge fragments, outputting knowledge diagnostic content for reference, including:

[0171] Employing a composite rule based on TF-IDF and part-of-speech tagging, from business data text Q n Extract the weighted keywords that meet the preset conditions and construct a keyword set: Kw1, Kw2, ..., Kw M Among them, Kw m Indicates user question Q n The m-th keyword in the text, where m∈[1,M]; the keyword is converted into corresponding word vectors Kf1,Kf2,...,Kf using a pre-trained Word2Vec model. M ;

[0172] The design uses a graph-based data construction method, treating each keyword as a node and its feature vector as a node feature. Cosine similarity is used to calculate the pairwise similarity between node features. When the similarity exceeds a preset threshold, the edge is considered. th At this point, an edge is constructed between the two nodes, and the similarity calculation value is used as the feature of the edge. The final constructed graph data is Graph(V,E), where V represents the set of all nodes, and the node feature corresponding to the i-th node is n. i E represents the set of all edges, and the edge between the i-th node and the j-th node has the characteristic e. ij ;

[0173] The constructed graph data Graph(V,E) is input into a four-layer stacked graph Transformer encoder, each layer including a multi-head graph attention mechanism, a feedforward network, and a global average pooling operation;

[0174] The target model label output by the trained large model adaptive selector. and the predicted Top-K candidate set The concatenation is then performed, followed by input into a two-layer fully connected ReLU network for high-order feature extraction, resulting in the scheduling label semantic features F. label ;

[0175] The main channel output feature F graph With auxiliary channel feature F label F Emb Channel splicing is performed to form a fusion feature F Con ; Fuse features F ConThe higher-order representations are extracted from two sets of MLP-ReLU layers, and a Dropout layer is added to prevent overfitting. Finally, a sigmoid activation function is used for multi-label classification, and the confidence prediction vector of the knowledge fragment index label is output. It also calls the corresponding structured knowledge matching text based on the prediction results of the knowledge fragment index tags.

[0176] Binary Cross-Entropy is used as the multi-label supervised loss function to optimize the model parameters. The training process is combined with an early stopping mechanism and the Adam optimizer to ensure that the model achieves the best balance between generalization performance and matching accuracy. Finally, the model can locate the most relevant knowledge fragments in the structured knowledge base for the input question, thereby improving the relevance and contextual consistency of the downstream analysis results.

[0177] In some embodiments, based on the target large model, the raw power generation data and matched knowledge text are jointly input, and data analysis and governance suggestions are generated by constructing prompt templates; simultaneously, user feedback information is collected to optimize the model selector and knowledge matching module respectively, and user profiles are constructed to achieve dynamic updates and personalized optimization of the system, including:

[0178] The optimal target large model is selected. To generate the model, text matching data from the knowledge base is introduced. As a knowledge enhancement prompt, it compensates for the knowledge omissions and fuzzy expressions in the model;

[0179] Based on users' subjective ratings of the model scheduling accuracy, Ev n A model selector tuning path based on user feedback ratings is constructed. After receiving the system output, users evaluate the accuracy of model selection based on a rating mechanism that includes scheduling rationality, semantic understanding, and business style suitability, and obtain an evaluation score Ev. n ;

[0180] Based on user feedback text content, a semantically driven graph construction strategy optimization mechanism is established. If the target tag appears in the user feedback text content, the system identifies relevant entities and relationships based on the target tag, thus preventing deviations in the construction of structured knowledge graphs.

[0181] For each user, we maintain historical data types, model call preferences, and knowledge matching habits and behavioral characteristics, and realize the semantic mapping between user behavior patterns and model strategies by constructing a semantic similarity matching matrix.

[0182] In some embodiments, the above-described construction of a multi-source corpus dataset covering multiple core business modules of the power generation industry includes equipment operation management, dispatch instruction execution, safety and compliance supervision, technical support services, and energy regulation and operation and maintenance management, including:

[0183] Initial business text data was collected from multiple information sources, including power plant management systems, monitoring and maintenance platforms, safety manuals, technical guidelines, expert response systems, and duty work order databases, forming raw corpora from multiple vertical domains: Text1, Text2, ..., Text N , where Text n Let n represent the initial text data of the nth domain, where n∈[1,N];

[0184] Text n Divided into multiple sub-segments and This represents a sub-segment of text containing T characters; each segment The first 1 to t words are used as input The last t+1 to T words are output. Then, corresponding operations are performed on all sub-segments of data to ultimately form a domain autoregressive dataset for all vertical domains;

[0185] Based on the initial text data Text n Define or filter business requirement text data Q n This was used as input to a domain-specific question-and-answer dataset; and the corresponding answer text A was labeled by professionals. n This is used as the output of a domain question-answering dataset to create domain question-answering datasets for all vertical domains.

