Energy-saving and carbon-reducing dynamic decision-making method, system and device based on artificial intelligence large model, and storage medium

Abstract

The invention discloses an energy-saving and carbon-reducing dynamic decision-making method and system based on an artificial intelligence large model, and relates to the technical field of energy-saving control and artificial intelligence. Comprising the steps of collecting and storing structure data and vector data, and constructing a data warehouse; cleaning the data, removing taint data, executing intelligent recovery and rewriting, and pushing and updating the cleaned data to a knowledge base; when a user inputs or is triggered abnormally, performing problem analysis by using the first LLM, performing full retrieval on the knowledge base by using the second LLM in combination with an RAG technology, and performing comprehensive analysis on a retrieval result by using the third LLM; and finally, a natural language report, an operation suggestion or an executable regulation and control instruction is generated, and regulation and control are executed through IOT or an artificial channel according to the authority rule. Through intelligent rewriting and recovery of abnormal data and three-level LLM cooperative processing, data quality and decision accuracy are significantly improved, and dynamic closed-loop optimization and knowledge iteration in an energy-saving and carbon-reducing scene are realized.

Description

Technical Field

[0001] This invention relates to the fields of energy-saving control and artificial intelligence, and in particular to a dynamic decision-making method, system, device and storage medium for energy saving and carbon reduction based on a large artificial intelligence model. Background Technology

[0002] With the rapid development of IoT technology, traditional IoT systems can now achieve interconnection and data collection and transmission of physical devices through various sensors, communication modules, and network protocols. Existing technologies typically employ centralized or distributed data platforms to aggregate the collected data and perform simple data processing, analysis, and early warning notifications according to preset rules. For example, in fields such as industrial monitoring, environmental monitoring, and smart buildings, the system can monitor equipment status, environmental parameters, and other information in real time, and notify relevant personnel via SMS, email, or platform alarms when data anomalies occur.

[0003] However, such systems still have significant limitations in practical applications. First, the system's decision-making loop heavily relies on human intervention. Even after an alert is issued, operators still need to manually assess and decide on subsequent operational logic, such as adjusting equipment parameters or activating emergency plans. This approach not only fails to effectively save labor costs but also limits the system's response speed to human reaction time, making it prone to delayed responses in emergency failure scenarios, thus affecting the system's real-time performance and reliability.

[0004] Furthermore, the fault handling and decision-making process heavily relies on the personal experience of operators. Due to the lack of systematic summarization and effective accumulation of historical handling cases and expert experience, it is difficult to form a standardized decision support knowledge base. When faced with complex or rare anomalies, personnel often have to rely solely on their own experience to make judgments, resulting in low processing efficiency and a high risk of decision-making bias due to insufficient experience, thus affecting the overall efficiency and accuracy of fault recovery.

[0005] Therefore, while achieving data interconnection, the current IoT technology framework urgently needs to break through the bottleneck of relying on manual decision-making. By introducing more intelligent data analysis, decision support, and automated execution mechanisms, the system's autonomous response capability, processing efficiency, and reliability can be improved, thereby truly achieving cost reduction, efficiency improvement, and intelligent energy-saving management. Summary of the Invention

[0006] The present invention aims to overcome at least one of the defects of the prior art and provide a dynamic decision-making method for energy conservation and carbon reduction based on a large artificial intelligence model, which is used to solve the technical problem that energy conservation and emission reduction based on human decision-making in the Internet of Things is not timely.

