Energy management multi-agent collaborative optimization method and system based on large model agent architecture

Abstract

The application discloses an energy management multi-agent collaborative optimization method and system based on a large model agent architecture, which comprises a multi-agent architecture module, a model processing module, a data acquisition and management module, a system integration and interaction module, and a collaborative optimization control module; the model processing module contains a large model and multiple small models, realizing task processing and analysis; the data acquisition and management module collects and processes multi-source data; the system integration and interaction module realizes multi-platform integration and visual interaction; and the collaborative optimization control module guarantees the collaborative work of agents and system optimization. The application has the advantages of autonomy and decentralized control, heterogeneity and modularity, intelligent collaboration, etc., can accurately diagnose instrument faults, optimize energy management, intelligently control production, improve enterprise energy utilization efficiency, instrument operation stability and overall operation management level, help enterprises realize green and low-carbon transformation, and is suitable for energy management and energy equipment maintenance scenes of various industrial production enterprises.

Description

TECHNICAL FIELD

[0001] The present application mainly relates to the field of energy management and instrument fault diagnosis in the field of industrial production, focusing on optimizing energy utilization efficiency and instrument operation state monitoring through intelligent means, and specifically applied to energy consumption optimization, instrument maintenance, and real-time control and adjustment of energy equipment. Specifically relates to an energy management multi-agent collaborative optimization method and system based on a large model agent architecture. BACKGROUND

[0002] Energy, as a key supporting element of enterprise production and manufacturing activities, accounts for a significant proportion of the total cost of the enterprise, and this proportion is particularly prominent for high-energy-consuming enterprises.

[0003] Current enterprises in the industrial manufacturing industry are facing many difficulties. In terms of energy equipment management, the energy equipment management system in the industrial production site generally relies too much on manual experience and rule-based diagnosis methods. This mode has many disadvantages. First, maintenance experience is difficult to effectively accumulate, reuse, and system integrate, and knowledge loss occurs frequently, and there is a serious problem of repeated construction. Second, due to the lack of effective prediction means, the energy instrument fault point cannot be known in advance, resulting in the inability of energy equipment to be maintained in a timely manner, and unexpected shutdown events occur from time to time, which not only seriously disrupts the normal production order and reduces production efficiency, but also causes a lot of unnecessary waste of energy.

[0004] In the field of energy management, the existing energy management system also has obvious shortcomings. First, in terms of real-time and computational efficiency, the model often has computational delay when dealing with complex industrial scenarios, which cannot meet the strict requirements of industrial production for millisecond-level real-time response, which seriously restricts the overall performance of the system in the industrial environment with extremely high real-time requirements. Second, there is a lot of room for improvement in the accuracy of energy prediction, and there is an optimization problem between real-time and prediction period, making it difficult to ensure high accuracy while ensuring prediction timeliness. Third, the sensitivity and specificity of energy consumption anomaly detection need to be further enhanced, making it difficult for the system to quickly and accurately identify energy consumption anomalies and take effective measures in a timely manner. In addition, the degree of cross-domain knowledge fusion is low, making it difficult for the energy management system to comprehensively consider instrument operating status, production plans, environmental factors, and other information, and it is difficult to develop comprehensive, accurate, and efficient energy management strategies.

[0005] In summary, the existing industrial manufacturing management system has significant defects in energy equipment maintenance and energy management, and there is an urgent need for an innovative and efficient solution to promote the intelligent, efficient, and sustainable development of industrial production. SUMMARY

[0006] In view of the deficiencies of the prior art, the present application provides an energy management multi-agent collaborative optimization method and system based on a large model agent architecture; for realizing intelligent energy management optimization and energy equipment control in industrial production and energy management scenarios. The system uses a large model combined with a small model architecture, adopts task decomposition and deep learning technology, and has specific functional capabilities such as instrument fault diagnosis, energy use analysis (including energy use prediction, high energy consumption identification, and energy use anomaly diagnosis), and production energy use implementation control. The system includes multiple agents, each of which is responsible for a different workshop or energy node, and provides accurate energy management recommendations and energy equipment operation optimization control through data acquisition, analysis, and feedback.

[0007] To achieve the above-mentioned purpose, the present application provides a first aspect, which provides an energy management multi-agent collaborative optimization method based on a large model agent architecture, characterized by comprising the following steps:

[0008] S1 task input and preliminary processing step: receiving user input tasks, performing Prompt query on the input tasks, converting them into a form suitable for large model processing, planning and decomposing the tasks by the large model, determining the subtasks to be executed, determining the data types to be collected according to the subtasks, collecting instrument running state data, energy consumption data, and production environment data from a data source system including sensors, intelligent instruments, and enterprise internal production databases, cleaning the collected data to remove outliers and noise, and performing preprocessing including normalization and feature extraction;