[0186] In some embodiments, the above-mentioned three sub-objectives—data type identification, business scenario attribution judgment, and model capability matching—are used to construct a multi-perspective semantic representation structure that adapts to the needs of complex tasks, including:

[0187] Using a pre-trained bert-base-chinese model to process business data text data Q n Encode the sequence of context-dependent word vectors F. base ;

[0188] The data vector F is processed through a five-layer cross-attention mechanism. glo With scene category vector Layer-by-layer alignment modeling is used to capture semantic relevance, and the output of each layer is processed by the ReLU activation function to enhance the semantic nonlinearity expression capability.

[0189] Using the chinese-roberta-wwm-ext model to process business data text data Q n Perform sentence-level modeling to obtain the initial semantic embedding vector F glo Construct a set of business scenario category vectors Let n represent the category vector in the nth scene domain, where n∈[1,N];

[0190] The data vector F is processed through a five-layer cross-attention mechanism. glo With scene category vector Layer-by-layer alignment modeling is used to capture semantic relevance. The output of each layer is processed by the ReLU activation function to enhance the semantic nonlinearity representation capability; the final output scene feature vector F Sap Characterize the semantic attribution features of data within the business structure;

[0191] The pre-trained model Qwen1.5-0.5B, which has instruction understanding capabilities, was used to process the data text Q. n Embedding is performed to extract semantic representation vector F with operation-oriented features. task Preserve the original semantic information of the data at the structural and logical intent level;

[0192] The task semantic representation vector F task The data is input into a five-layer stacked Transformer encoder structure, which uses a multi-head attention mechanism to model different combinations of data elements and their preferences for different models. Each Transformer output is followed by a ReLU activation function to ultimately obtain the fitting feature vector F representing the data-model matching. Mat This is used in the subsequent semantic fusion process.

[0193] Combination Figure 8 As shown in the illustration, this disclosure also provides a data-intelligent governance device 800 for the power generation industry based on a large model, including a processor 804 and a memory 801. Optionally, the system may further include a communication interface 802 and a bus 803. The processor 804, communication interface 802, and memory 801 can communicate with each other via the bus 803. The communication interface 802 can be used for information transmission. The processor 804 can call logical instructions in the memory 801 to execute the data-intelligent governance method for the power generation industry based on a large model described in the above embodiments.

[0194] Furthermore, the logic instructions in the aforementioned memory 801 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium.

[0195] The memory 801, as a computer-readable storage medium, can be used to store software programs and computer-executable programs, such as program instructions / modules corresponding to the methods in the embodiments of this disclosure. The processor 804 executes functional applications and data processing by running the program instructions / modules stored in the memory 801, thereby realizing the data-intelligent governance method for the power generation industry based on a large model as described in the above embodiments.

[0196] The memory 801 may include a program storage area and a data storage area. The program storage area may store the operating system and application programs required for at least one function; the data storage area may store data created based on the use of the terminal device. Furthermore, the memory 801 may include high-speed random access memory and may also include non-volatile memory.

[0197] This disclosure provides a computer-readable storage medium storing computer-executable instructions configured to execute a data-intelligent governance method for the power generation industry based on a large model.

[0198] The aforementioned computer-readable storage medium may be a transient computer-readable storage medium or a non-transitory computer-readable storage medium.

[0199] The technical solutions of this disclosure can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes one or more instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of this disclosure. The aforementioned storage medium can be a non-transitory storage medium, including: a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk, and other media capable of storing program code; it can also be a transient storage medium.