[0007] This invention provides a method for IoT data sharing and data quality inspection based on a large model for government affairs, including: S1: Data is collected through the intelligent control system and transmitted to the business system via IoT for structured data and vector data storage; S2: Combine structured data and vector data to build a data warehouse; S3: After cleaning the data and removing tainted data, push the document to update the knowledge base; S4: Based on the knowledge base, when a user input or a business system data collection configuration triggers an anomaly, the input question or anomaly is pushed to the first LLM, analyzed, and then pushed to the second LLM. The second LLM uses RAG search enhancement capabilities to perform a full search of the knowledge base, obtains the knowledge search results, and pushes them to the third LLM to analyze the conclusions and push them to the data control system. S5: After receiving the conclusions from the data control system, perform comprehensive analysis and expression generation to generate user-oriented natural language analysis reports, operation suggestions, or executable control instructions; S6: The business system pre-configures data operation permissions, and after actively pushing the analysis results to the data control system, it processes them through two branches: the IoT data transmission channel or the manual channel, according to the permission rules. S7: Based on the control command, after completing the instruction control of the environmental infrastructure equipment through the intelligent control system, the operation data is collected and transmitted back to the business system via IoT; S8: Clean the returned operation data, push updates to the knowledge base, and perform iterative optimization and data accumulation.

[0008] By constructing a special library of abnormal data for standardized fault scenarios and designing specific system prompts based on incomplete data types, combined with model re-ranking and multi-dimensional quality verification, high-quality recovery and standardized reconstruction of previously ignored abnormal, unstructured, and noisy data in business systems are achieved. This significantly improves the semantic integrity, format standardization, and factual consistency of tainted data, greatly expands the effective coverage of the knowledge base, and reduces the information loss rate in the data collection process. Simultaneously, through a three-level LLM collaborative architecture, based on the first LLM's problem intent parsing, the second LLM's RAG vector retrieval enhancement, and the third LLM's conclusion deduction technology, intelligent decision-making is achieved throughout the entire process from anomaly triggering to knowledge matching and strategy generation. This significantly improves the dynamic response speed, retrieval recall accuracy, and executability of control commands in complex energy-saving scenarios, forming a dynamic self-optimizing system for energy conservation and carbon reduction: "data cleaning—knowledge accumulation—intelligent decision-making—closed-loop control—iterative optimization," achieving truly unmanned energy conservation and emission reduction.

[0009] In step S1, the basic data acquisition capabilities of the business system are utilized to collect data through the intelligent control system and transmit it to the business system via IoT for structured data and vector data storage.

[0010] The structured data is traditional database data.

[0011] The vector data refers to data that has undergone vectorization and is suitable for AI processing, such as text and image features, embeddings, etc.

[0012] The business system performs real-time or batch processing on structured and vector data and operates according to preset business logic and rules.

[0013] In step S2, the data warehouse is responsible for storing raw business data, cleaned data, and anomaly records, supporting business traceability and preliminary analysis, and providing support for traditional data analysis and reporting.

[0014] The data cleaning described in step S3 includes: the system intelligently reclaims and rewrites abnormal data that was originally ignored, as well as unstructured or noisy data generated during the operation of the business system. The core objective is to distinguish between completely meaningless noise and incomplete data containing potential business logic from abnormal or low-quality fragments.

[0015] The intelligent data recycling and rewriting includes: S31: Establish a temporary storage library to record the source and original context of each abnormal or noisy data; S32: Perform data segmentation, deduplication, and key information masking on the data in the special database; S33: For the deduplicated data, design specific system prompts for different types of missing data and rewrite them into semantically correct data; S34: Use a reordering strategy for the rewritten data, that is, use the model to generate multiple candidates, and then use a rule or classification model to select the best one. S35: Perform format validation, factual consistency check, and quality scoring on the reordered data.

[0016] After data cleaning, during the process of building a knowledge base by pushing data from the data warehouse, LLM-based data processing enhancement capabilities are used to further clean, segment, and integrate the data, forming a vectorized knowledge base that covers all cleaned data collected by the business system.

[0017] The knowledge base stores refined, structured knowledge that can be directly accessed by AI. Abnormal business data first enters the data warehouse, where it is cleaned and integrated, and then selectively pushed to the knowledge base. The knowledge base receives incremental information through updates pushed from the data warehouse, indirectly connecting with business systems to serve LLM enhancement processing and RAG search.

[0018] In step S4, based on the knowledge base, when a user inputs or the business system data collection configuration triggers an anomaly, the input question or anomaly is pushed to the first LLM, analyzed, and then pushed to the second LLM to use RAG search enhancement capabilities to perform a full search of the knowledge base, obtain knowledge retrieval results, and then pushed to the third LLM to analyze the conclusions and push them to the data control system.