[0009] S2 data collection and model analysis step: calling corresponding models for different subtasks, calling a fault diagnosis small model group including a DBSCAN model and an AutoEncoder model for instrument fault diagnosis subtasks, using the DBSCAN model to locate the fault position through instrument running data clustering analysis, and using the AutoEncoder model to determine the fault by learning the normal operation mode of the instrument to determine the data deviation; calling an energy use analysis small model group including an ARIMA model for energy consumption analysis subtasks, selecting appropriate methods for energy use prediction, high energy consumption identification, and energy use anomaly detection according to the characteristics of the energy consumption data, using a nonlinear trend prediction method if the data has a trend, using a seasonal autoregressive model if it has a seasonality, and using a moving average method if it has no trend and seasonality, and then returning the model analysis results to the large model;

[0010] S3 result integration and control execution step: the large model integrates the small model results to perform root cause analysis, excavates the root cause from the overall system perspective, generates optimization strategy suggestions including instrument fault handling suggestions and energy distribution strategies based on root cause analysis, generates controller instructions according to the optimization strategy suggestions, realizes dynamic control of energy equipment and energy management system, and the system can dynamically adjust the control strategy according to the equipment aging, production plan change and energy supply fluctuation.

[0011] Further, in step S1, the user input task is in natural language form, specifically the energy management and instrument state related requirements of a certain workshop or production link, the large model plans and disassembles the task based on the knowledge and ability obtained through incremental pre-training and instruction fine-tuning, and fully considers the logical relationship between subtasks during disassembly. The sensors in the data source system include temperature sensors, pressure sensors, intelligent instruments for measuring energy consumption data, enterprise internal production databases store instrument running state data and production plan data, statistical analysis methods are used for data cleaning to remove abnormal data, specific algorithms are used for normalization processing to make the data comparable, and principal component analysis methods are used for feature extraction.

[0012] Further, in step S2, the DBSCAN model determines the clustering parameters according to the data density distribution during clustering analysis, the AutoEncoder model learns the normal operation mode by constructing a multi-layer neural network, and the ARIMA model determines the model parameters for prediction according to the energy consumption data autocorrelation and seasonal characteristics. During energy consumption analysis, historical energy consumption and production plan data are combined for comprehensive judgment.

[0013] Further, in step S3, during the root cause analysis process, the large model analyzes the cause and effect relationship between faults and energy consumption anomalies in combination with equipment operation principles, production plan requirements and historical data experience, and generates optimization strategy suggestions by comprehensively considering enterprise production targets, cost limitations and energy supply conditions. The controller instructions are transmitted to the PLC through a specific communication protocol, the PLC controls the start-stop, speed and power parameters of the energy equipment according to the instructions, and the system evaluates and adjusts the effect through real-time monitoring data and feedback mechanism.

[0014] Another aspect of the present application provides an energy management multi-agent collaborative optimization system based on a large model agent architecture, which is used to realize the above-mentioned method, comprising:

[0015] Multi-agent architecture module: composed of multiple agents, each agent is responsible for a specific energy equipment or energy node, based on a large model agent architecture, has independent perception, decision and execution capabilities, and cooperates with other agents through network communication to realize information sharing, including sharing monitoring data, diagnosis results and optimization strategies;

[0016] The model processing module includes a large model and a small model. The large model is based on Qwen2.5 and is pre-trained with incremental knowledge in energy management and energy equipment maintenance. It uses instruction fine-tuning technology to understand and execute task instructions, and uses prompt engineering to realize chain reasoning to decompose complex tasks. It also has a tool calling function to connect external tools such as PLC controllers and energy consumption analysis tools. The small model includes a DBSCAN model for low-dimensional fault diagnosis, an AutoEncoder model for high-dimensional data fault detection, and a moving average method model, a seasonal autoregressive model, and a nonlinear trend prediction method model for time series data analysis in energy management, which assists the large model in fine-grained analysis.

[0017] The data acquisition and management module collects instrument operation state data, energy consumption data, and production environment data through sensors and intelligent meters. The collected data is cleaned, pre-processed, and stored. Advanced algorithms are used to remove outliers and noise. Pre-processing includes data normalization and feature extraction. The data storage management system uses efficient structures and algorithms to classify and manage historical data. Big data technology and cloud computing platforms are used to process and analyze data.

[0018] The system integration and interaction module seamlessly integrates with one or more production data platforms such as MES, APS, MRO, SCADA, and IoT. It connects with other systems through standardized interfaces, develops a user interface with visualization functions, and displays instrument operation state, energy consumption data, analysis results, and optimization strategy information. It allows operators to set parameters, issue tasks, and monitor the system, while also having user permission management functions to ensure data security and operation standards.

[0019] The collaborative optimization control module is used for information sharing, collaboration, and dynamic task allocation among agents. It establishes negotiation and coordination mechanisms to solve resource conflicts and task competition problems. The dynamic task allocation mechanism allocates tasks based on task priority, urgency, agent load, and processing capacity. It can adjust collaborative strategies and optimization objectives based on device aging, production plan changes, and energy supply fluctuations. It uses a performance evaluation index system to monitor and evaluate key indicators such as energy utilization efficiency, instrument operation stability, task processing timeliness, and adjusts collaborative strategies and optimization parameters automatically based on the results.