[0200] The foregoing description and accompanying drawings fully illustrate embodiments of the present disclosure to enable those skilled in the art to practice them. Other embodiments may include structural, logical, electrical, procedural, and other changes. The embodiments represent only possible variations. Individual components and functions are optional unless explicitly required, and the order of operation may vary. Parts and features of some embodiments may be included in or replace parts and features of other embodiments. As used in the description of the embodiments, the singular forms “a,” “an,” and “(the)” are intended to equally include the plural forms unless the context clearly indicates otherwise. Similarly, the term “and / or” as used in this application means including one or more of the associated listed items and all possible combinations thereof. Additionally, when used in this application, the terms “comprise” and its variations “comprises” and / or “comprising” refer to the presence of stated features, integrals, steps, operations, elements, and / or components, but do not preclude the presence or addition of one or more other features, integrals, steps, operations, elements, components, and / or groups thereof. Unless otherwise specified, an element defined by the phrase "comprising a..." does not exclude the presence of other identical elements in the process, method, or apparatus that includes the element. In this document, each embodiment may focus on describing the differences from other embodiments, and similar or identical parts between embodiments can be referred to mutually. For methods, products, etc., disclosed in the embodiments, if they correspond to the method section disclosed in the embodiments, then the relevant parts can be referred to the description of the method section.

[0201] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A data-driven intelligent governance method for the power generation industry based on a large model, characterized in that, The method includes: We collected structured, semi-structured, and unstructured enterprise corpora, combined them with historical data diagnostic records and knowledge fragments, constructed a multi-source heterogeneous corpus dataset, and fine-tuned the training on data from different industry sub-domains based on a general pre-trained language model, forming multiple vertical large language models for the power generation industry with business adaptability. For the power generation business data provided by users, a multi-perspective semantic representation structure is constructed. The optimal target large model, the Top-K candidate model set and their semantic embedding vectors are output through a supervised learning model for subsequent knowledge matching and data diagnostic content generation. The graph structure semantic representation is constructed using key features extracted from power generation data. The output embedding vector, the identifier corresponding to the target large model, and the Top-K candidate model information are integrated. The graph neural network is used to achieve accurate matching of structured knowledge fragments and output knowledge diagnostic content for reference. Based on the target large model, the raw power generation data and the matched knowledge text are jointly input, and data analysis and governance suggestions are generated by constructing prompt templates. At the same time, user feedback information is collected to optimize the model selector and knowledge matching module respectively, and user profiles are constructed to realize the dynamic updating and personalized optimization of the system.

2. The method according to claim 1, characterized in that, The collected structured, semi-structured, and unstructured enterprise corpora, combined with historical data diagnostic records and knowledge fragments, construct a multi-source heterogeneous corpus dataset. Based on a general pre-trained language model, fine-tuning and training are performed on data from different industry sub-domains to form multiple vertical large-scale language models for the power generation industry with business adaptability, including: Construct a multi-source corpus dataset covering multiple core business modules of the power generation industry, wherein the multi-source corpus dataset includes equipment operation management, dispatch instruction execution, safety and compliance supervision, technical support services, and energy regulation and operation and maintenance management; Based on the initial text data Text n Define or filter business requirement text data Q n The corresponding answer text A was labeled by professionals. n ; Based on the constructed business requirement text data Q n Based on this, experts were organized to label the corresponding optimal fit large language model and Top-K candidate model labels S. n This forms structured input-output pairs; The basic model ChatGLM2-6B was chosen as the starting point for pre-training for subsequent fine-tuning in the vertical domain. Based on the constructed multi-source heterogeneous corpus dataset and the selected base model ChatGLM2-6B, the models for each electronic domain were fine-tuned and optimized to form multiple vertical large language models with domain adaptability.

3. The method according to claim 2, characterized in that, The aforementioned method constructs a multi-perspective semantic representation structure based on the power generation business data provided by the user. Through a supervised learning model, it outputs the optimal target large model, a Top-K candidate model set, and their semantic embedding vectors for subsequent knowledge matching and data diagnostic content generation, including: To address three sub-objectives—data type identification, business scenario attribution judgment, and model capability matching—a multi-perspective semantic representation structure adapted to the needs of complex tasks is constructed. A multi-perspective semantic fusion mechanism is designed to deeply fuse the feature vectors obtained from the paths corresponding to the three types of sub-objectives, in order to construct a high-order problem semantic expression with multiple semantic dimensions of task understanding, scene attribution and capability adaptation. In the fusion semantic embedding F Emb Based on this, a two-branch model prediction structure is designed, which includes a prediction branch for the Top-K candidate model set and a prediction branch for the best target model. Based on the constructed optimally adapted large language model labels and the corresponding dataset, a joint multi-task supervised training strategy is adopted to optimize the two sub-tasks of Top-K prediction and target model classification.