[0019] The first LLM (Local Management Model) is responsible for problem analysis. As the system entry point, it instantly understands user intent and business anomalies, transforming unstructured natural language input or abnormal signals into structured task or problem descriptions that the system can understand. The second LLM is responsible for data processing enhancement, located in the knowledge base stage. It deeply processes information, creating high-quality knowledge. Utilizing cleaned data pushed from the data warehouse, it summarizes, generalizes, correlates, and formats the data to generate structured or vectorized knowledge that is easy to retrieve and understand. The third LLM is responsible for conclusion analysis. After receiving the conclusions from the data control system, it performs final comprehensive judgment and expression generation. Combining information from all stages (original problem, retrieved knowledge, and control data results), it generates user-oriented natural language analysis reports, operational suggestions, or executable control instructions.

[0020] In step S6, the business system pre-configures data operation permissions and processes data through either the IoT data transmission channel or the manual channel according to the permission rules.

[0021] The IoT data transmission channel processing includes the business system actively sending control commands through the IoT data transmission channel to autonomously and dynamically control the edge infrastructure equipment, thereby completing the business closed loop.

[0022] The manual channel processing includes the business system providing analysis results to the user through the data control system, which serves as a summary of historical experience and suggestions for subsequent manual operations, and the business loop is completed by manual processing.

[0023] This invention also provides a dynamic decision-making system for energy conservation and carbon reduction based on a large artificial intelligence model, characterized in that it includes: The data acquisition module is used for: acquiring data through the intelligent control system; The data storage module is used for: storing structured data and vector data transmitted to the business system via IoT devices; The data cleaning module is used for: data cleaning, intelligent data recycling and rewriting, building data warehouses, and pushing knowledge bases. The data processing module is used to: obtain user input questions or system anomaly information, push them to the LLM for basic analysis, and then the LLM uses RAG search enhancement capabilities to perform a full search of the knowledge base, retrieve related data from the knowledge base and provide it to the LLM for basic data analysis, and obtain analysis results. The access control module is used to process analysis results through either an IoT data transmission channel or a manual channel, depending on the data access permissions. The results analysis module is used to: after receiving the conclusions from the data control system, perform final comprehensive analysis and expression generation, and generate user-oriented natural language analysis reports, operation suggestions, or executable control instructions.

[0024] The present invention also provides an apparatus comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, when the processor executes the program, it implements the energy-saving and carbon-reduction dynamic decision-making method based on an artificial intelligence large model as described in any one of claims 1 to 7.

[0025] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, when the computer program is executed by a processor, it implements the energy-saving and carbon-reduction dynamic decision-making method based on an artificial intelligence large model as described in any one of claims 1 to 7.

[0026] This invention provides a dynamic decision-making method for energy conservation and carbon reduction based on a large-scale artificial intelligence model. It uses Large-Scale Language Modeling (LLM) and Retrieval Enhancement Group (RAG) as the core foundation for question answering and decision-making, constructing an intelligent IoT dynamic decision-making and application system. Within a single system, multiple LLMs are employed. The first LLM analyzes the problem, quickly and accurately identifying the core issue. The second LLM, combined with RAG technology, performs a full search of the knowledge base, comprehensively acquiring relevant information and avoiding omissions. The third LLM comprehensively analyzes the search results, further integrating and refining key content. These three LLMs work collaboratively, along with intelligent rewriting and recycling of abnormal data, effectively improving data quality, ensuring the reliability of decision-making basis, and significantly improving decision accuracy. Ultimately, this achieves dynamic closed-loop optimization and knowledge iteration in energy conservation and carbon reduction scenarios, providing strong support for the efficient operation and continuous improvement of the system.