[0020] Further, the large model constructs a decision tree according to logical steps when performing chain reasoning, adjusts reasoning paths and decision strategies based on real-time data and feedback, and connects external tools through secure interface technology and interacts with data. Incremental pre-training uses continuous learning algorithms to update the knowledge system to adapt to environmental changes.

[0021] Further, the DBSCAN model adjusts clustering parameters according to data distribution to improve positioning accuracy in low-dimensional fault diagnosis, the AutoEncoder model enhances abnormal pattern recognition ability through reasonable encoding and decoding structure, the moving average method model dynamically adjusts the size of the time window according to the data, the seasonal autoregressive model optimizes parameters according to seasonal cycles, and the nonlinear trend prediction method model uses fitting algorithm to adapt to trend changes.

[0022] Further, the sensors and smart meters in the data acquisition and management module have high precision, high reliability and real-time transmission capability, the data cleaning algorithm intelligently identifies and processes abnormal data, the data storage management system supports fast retrieval, backup and recovery, the big data technology and cloud computing platform use distributed computing architecture and parallel processing algorithm to improve efficiency, and can intelligently cache and optimize storage strategies.

[0023] Further, the standardized interface in the system integration and interaction module complies with industry standard specifications to ensure compatibility and interoperability, the user interface visualization uses intuitive charts and dynamic display for easy operation, and the user permission management function is perfect to ensure data security and operation specification.

[0024] Further, the negotiation and coordination mechanism in the collaborative optimization control module uses a distributed consensus algorithm to ensure decision consistency, the dynamic task allocation mechanism uses intelligent algorithms to evaluate task and agent state, the resource conflict and task competition resolution strategy is based on priority and fairness principles, and the performance evaluation index system comprehensively monitors key indicators and automatically adjusts strategies and parameters.

[0025] The above technical scheme is adopted in the present application, which has at least the following beneficial effects:

[0026] 1. Autonomy and decentralized control

[0027] (1) Existing technology: There may be certain limitations in autonomy and decentralized control. Although existing large models have cloud edge end collaboration, autonomous control, and data and model security features, ensuring their stability and security in actual application, there are limitations in autonomous decision-making and decentralized control capabilities, and in some complex energy management scenarios, more central coordination and control are needed, thereby limiting their adaptability and flexibility to some extent.

[0028] (2) The present invention: It shows significant technical advantages in autonomy and decentralized control. The model is based on an advanced multi-agent system architecture, giving each sub-model high independence and autonomous decision-making ability. Sub-models can independently perform complex reasoning, planning, and select the most suitable strategy to deal with various sub-problems. Without a central control layer, sub-models can still make effective decisions and actions, significantly improving the robustness and flexibility of the entire system, allowing it to remain stable in the face of dynamic changes in the environment.

[0029] 2. Heterogeneity and modularity

[0030] (1) Existing large model technology: Although it also has modular design, it has poor flexibility in heterogeneity, affecting the adaptability in diversified energy management tasks.

[0031] (2) The present invention: Adopting the principle of heterogeneity, each agent constituting the system can have specialized structure and function according to the nature of the task it undertakes. This design not only gives the system high adaptability and flexibility, but also enables each agent to provide customized solutions to specific problems. In addition, the modular design of the model ensures the independence of each component, facilitating independent development, testing and reuse, greatly improving the efficiency of the development process, and enhancing the maintainability and scalability of the system.

[0032] 3. Intelligent collaboration

[0033] (1) Existing large model technology: Although it has made achievements in modular design, it has great limitations in dynamic asynchronous computing and self-organization collaboration, limiting its efficiency in handling energy management tasks that require fast response and flexible resource scheduling.

[0034] (2) The present invention: It adopts a dynamic asynchronous computing mode in the computing architecture, which significantly improves the parallel processing capability and response speed of the system. The model can dynamically adjust resource allocation and execution order according to the urgency of the task and computing demand, realizing non-synchronous high-efficiency computing. In addition, the energy consumption large model also has excellent self-organizing ability, forming an effective cooperation mechanism through local interaction, enabling large and small models to work together through negotiation, coordination and cooperation to solve complex problems, exhibiting strong collaborative processing capability. BRIEF DESCRIPTION OF DRAWINGS

[0035] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings needed to be used in the embodiment or prior art description. Obviously, the drawings in the following description are only some embodiments of the present invention, and those skilled in the art can obtain other drawings according to these drawings without creative labor.

[0036] Figure 1 is the flow chart of the energy management multi-agent collaborative optimization method of the present application;

[0037] Figure 2 is the flow chart of the time series prediction model group decision-making of the present application. DETAILED DESCRIPTION

[0038] The exemplary embodiments will be described in detail herein with reference to the accompanying drawings. In the following description, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The implementations described in the following exemplary embodiments do not represent all implementations consistent with the present application. Instead, they only represent examples of devices and methods consistent with some aspects of the present application, as detailed in the appended claims.