4. The method according to claim 1, characterized in that, The process involves constructing a graph-structured semantic representation using key features extracted from power generation data, fusing the output embedding vector, the identifier corresponding to the target large model, and Top-K candidate model information, and then using a graph neural network to achieve accurate matching of structured knowledge fragments. The output includes knowledge diagnostic content for reference, including: Employing a composite rule based on TF-IDF and part-of-speech tagging, from business data text Q n Extract the weighted keywords that meet the preset conditions and construct a keyword set: Kw1, Kw2, ..., Kw M Among them, Kw m Indicates user question Q n The m-th keyword in the text, where m∈[1,M]; the keyword is converted into corresponding word vectors Kf1,Kf2,...,Kf using a pre-trained Word2Vec model. M ; The design uses a graph-based data construction method, treating each keyword as a node and its feature vector as a node feature. Cosine similarity is used to calculate the pairwise similarity between node features. When the similarity exceeds a preset threshold, the edge is considered. th At this point, an edge is constructed between the two nodes, and the similarity calculation value is used as the feature of the edge. The final constructed graph data is Graph(V,E), where V represents the set of all nodes, and the node feature corresponding to the i-th node is n. i E represents the set of all edges, and the edge between the i-th node and the j-th node has the characteristic e. ij ; The constructed graph data Graph(V,E) is input into a four-layer stacked graph Transformer encoder, each layer including a multi-head graph attention mechanism, a feedforward network, and a global average pooling operation; The target model label output by the trained large model adaptive selector. and the predicted Top-K candidate set The concatenation is then performed, followed by input into a two-layer fully connected ReLU network for high-order feature extraction, resulting in the scheduling label semantic features F. label ; The main channel output feature F graph With auxiliary channel feature F label F Emb Channel splicing is performed to form a fusion feature F Con ; Fuse features F Con The higher-order representations are extracted from two sets of MLP-ReLU layers, and a Dropout layer is added to prevent overfitting. Finally, a sigmoid activation function is used for multi-label classification, and the confidence prediction vector of the knowledge fragment index label is output. It also calls the corresponding structured knowledge matching text based on the prediction results of the knowledge fragment index tags. Binary Cross-Entropy is used as the multi-label supervised loss function to optimize the model parameters. The training process is combined with an early stopping mechanism and the Adam optimizer to ensure that the model achieves the best balance between generalization performance and matching accuracy. Finally, the model can locate the most relevant knowledge fragments in the structured knowledge base for the input question, thereby improving the relevance and contextual consistency of the downstream analysis results.

5. The method according to claim 4, characterized in that, Based on the target large model, the raw power generation data and the matched knowledge text are jointly input, and data analysis and governance suggestions are generated by constructing prompt templates. Simultaneously, user feedback is collected to optimize the model selector and knowledge matching module, and user profiles are built to enable dynamic system updates and personalized tuning, including: The optimal target large model is selected. To generate the model, text matching data from the knowledge base is introduced. As a knowledge enhancement prompt, it compensates for the knowledge omissions and fuzzy expressions in the model; Based on users' subjective ratings of the model scheduling accuracy, Ev n A model selector tuning path based on user feedback ratings is constructed. After receiving the system output, users evaluate the accuracy of model selection based on a rating mechanism that includes scheduling rationality, semantic understanding, and business style suitability, and obtain an evaluation score Ev. n ; Based on user feedback text content, a semantically driven graph construction strategy optimization mechanism is established. If the target tag appears in the user feedback text content, the system identifies relevant entities and relationships based on the target tag, thus preventing deviations in the construction of structured knowledge graphs. For each user, we maintain historical data types, model call preferences, and knowledge matching habits and behavioral characteristics, and realize the semantic mapping between user behavior patterns and model strategies by constructing a semantic similarity matching matrix.