[0027] Users only need to describe the type, format, and application scenario of the required data through the front-end interactive interface, and the data sharing model can understand their intentions and automatically generate standardized and complete data request schemes, thereby significantly reducing the threshold for data-using departments to obtain and use data and improving the efficiency and user-friendliness of data services. Attached Figure Description

[0028] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0029] Figure 1 This is a schematic diagram illustrating the steps of the energy-saving and carbon-reduction dynamic decision-making method based on a large artificial intelligence model.

[0030] Figure 2 These are the prompts and example diagrams for the data recycling and utilization rewriting module in Example 3. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings: This invention provides an embodiment of a dynamic decision-making method and system for energy conservation and carbon reduction based on a large artificial intelligence model. It should be noted that although the logical order is shown in the flowchart, under certain data conditions, the steps shown or described may be completed in a different order than that shown here.

[0032] Example 1: Refer to Figures 1-2 like Figure 1 As shown, this invention provides a dynamic decision-making method for energy conservation and carbon reduction based on a large-scale artificial intelligence model, including: S1: Data is collected through the intelligent control system and transmitted to the business system via IoT for structured data and vector data storage; S2: Combine structured data and vector data to build a data warehouse; S3: After cleaning the data and removing tainted data, push the document to update the knowledge base; S4: Based on the knowledge base, when a user input or a business system data collection configuration triggers an anomaly, the input question or anomaly is pushed to the first LLM, analyzed, and then pushed to the second LLM. The second LLM uses RAG search enhancement capabilities to perform a full search of the knowledge base, obtains the knowledge search results, and pushes them to the third LLM to analyze the conclusions and push them to the data control system. S5: After receiving the conclusions from the data control system, perform comprehensive analysis and expression generation to generate user-oriented natural language analysis reports, operation suggestions, or executable control instructions; S6: The business system pre-configures data operation permissions, and after actively pushing the analysis results to the data control system, it processes them through two branches: the IoT data transmission channel or the manual channel, according to the permission rules. S7: Based on the control command, after completing the instruction control of the environmental infrastructure equipment through the intelligent control system, the operation data is collected and transmitted back to the business system via IoT; S8: Clean the returned operation data, push updates to the knowledge base, and perform iterative optimization and data accumulation.

[0033] By constructing a special library of abnormal data for standardized fault scenarios and designing specific system prompts based on incomplete data types, combined with model re-ranking and multi-dimensional quality verification, high-quality recovery and standardized reconstruction of previously ignored abnormal, unstructured, and noisy data in business systems are achieved. This significantly improves the semantic integrity, format standardization, and factual consistency of tainted data, greatly expands the effective coverage of the knowledge base, and reduces the information loss rate in the data collection process. Simultaneously, through a three-level LLM collaborative architecture, based on the first LLM's problem intent parsing, the second LLM's RAG vector retrieval enhancement, and the third LLM's conclusion deduction technology, intelligent decision-making is achieved throughout the entire process from anomaly triggering to knowledge matching and strategy generation. This significantly improves the dynamic response speed, retrieval recall accuracy, and executability of control commands in complex energy-saving scenarios, forming a dynamic self-optimizing system for energy conservation and carbon reduction: "data cleaning—knowledge accumulation—intelligent decision-making—closed-loop control—iterative optimization," achieving truly unmanned energy conservation and emission reduction.

[0034] In step S1, the basic data acquisition capabilities of the business system are utilized to collect data through the intelligent control system and transmit it to the business system via IoT for structured data and vector data storage.

[0035] The structured data is traditional database data.

[0036] The vector data refers to data that has undergone vectorization and is suitable for AI processing, such as text and image features, embeddings, etc.

[0037] The business system performs real-time or batch processing on structured and vector data and operates according to preset business logic and rules.

[0038] In step S2, the data warehouse is responsible for storing raw business data, cleaned data, and anomaly records, supporting business traceability and preliminary analysis, and providing support for traditional data analysis and reporting.

[0039] The data cleaning described in step S3 includes: the system intelligently reclaims and rewrites abnormal data that was originally ignored, as well as unstructured or noisy data generated during the operation of the business system. The core objective is to distinguish between completely meaningless noise and incomplete data containing potential business logic from abnormal or low-quality fragments.