[0039] Embodiment One

[0040] Referring to Figure 1 and Figure 2 , the first aspect of the present embodiment provides an energy management multi-agent collaborative optimization method based on a large model agent architecture, comprising the following steps:

[0041] S1 task input and preliminary processing step: receiving user input tasks, querying the input tasks with prompts, converting them into a form suitable for large model processing, the large model planning and disassembling the tasks, determining the subtasks to be executed, determining the data types to be collected according to the subtasks, collecting instrument running state data, energy consumption data and production environment data from the data source system containing sensors, intelligent instruments and enterprise internal production databases, cleaning the collected data to remove outliers and noise, and performing preprocessing including normalization and feature extraction;

[0042] S2 data collection and model analysis step: calling corresponding models for different subtasks, calling a small model group containing a DBSCAN model and an AutoEncoder model for instrument fault diagnosis subtasks, the DBSCAN model locating the fault position through instrument running data clustering analysis, and the AutoEncoder model determining the fault by learning the normal operation mode of the instrument to determine the data deviation; calling a small model group containing an ARIMA model for energy consumption analysis subtasks, selecting appropriate methods for energy consumption prediction, high energy consumption identification and energy consumption anomaly detection according to the characteristics of the energy consumption data, if the data has a trend, using a nonlinear trend prediction method, if it has a seasonality, using a seasonal autoregressive model, if it has no trend and seasonality, using a moving average method, and then returning the model analysis results to the large model;

[0043] S3 result integration and control execution step: the large model integrates the small model results to perform root cause analysis, excavates the root cause from the overall system perspective, generates optimization strategy suggestions including instrument fault handling suggestions and energy distribution strategies based on root cause analysis, generates controller instructions according to the optimization strategy suggestions, realizes dynamic control of energy equipment and energy management system, and the system can dynamically adjust the control strategy according to the equipment aging, production plan change and energy supply fluctuation.

[0044] As a preferred embodiment, in step S1 of the embodiment, the user input task is in natural language form, specifically the energy management and instrument state related requirements of a workshop or production link, the large model plans and disassembles the task based on the knowledge and ability obtained through incremental pre-training and instruction fine-tuning, fully considers the logical relationship between subtasks during disassembly, the sensors in the data source system include temperature sensors and pressure sensors, intelligent instruments are used to measure energy consumption data, enterprise internal production databases store instrument running state and production plan data, statistical analysis methods are used for data cleaning to remove abnormal data, specific algorithms are used for normalization processing to make data comparable, and principal component analysis methods are used for feature extraction.

[0045] As a preferred embodiment, in step S2 of the embodiment, the DBSCAN model determines clustering parameters according to data density distribution during clustering analysis, the AutoEncoder model learns normal operation mode by constructing a multi-layer neural network, the ARIMA model determines model parameters for prediction according to the autocorrelation and seasonal characteristics of energy consumption data, and historical energy consumption and production plan data are combined for comprehensive judgment during energy consumption analysis.

[0046] As a preferred embodiment, in step S3 of the embodiment, during the root cause analysis process, the large model analyzes the cause and effect relationship between faults and energy consumption anomalies in combination with equipment operation principles, production plan requirements and historical data experience, considers enterprise production targets, cost limitations and energy supply conditions when generating optimization strategy suggestions, the controller instructions are transmitted to the PLC through a specific communication protocol, the PLC controls the start and stop, speed and power parameters of the energy equipment according to the instructions, and the system evaluates and adjusts the effect through real-time monitoring data and feedback mechanism.

[0047] The second aspect of the embodiment provides an energy management multi-agent collaborative optimization system based on a large model agent architecture, which is used to implement the above method, and includes:

[0048] Multi-agent architecture module: composed of multiple agents, each agent is responsible for a specific energy device or energy node, constructed based on a large model agent architecture, has independent perception, decision and execution capabilities, and cooperates with other agents through network communication to realize information sharing, including sharing monitoring data, diagnosis results and optimization strategies;

[0049] Model processing module: contains large model and small model, large model is based on Qwen2.5, incrementally pre-trained into energy management and energy equipment maintenance professional knowledge, understands and executes task instructions by using instruction fine-tuning technology, realizes thinking chain reasoning decomposition of complex tasks with the help of prompt engineering, and has tool calling function to connect external tools such as PLC controller and energy consumption analysis tool, small model includes DBSCAN model for low-dimensional fault diagnosis, AutoEncoder model for high-dimensional data fault detection, and moving average method model, seasonal autoregressive model and nonlinear trend prediction method model for time series data analysis in energy management, assisting large model for fine-grained analysis;

[0050] Data acquisition and management module: collects instrument running state data, energy consumption data and production environment data through sensors and intelligent instruments, cleans, preprocesses and stores the collected data, data cleaning uses advanced algorithms to remove outliers and noise, preprocessing includes data normalization and feature extraction, data storage management system uses efficient structure and algorithm to classify and store and manage historical data, processes and analyzes data with the help of big data technology and cloud computing platform;

[0051] System integration and interaction module: seamless integration with one or more production data platforms of MES, APS, MRO, SCADA and IoT, through standardized interface and other systems, develop user interface with visualization function, display instrument running state, energy consumption data, analysis results and optimization strategy information, convenient for operators to set parameters, issue tasks and monitor system, at the same time, has user permission management function to ensure data security and operation specification.