6. The method according to claim 2, characterized in that, The construction of the multi-source corpus dataset covers multiple core business modules of the power generation industry. This multi-source corpus dataset includes data on equipment operation management, dispatch instruction execution, safety and compliance supervision, technical support services, and energy regulation and maintenance management, including: Initial business text data was collected from multiple information sources, including power plant management systems, monitoring and maintenance platforms, safety manuals, technical guidelines, expert response systems, and duty work order databases, forming raw corpora from multiple vertical domains: Text1, Text2, ..., Text N , where Text n Let n represent the initial text data of the nth domain, where n∈[1,N]; Text n Divided into multiple sub-segments and This represents a sub-segment of text containing T characters; each segment The first 1 to t words are used as input The last t+1 to T words are output. Then, corresponding operations are performed on all sub-segments of data to ultimately form a domain autoregressive dataset for all vertical domains; Based on the initial text data Text n Define or filter business requirement text data Q n This was used as input to a domain-specific question-and-answer dataset; and the corresponding answer text A was labeled by professionals. n This is used as the output of a domain question-answering dataset to create domain question-answering datasets for all vertical domains.

7. The method according to claim 3, characterized in that, The above describes the construction of a multi-perspective semantic representation structure adapted to the needs of complex tasks, addressing three sub-objectives: data type identification, business scenario attribution judgment, and model capability matching. Using a pre-trained bert-base-chinese model to process business data text data Q n Encode the sequence of context-dependent word vectors F. base ; The data vector F is processed through a five-layer cross-attention mechanism. gko With scene category vector Layer-by-layer alignment modeling is used to capture semantic relevance, and the output of each layer is processed by the ReLU activation function to enhance the semantic nonlinearity expression capability. Using the chinese-roberta-wwm-ext model to process business data text data Q n Perform sentence-level modeling to obtain the initial semantic embedding vector F glo Construct a set of business scenario category vectors Let n represent the category vector in the nth scene domain, where n∈[1,N]; The data vector F is processed through a five-layer cross-attention mechanism. glo With scene category vector Layer-by-layer alignment modeling is used to capture semantic relevance. The output of each layer is processed by the ReLU activation function to enhance the semantic nonlinearity representation capability; the final output scene feature vector F Sap Characterize the semantic attribution features of data within the business structure; The pre-trained model Qwen1.5-0.5B, which has instruction understanding capabilities, was used to process the data text Q. n Embedding is performed to extract semantic representation vector F with operation-oriented features. task Preserve the original semantic information of the data at the structural and logical intent level; The task semantic representation vector F task The data is input into a five-layer stacked Transformer encoder structure, which uses a multi-head attention mechanism to model different combinations of data elements and their preferences for different models. Each Transformer output is followed by a ReLU activation function to ultimately obtain the fitting feature vector F representing the data-model matching. mat This is used in the subsequent semantic fusion process.

8. A data-driven intelligent governance device for the power generation industry based on a large model, characterized in that: The device includes: The generation module is used to collect structured, semi-structured and unstructured enterprise corpora, combine historical data diagnostic records and knowledge fragments to construct a multi-source heterogeneous corpus dataset, and fine-tune and train it on data from different industry sub-domains based on a general pre-trained language model to form multiple vertical large language models for the power generation industry with business adaptability. The data processing module is used to construct a multi-perspective semantic representation structure for the power generation business data provided by users. It outputs the optimal target large model, the Top-K candidate model set and their semantic embedding vectors through a supervised learning model, which are used for subsequent knowledge matching and data diagnostic content generation. The output module is used to construct a graph structure semantic representation based on key features extracted from power generation data, and to integrate the output embedding vector, the identifier corresponding to the target large model, and the Top-K candidate model information. Through graph neural networks, it achieves accurate matching of structured knowledge fragments and outputs knowledge diagnostic content for reference. The training module is used to input user questions and knowledge text together based on the target model and Top-K candidate model, and generate target content through prompts; at the same time, it collects user feedback information, optimizes the model selector and knowledge matching module respectively, builds user behavior profiles, and realizes dynamic system updates and personalized optimization.

9. An electronic device, characterized in that, include: At least one processor; and a memory communicatively connected to the at least one processor; The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method of any one of claims 1-7.

10. A non-transitory computer-readable storage medium storing computer instructions, characterized in that, The computer instructions are used to cause the computer to perform the method according to any one of claims 1-7.

Images