[0040] The intelligent data recycling and rewriting includes: S31: Establish a temporary storage library to record the source and original context of each abnormal or noisy data; S32: Perform data segmentation, deduplication, and key information masking on the data in the special database; S33: For the deduplicated data, design specific system prompts for different types of missing data and rewrite them into semantically correct data; S34: Use a reordering strategy for the rewritten data, that is, use the model to generate multiple candidates, and then use a rule or classification model to select the best one. S35: Perform format validation, factual consistency check, and quality scoring on the reordered data.

[0041] After data cleaning, during the process of building a knowledge base by pushing data from the data warehouse, LLM-based data processing enhancement capabilities are used to further clean, segment, and integrate the data, forming a vectorized knowledge base that covers all cleaned data collected by the business system.

[0042] The knowledge base stores refined, structured knowledge that can be directly accessed by AI. Business data containing anomalies first enters the data warehouse, and after cleaning and integration, it is selectively pushed to the knowledge base. This avoids the knowledge base being polluted by massive amounts of raw data, ensuring its high-quality and high-density content.

[0043] The knowledge base receives incremental information through push updates from the data warehouse, indirectly connecting with business systems. This reduces the real-time load on the knowledge base while ensuring the systematic nature and traceability of knowledge updates.

[0044] The data warehouse independently expands storage and computing capabilities, while the knowledge base focuses on improving retrieval efficiency and knowledge organization. Through push and retrieval interactions, the data warehouse and knowledge base enhance the system's flexibility and maintainability.

[0045] In step S4, when a user input or a data collection configuration of the business system triggers an anomaly, the collected issues are pushed to the first LLM for basic analysis. The second LLM then uses RAG search enhancement capabilities to perform a full search of the knowledge base, retrieves relevant data from the knowledge base, and provides it to the third LLM for basic data analysis. The analysis results cover methods for dealing with the current situation, subsequent precautions, and operational suggestions.

[0046] The anomalies include: 1. Technical anomalies: Traditional "errors" such as code errors, service outages, and API timeouts.

[0047] 2. Business Rule Anomalies: When a business system processes data, if the data violates preset business rules, an "abnormal situation" will be triggered even if the program itself does not "report an error". For example: in a risk control system, the characteristic value of a transaction touches the risk rule threshold; in a supply chain system, the inventory level of a warehouse falls below the safety level; in an operations system, user activity suddenly drops sharply.

[0048] 3. Data quality anomalies: Dirty data, missing values, or logically contradictory data that cannot be automatically repaired and are discovered during the initial data cleaning process are reported as "abnormal situations".

[0049] 4. Predictive anomalies: The system issues early warning signals based on historical data or model predictions, anticipating potential problems at a future point in time.

[0050] When processing collected structural and vector data, a single LLM can only perform data cleaning according to preset general rules. It is difficult to perform detailed and accurate processing on the complex and diverse data characteristics in the field of energy conservation and carbon reduction. Therefore, the identification of tainted data may not be comprehensive enough. When intelligently recycling and rewriting abnormal data, it also lacks in-depth understanding of specific business scenarios, resulting in some valuable data being wrongly discarded or rewritten inaccurately.

[0051] This invention employs multiple LLM collaborative processing. The first LLM, during the problem analysis phase, can make a preliminary judgment on the overall characteristics and potential problems of the data, providing guidance for subsequent data cleaning. The second LLM, combined with search enhancement RAG technology, can utilize the rich historical data and experience in the knowledge base to more accurately identify tainted data and intelligently recycle and rewrite abnormal data with reference to relevant knowledge, thereby improving data quality and providing a more reliable data foundation for subsequent decision-making.

[0052] In step S5, after receiving the conclusions from the data control system, a comprehensive analysis and expression generation are performed. The third LLM combines the original question, the retrieved knowledge, and the information from the control data results to generate a user-oriented natural language analysis report, operation suggestions, or executable control instructions.