[0052] Collaborative optimization control module: for information sharing, collaborative cooperation and dynamic task allocation among agents, establishes negotiation and coordination mechanism to solve resource conflict and task competition problem, dynamic task allocation mechanism allocates tasks according to task priority, urgency, agent load and processing capacity, etc., and can adjust collaborative strategy and optimization target according to equipment aging, production plan change and energy supply fluctuation, etc., adopts performance evaluation index system to monitor and evaluate key indicators such as energy utilization efficiency, instrument running stability, task processing timeliness, etc. in real time, adjusts collaborative strategy and optimization parameters automatically according to the results.

[0053] As a preferred embodiment, the large model in this embodiment constructs decision tree according to logical steps when thinking chain reasoning, adjusts reasoning path and decision strategy according to real-time data and feedback, tool calling function connects external tools through security interface technology and interacts data, and incrementally pre-training uses continuous learning algorithm to update knowledge system to adapt to environmental changes.

[0054] As a preferred embodiment, the DBSCAN model in this embodiment adjusts clustering parameters according to data distribution to improve positioning accuracy in low-dimensional fault diagnosis, the AutoEncoder model enhances abnormal pattern recognition ability through reasonable encoding and decoding structure, the moving average method model dynamically adjusts the size of the time window according to the data, the seasonal autoregressive model optimizes parameters according to seasonal cycles, and the nonlinear trend prediction method model uses fitting algorithm to adapt to trend changes.

[0055] As a preferred embodiment, the sensors and smart meters in the data acquisition and management module in this embodiment have high precision, high reliability and real-time transmission capability, the data cleaning algorithm intelligently identifies and processes abnormal data, the data storage management system supports fast retrieval, backup and recovery, the big data technology and cloud computing platform use distributed computing architecture and parallel processing algorithm to improve efficiency, and can intelligently cache and optimize storage strategies.

[0056] As a preferred embodiment, the standardized interface in the system integration and interaction module in this embodiment complies with industry standard specifications to ensure compatibility and interoperability, the user interface visualization uses intuitive charts and dynamic display for easy operation, and the user permission management function is perfect to ensure data security and operation specification.

[0057] As a preferred embodiment, the negotiation and coordination mechanism in the collaborative optimization control module in this embodiment uses a distributed consensus algorithm to ensure decision consistency, the dynamic task allocation mechanism uses intelligent algorithms to evaluate task and agent state, the resource conflict and task competition resolution strategy is based on priority and fairness principles, and the performance evaluation index system comprehensively monitors key indicators and automatically adjusts strategies and parameters.

[0058] Embodiment Two

[0059] As Figure 1 The system flowchart analysis includes:

[0060] 1. Overall flow architecture and logical relationship;

[0061] This flowchart clearly shows the overall operation process of the energy management multi-agent system based on the large model agent architecture. Starting from natural language input tasks, through a series of ordered steps, dynamic adjustment is finally realized. The steps are closely connected, the logic is clear, and they form a complete closed-loop system.

[0062] It embodies the design idea of the system driven by tasks, which realizes energy management and energy equipment control optimization through the collaborative work of multiple modules. Each module has its own function and cooperates with each other to complete complex energy management tasks.

[0063] 2. Function of each flow node and data flow direction;

[0064] Natural language input task (prompt query): As the starting point of the system, users input tasks in natural language, which reflects the user-friendly interaction of the system and makes it convenient for operators to issue various energy management and energy equipment related instructions, such as instrument fault diagnosis requests, energy consumption optimization needs, etc. The input tasks will enter the subsequent large model processing stage.

[0065] Large Model Processing (LLM Large Model): The large model is the core decision-making unit of the system. After receiving user input tasks, it first deeply analyzes the user's intent through deep learning and contextual understanding techniques. This step is the foundation for subsequent accurate operations. Then, the large model breaks down complex tasks into specific sub-tasks. For example, the instrument fault diagnosis task is decomposed into sub-tasks such as data acquisition, fault location, and root cause analysis, preparing for the subsequent targeted invocation of different models and data sources.

[0066] Data Acquisition and Processing: Based on the sub-task requirements decomposed from the large model, relevant data is collected from data source systems. These data sources are broad, including sensors, smart meters, and internal enterprise production databases, ensuring comprehensive and accurate acquisition of instrument operating status data, energy consumption data, and production environment data. The collected data undergoes cleaning and preprocessing. Cleaning removes outliers and noise interference, while preprocessing includes normalization to ensure data comparability and feature extraction for subsequent model analysis. The processed data is then provided to the corresponding smaller models for analysis.