[0053] When faced with complex issues in the field of energy conservation and carbon reduction, a single LLM lacks a deep understanding and detailed knowledge of the field. They can only analyze the problems at a superficial level, making it difficult to delve into the essence and root causes of the problems, resulting in inaccurate and incomplete problem analysis.

[0054] In this invention, the third LLM combines the problem analysis results of the first LLM to evaluate and weigh the retrieved information from multiple perspectives. For example, when considering energy-saving and carbon-reducing measures, it is necessary to consider not only the effect of reducing energy consumption, but also the impact on equipment operation, cost input, and other factors. Then, through comprehensive analysis, more scientific and reasonable natural language reports, operation suggestions, or executable control instructions are generated to improve the accuracy and feasibility of decision-making.

[0055] In step S6, the business system pre-configures data operation permissions and processes data through either the IoT data transmission channel or the manual channel according to the permission rules.

[0056] The IoT data transmission channel processing includes the business system actively sending control commands through the IoT data transmission channel to autonomously and dynamically control the edge infrastructure equipment, thereby completing the business closed loop.

[0057] The manual channel processing includes the business system providing analysis results to the user through the data control system, which serves as a summary of historical experience and suggestions for subsequent manual operations, and the business loop is completed by manual processing.

[0058] Example 2: The specific screening steps for data recycling and utilization include: S321, A rule engine based on regular expressions, pattern matching, and business thresholds quickly removes obvious noise; For structured SQL fragments, check if they contain keywords such as SELECT, WHERE, JOIN, etc., but the structure is incomplete (e.g., missing FROM clause). If so, retain them. For vectorized embeddings, the vector magnitude (L2 norm) is calculated. If the magnitude is close to 0 or is NaN, it is considered invalid noise. The specific formula is as follows: is_valid={True, if∥v∥2>θ False ,otherwise} Where θ is an empirical threshold (e.g., 0.05); For log anomalies: If a log matches a known error code pattern (such as HTTP 5xx, ERROR, Exception) but more than 50% of the context information (such as timestamp, device ID) is missing, it is marked as "incomplete".

[0059] S322. Using unsupervised and supervised learning, identify fragments with potential business logic from the data retained in S321, cluster the fragments. Outliers are usually noise, while data that can form small clusters, even if incomplete, may represent a certain type of recurring business scenario. If historical labeled data is available, a binary classification model is trained. If the features include extraction length, information entropy, keyword frequency, syntactic complexity, and semantic similarity to existing entries in the knowledge base, then the "meaningful / meaningless" probability is output. The decision formula is as follows: (\text{keep} = \begin{cases} \text{True},&\text{if} P(\text{meaningful}); The criteria for determining meaningful incomplete data are: 1. Business relevance: There is an identifiable association with core business entities (users, devices); 2. Pattern identifiability: Although incomplete, it exhibits a certain inductive pattern (such as specific error types or similar SQL fragments).

[0060] 3. Information gain: Includes information that is not yet fully covered or is new in the knowledge base; 4. Repairability: With context (such as concurrent logs, database schema) or large model, there is a high probability that it can be accurately completed.

[0061] Example 3: The specific steps of the data recycling and utilization rewriting module include: S331. Design specific system prompts and examples for different types of incompleteness, such as... Figure 2 As shown.

[0062] S332. Use the LLM (DeepSeek) API. To control cost and quality, a "re-ranking" strategy is adopted, which involves generating multiple candidates using a smaller model (such as 7B), and then selecting the optimal candidate using a rule-based or smaller classification model. S333. Post-processing and validation of the data, including: Format validation: Use regular expressions to check if the output conforms to the preset template; Fact consistency check: Compare the knowledge generated by the LLM with the original context in the data warehouse, and calculate the overlap rate of key entities (such as order number, device ID), where represents the entity set. A low overlap rate (e.g., <0.5) may indicate that the LLM is "fabricated"; Quality scoring: Train a regression model (features: fluency of generated text, confidence, similarity to the knowledge base) or use an NLI (Natural Language Inference) model to determine whether the generated content has semantic implications as the original fragment.