[0067] Small model analysis: For different types of tasks, the system calls different small models for specialized analysis. In fault diagnosis, models such as DBSCAN (density-based clustering algorithm) and AutoEncoder are used. The DBSCAN model can accurately locate the fault location of the instrument through cluster analysis of instrument operation data, while the AutoEncoder model judges the fault situation by learning the normal operation mode of the instrument and comparing the deviation of real-time data. In energy consumption analysis, models such as ARIMA (time series forecasting model) are used for energy consumption prediction, high energy consumption identification, and energy consumption anomaly detection. After the small model analysis is completed, the results are returned to the large model.

[0068] Large-scale model integration and decision-making (root cause analysis, optimization strategy recommendations): After receiving the analysis results from the smaller models, the large-scale model performs a comprehensive root cause analysis, delving into the root causes of problems from a holistic system perspective, going beyond mere surface-level fault phenomena. Based on the root cause analysis results, the large-scale model further generates comprehensive optimization strategy recommendations, covering multiple aspects such as instrument fault handling suggestions, energy equipment operating parameter adjustment strategies, production process optimization schemes, and energy allocation strategies, providing detailed guidance for the system's optimized control.

[0069] Control execution (generate controller instructions, dynamic adjustment): Finally, according to the optimization strategy suggestion to generate controller instructions, through the connection with PLC (programmable logic controller), the instructions are transmitted to the energy equipment and energy management system, realizing the real-time and dynamic adjustment of the energy equipment operation and energy use. Moreover, the system can continuously adjust the control strategy according to the actual situation such as equipment aging, production plan change and energy supply fluctuation, etc., to ensure that the system always maintains the optimal operating state.

[0070] As Figure 2 The time series prediction model group decision flowchart analysis includes:

[0071] 1. Time series data processing logic;

[0072] The flowchart focuses on the processing of time series data, and according to the characteristics of time series data such as energy consumption data in energy management, a set of scientific and reasonable prediction model decision-making process is designed. The core logic is to classify and process according to the trend and periodicity of data, to select the most suitable prediction method, and to improve the accuracy and effectiveness of prediction.

[0073] It embodies the system's fine management and targeted processing ability for different types of time series data, ensuring accurate energy consumption prediction and anomaly detection under various complex data patterns, and providing reliable decision-making basis for energy management.

[0074] 2. Different situation processing flow and model selection basis;

[0075] Trend judgment and model selection: the system first judges the trend of time series data. If the data shows a clear trend, such as energy consumption gradually rising or falling over time, a nonlinear trend prediction method is selected for prediction. This is because the nonlinear trend prediction method can better fit the nonlinear trend of the data and capture the complex rules of energy consumption changes over time, thus providing more accurate prediction results.

[0076] Periodic judgment and model selection (in the absence of trend): when there is no trend in the data, the periodic judgment link is entered. If the data has obvious periodicity, such as energy consumption data fluctuating according to daily, weekly, monthly, etc. Periodic rules, the system uses a seasonal autoregressive model for prediction. The seasonal autoregressive model can fully consider the seasonal periodic characteristics of the data, use the periodic rules in the historical data to predict future data, and effectively deal with the impact of periodic changes.

[0077] Trendless and aperiodic case processing: If the data has neither trend nor periodicity, the moving average method is used for prediction. Moving average method smooths data fluctuations by calculating the average value of data within a certain time window, which has good prediction effect on stationary time series data, and is suitable for cases where energy consumption data is relatively stable and has no obvious regular changes. Through this case-by-case processing method, the system can select the most suitable prediction model according to the actual characteristics of the data, maximize the prediction accuracy, and provide strong support for energy management decisions.

[0078] Embodiment three

[0079] Please refer to Figure 1 and Figure 2 The embodiment provides an implementation of an energy management multi-agent collaborative optimization method based on a large model agent architecture, including the following steps:

[0080] I. System initialization and preparation

[0081] An energy management multi-agent system based on a large model agent architecture is constructed to ensure that the system includes scene function modules, large language model modules, production data docking and control system integration modules, industry standard docking modules, and other main components. Stable communication connections are established between the modules to achieve smooth transmission of data and instructions.

[0082] The large model (based on ChatGLM3-6B or Qwen2.5) is initialized and set, including completing incremental pre-training, integrating professional knowledge in energy management, energy equipment maintenance, and other specific fields into the model parameters, so that it has the basic ability to understand and handle industrial scene tasks. At the same time, for different task scenarios, the large model is optimized using instruction fine-tuning technology to ensure that it can accurately understand and execute complex instructions such as "fault analysis" and "energy optimization". Set the prompt engineering related parameters, so that the large model can perform chain reasoning, effectively decompose and solve complex tasks. Configure the tool calling function of the large model, establish a connection interface with external tools (such as PLC controllers, energy consumption analysis tools, etc.) through AIAgent technology, and ensure that the operation of external tools can be realized.

[0083] Deploy small models for specific tasks, such as DBSCAN models for low-dimensional fault diagnosis, AutoEncoder models for high-dimensional data fault detection, and moving average method models, seasonal autoregressive models, and nonlinear trend prediction method models for time series data analysis in energy management. Ensure that the small models and large models establish effective data interaction channels to receive large model instructions and return analysis results.

[0084] Connect to data source systems, including various sensors (such as temperature sensors, pressure sensors, etc.), smart meters (for measuring energy consumption data, etc.), and the enterprise's internal production database, to ensure the system can collect real-time information such as instrument operating status data, energy consumption data, and production environment data. Calibrate and test the data acquisition equipment to ensure data accuracy and reliability. Set up a data storage management system, utilizing efficient data structures and algorithms to classify, store, and manage the collected historical data, facilitating subsequent data retrieval, analysis, and mining.

[0085] II. Task Input and Processing

[0086] Users input tasks (prompt queries) in natural language, such as "Analyze the reasons for excessive energy consumption in a certain production workshop and propose optimization solutions." The input tasks are then fed into the LLM large model through the system interface.

[0087] After receiving a task, the LLM large model uses deep learning and contextual understanding techniques to perform in-depth analysis of user intent. Through semantic understanding, syntactic parsing, and matching with pre-trained knowledge of natural language text, it accurately grasps the core content of the user's needs. For example, it identifies key information such as the instrument range involved in the task, energy consumption-related issues, and expected optimization goals.

[0088] Based on the analysis results, the large model breaks down complex tasks into specific sub-tasks. For example, regarding the aforementioned problem of excessive energy consumption, it breaks down the task into a data acquisition sub-task (including collecting real-time energy consumption data, operating parameters, ambient temperature, and other relevant data from the production workshop), a fault diagnosis sub-task (determining whether instrument malfunctions are causing abnormal energy consumption), and an energy consumption analysis sub-task (analyzing the trends and periodic characteristics of energy consumption data to identify high-energy-consumption periods and processes). Simultaneously, the data types and data sources required for each sub-task are determined to prepare for subsequent data acquisition and model invocation.

[0089] III. Data Acquisition and Preprocessing

[0090] Based on the sub-task requirements derived from the large model, the system collects relevant data from the data source system. It acquires real-time instrument operating status data (such as temperature, pressure, and flow rate) through sensors, accurately measures energy consumption data using smart instruments, and simultaneously extracts historical operating data and production plan data related to the workshop from the enterprise's internal production database.

[0091] The collected data enters the data processing stage. First, data cleaning is performed, using advanced algorithms (such as outlier detection algorithms based on statistical analysis) to identify and remove outliers and noise interference, ensuring the accuracy and reliability of the data. For example, data points that significantly deviate from the normal energy consumption range are corrected or deleted.

[0092] The pre-processing operation is performed on the cleaned data. Normalization algorithms (such as minimum-maximum normalization or mean-standard deviation normalization) are used to convert data of different types and ranges into a unified numerical interval, making the data comparable and facilitating subsequent model analysis. At the same time, feature extraction techniques (such as principal component analysis PCA and other methods) are used to extract key features from the original data, reduce data dimensionality, improve model calculation efficiency, and preserve the main information of the data. The pre-processed data will be provided to the corresponding small models for analysis according to the sub-task requirements.

[0093] IV. Model analysis and diagnosis

[0094] Fault diagnosis small model group

[0095] For the instrument fault diagnosis sub-task, the system calls the DBSCAN model and AutoEncoder model in the fault diagnosis small model group.

[0096] The DBSCAN model receives instrument operation data and determines clustering parameters based on data density distribution to perform clustering analysis. By identifying dense and sparse regions of data points, the instrument operation state is divided into normal and possible fault states. For example, if the data points of a certain instrument form a dense cluster in the normal operation state, and deviate from the cluster when an anomaly occurs, the DBSCAN model can accurately detect this change and locate the instrument fault position, such as determining that the temperature data cluster of a motor is abnormal, which may indicate a fault in the motor cooling system.

[0097] The AutoEncoder model learns the data feature representation of the instrument in the normal operation mode by constructing a multi-layer neural network. In real-time monitoring, the collected data is input into the model, and the deviation between the output and input data is calculated. If the deviation exceeds the set threshold, it is determined that the instrument has a fault. For example, for the attribute data of a certain instrument, the AutoEncoder model learns the normal operation mode, and when the instrument has faults such as inaccurate sampling and network offline, the deviation of the attribute data will trigger a fault alarm and provide evaluation information on the fault degree.

[0098] Energy consumption analysis small model group

[0099] In the energy consumption analysis sub-task, the system calls the ARIMA model and other related algorithm models in the energy consumption analysis small model group.

[0100] First, the trend and periodicity of energy consumption data are judged. If the data shows a clear trend (such as energy consumption gradually rising or falling over time), a nonlinear trend prediction method is used for prediction. This method builds a suitable nonlinear function model (such as a polynomial function, an exponential function, etc.) to fit the trend of energy consumption data, and predicts the future energy consumption trend based on historical data, providing a basis for enterprises to plan energy supply in advance.