[0063] S334. The verified data is labeled with metadata tags such as "Source: Recycling Enhancement", "Confidence: 0.92", and "Generation Timestamp".

[0064] S335. Convert the generated standardized text into vectors using the same embedding model (such as text-embedding-3-small).

[0065] S336. Store the standardized text and the transformed vector, along with the metadata, into the knowledge base for subsequent RAG retrieval.

[0066] Example 4: This invention also provides a dynamic decision-making system for energy conservation and carbon reduction based on a large artificial intelligence model, characterized in that it includes: The data acquisition module is used for: acquiring data through the intelligent control system; The data storage module is used for: storing structured data and vector data transmitted to the business system via IoT devices; The data cleaning module is used for: data cleaning, intelligent data recycling and rewriting, building data warehouses, and pushing knowledge bases. The data processing module is used to: obtain user input questions or system anomaly information, push the first LLM for question analysis, push the second LLM to use RAG search enhancement capabilities to perform a full search of the knowledge base, obtain knowledge retrieval results, and push the third LLM to analyze the conclusions. The access control module is used to process analysis results through either an IoT data transmission channel or a manual channel, depending on the data access permissions. The results analysis module is used to: after receiving the conclusions from the data control system, perform final comprehensive analysis and expression generation, and generate user-oriented natural language analysis reports, operation suggestions, or executable control instructions.

[0067] The processes in this system mainly involve hardware devices, IoT platforms, and business systems. The process logic is as follows: 1. There are two input methods: 1) "Events to be processed" or "signals to be watched" that are identified by programs or business rules during the operation of the business system and require understanding and decision-making by the first LLM; 2) User input issues; 2. Events of interest to the business system or questions entered by users are analyzed by pushing the first LLM. After analysis, the second LLM uses RAG search enhancement capabilities to perform a full search of the knowledge base to obtain knowledge retrieval results. The third LLM then analyzes the conclusions. 3. Upon receiving the conclusions from the data control system, and combining information from all stages (original problem, retrieved knowledge, and control data results), generate a user-oriented natural language analysis report, operational suggestions, or executable control instructions: 1) When permissions permit, the system can autonomously control the intelligent control system and manipulate basic environmental equipment through the data control system; 2) If permissions are not allowed, the data control system will return operation suggestions to the user and provide feedback to the human staff for historical experience summary and suggestions, so that the human staff can complete the subsequent business loop.

[0068] This invention employs multiple LLMs. The first LLM performs problem analysis, enabling rapid and accurate identification of the core issue. The second LLM, combined with Retrieval Enhancement Group (RAG) technology, performs a full search of the knowledge base, comprehensively acquiring relevant information and avoiding information omissions. The third LLM comprehensively analyzes the search results, further integrating and refining key content.

[0069] The above three LLM processes work together, along with the intelligent rewriting and recycling of abnormal data. During data processing and decision analysis, new data and experience can be continuously collected and organized into the knowledge base, thus effectively improving data quality, ensuring the reliability of decision-making basis, and significantly improving decision accuracy. Ultimately, this achieves dynamic closed-loop optimization and knowledge iteration in energy-saving and carbon-reduction scenarios, providing strong support for the long-term efficient operation and continuous improvement of the system.

Claims

1. A dynamic decision-making method for energy conservation and carbon reduction based on a large-scale artificial intelligence model, characterized in that, The method includes the following steps: S1: Data is collected through the intelligent control system and transmitted to the business system via IoT for structured data and vector data storage; S2: Combine structured data and vector data to build a data warehouse; S3: After cleaning the data and removing tainted data, push the document to update the knowledge base; S4: Based on the knowledge base, when a user input or a business system data collection configuration triggers an anomaly, the input question or anomaly is pushed to the first LLM, analyzed, and then pushed to the second LLM. The second LLM uses RAG search enhancement capabilities to perform a full search of the knowledge base, obtains the knowledge search results, and pushes them to the third LLM to analyze the conclusions and push them to the data control system. S5: After receiving the conclusions from the data control system, perform comprehensive analysis and expression generation to generate user-oriented natural language analysis reports, operation suggestions, or executable control instructions; S6: The business system pre-configures data operation permissions, and after actively pushing the analysis results to the data control system, it processes them through two branches: the IoT data transmission channel or the manual channel, according to the permission rules. S7: Based on the control command, after completing the instruction control of the environmental infrastructure equipment through the intelligent control system, the operation data is collected and transmitted back to the business system via IoT; S8: Clean the returned operation data, push updates to the knowledge base, and perform iterative optimization and data accumulation.

2. The energy-saving and carbon-reduction dynamic decision-making method based on a large artificial intelligence model according to claim 1, characterized in that, In step S3, data cleaning includes: The system intelligently recycles and rewrites previously ignored abnormal data, as well as unstructured or noisy data generated during the operation of business systems.

3. The energy-saving and carbon-reduction dynamic decision-making method based on a large artificial intelligence model according to claim 2, characterized in that, Intelligent data recycling and rewriting includes: S31: Establish a temporary storage library to record the source and original context of each abnormal or noisy data; S32: Perform data segmentation, deduplication, and key information masking on the data in the special database; S33: For the deduplicated data, design specific system prompts for different types of missing data and rewrite them into semantically correct data; S34: Use a reordering strategy for the rewritten data, that is, use the model to generate multiple candidates, and then use a rule or classification model to select the best one. S35: Perform format validation, factual consistency check, and quality scoring on the reordered data.

4. The energy-saving and carbon-reduction dynamic decision-making method based on a large artificial intelligence model according to claim 3, characterized in that: After data cleaning, during the process of building a knowledge base by pushing data from the data warehouse, the data is further cleaned, segmented, and integrated based on the enhanced data processing capabilities of the second LLM, forming a vectorized knowledge base that covers all cleaned data collected by the business system.

5. The energy-saving and carbon-reduction dynamic decision-making method based on a large artificial intelligence model according to claim 1, characterized in that, Build a data warehouse and a knowledge base separately: The data warehouse is responsible for storing raw business data, cleaned data, and anomaly records, providing support for traditional data analysis and reporting. The knowledge base stores cleaned, structured knowledge that can be directly accessed by AI, serving LLM enhancement processing and RAG search.

6. The energy-saving and carbon-reduction dynamic decision-making method based on a large artificial intelligence model according to claim 1, characterized in that, In step S6, the permission rules include: Based on the analysis results, within the scope of data permission operations, the business system actively sends control commands through the IoT data transmission channel to autonomously and dynamically control the edge infrastructure equipment, thus completing the business closed loop.

7. The energy-saving and carbon-reduction dynamic decision-making method based on a large artificial intelligence model according to claim 6, characterized in that: Based on the analysis results, if data access permissions do not permit operation, the business system provides the analysis results to the user through the data control system. This serves as a summary of historical experience and suggestions for subsequent manual operations, allowing for manual processing to complete the business loop.

8. A dynamic decision-making system for energy conservation and carbon reduction based on a large-scale artificial intelligence model, characterized in that, include: The data acquisition module is used for: acquiring data through the intelligent control system; The data storage module is used for: storing structured data and vector data transmitted to the business system via IoT devices; The data cleaning module is used for: data cleaning, intelligent data recycling and rewriting, building data warehouses, and pushing knowledge bases. The data processing module is used to: obtain user input questions or system anomaly information, push the first LLM for question analysis, push the second LLM to use RAG search enhancement capabilities to perform a full search of the knowledge base, obtain knowledge retrieval results, and push the third LLM to analyze the conclusions. The access control module is used to process analysis results through either an IoT data transmission channel or a manual channel, depending on the data access permissions. The results analysis module is used to: after receiving the conclusions from the data control system, perform final comprehensive analysis and expression generation, and generate user-oriented natural language analysis reports, operation suggestions, or executable control instructions.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the energy-saving and carbon-reduction dynamic decision-making method based on a large artificial intelligence model as described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the energy-saving and carbon-reduction dynamic decision-making method based on a large artificial intelligence model as described in any one of claims 1 to 7.

Images