[0101] If the energy consumption data has no trend but has a clear periodicity (such as daily, weekly, monthly, etc.), a seasonal autoregressive model is applied. This model considers the seasonal periodicity characteristics in historical data, establishes an autoregressive equation, and uses past period data to predict future energy consumption in the same period, helping enterprises to reasonably arrange production plans and reduce energy consumption during peak periods.

[0102] If the energy consumption data has neither trend nor periodicity, a moving average method is used for prediction. By calculating the average value of energy consumption data within a certain time window, the data fluctuations are smoothed, and the relatively stable energy consumption data is predicted for a short period, providing a reference for enterprises to adjust energy use strategies in real time. At the same time, based on these analysis models, the system can also identify high energy consumption and detect energy use anomalies, timely discover unreasonable energy consumption situations, and provide support for energy saving and consumption reduction.

[0103] Five, result integration and decision-making

[0104] After the fault diagnosis and energy consumption analysis small model completes the analysis, the results are returned to the LLM large model. The large model receives analysis results from different small models and conducts comprehensive root cause analysis. Combining device operation principles, production plan requirements, and historical data experience, it deeply excavates the root causes of problems from the overall system perspective. For example, if the instrument fault diagnosis small model detects that the energy consumption of a certain instrument is too high and there are abnormal operating parameters, the large model will further analyze the causal relationship between the abnormal parameters and the high energy consumption, consider factors such as device aging, environmental factors, improper operation, etc., and determine that the root cause of high energy consumption is the wear of a key component of the instrument, which leads to efficiency decline, rather than simply external environmental impact.

[0105] Based on the root cause analysis results, the large model generates comprehensive optimization strategy suggestions. The optimization strategy content covers instrument fault handling suggestions (such as specific solutions for repairing or replacing worn parts), energy device operating parameter adjustment strategies (such as optimizing energy device operating speed, temperature setting, etc.), and energy distribution strategies (such as reasonably distributing energy supply to different workshops, prioritizing key link energy demand, etc.). These optimization strategies aim to improve energy device operating efficiency, reduce energy consumption, and improve overall production benefits, while ensuring that the system operates in a stable and reliable state.

[0106] Six, control execution and dynamic adjustment

[0107] According to the optimization strategy suggestion, the LLM large model generates corresponding controller instructions. The instructions are transmitted to the PLC (Programmable Logic Controller) through a specific communication protocol. The PLC controls the operation parameters of the energy equipment and energy management system according to the received instructions. For example, adjusting the start-stop time, running speed, power output, etc. of the energy equipment, realizing precise control of energy use. At the same time, the system can monitor the changes of relevant data in real time according to the actual situation such as equipment aging, production plan changes and energy supply fluctuations, and dynamically adjust the control strategy. For example, when the energy equipment ages and its performance declines, the system automatically reduces the running load of the energy equipment, optimizes energy distribution, ensures the safe and reliable operation of the energy equipment, and avoids energy waste; if the production plan changes, the system adjusts the energy equipment operation parameters and energy supply strategy in time according to the new production plan requirements to ensure the smooth progress of production; when the energy supply is tight, the energy demand of key production links is prioritized, and the operation of non-critical equipment is adjusted reasonably to realize efficient use of energy, and ensure that the system always maintains the optimal running state, and constantly adapts to various complex and variable industrial production environments.

[0108] During the entire system operation process, the system performance is continuously monitored, instrument operation data, energy consumption data and user feedback information are collected, and the model is evaluated and optimized regularly. According to the new data and actual operation situation, the large model is incrementally trained, its knowledge and strategy are updated, the adaptability and processing capacity of the system to different tasks and scenes are improved, and the system always maintains an efficient, intelligent and stable running state, realizing continuous optimization of energy management and energy equipment control.

[0109] Although the embodiments of the present application have been shown and described above, it can be understood that the above-mentioned embodiments are exemplary and cannot be understood as limiting the present application, and those skilled in the art can make changes, modifications, replacements and variations to the above-mentioned embodiments within the scope of the present application.

Claims

1. A multi-agent cooperative optimization method for energy management based on a large model agent architecture, characterized in that: Includes the following steps: S1 Task Input and Preliminary Processing Steps: Receive user-inputted tasks, perform a Prompt query on the input tasks, transform them into a form suitable for large-scale model processing, the large-scale model plans and decomposes the tasks, determines the sub-tasks to be executed, determines the data types to be collected based on the sub-tasks, collects instrument operating status data, energy consumption data, and production environment data from data source systems including sensors, smart meters, and enterprise internal production databases, cleans the collected data to remove outliers and noise, and performs preprocessing including normalization and feature extraction; S2 Data Acquisition and Model Analysis Steps: For different subtasks, the corresponding models are invoked. For the instrument fault diagnosis subtask, a small fault diagnosis model group including the DBSCAN model and the AutoEncoder model is invoked. The DBSCAN model locates the fault location through cluster analysis of instrument operating data, while the AutoEncoder model determines the fault by judging data deviations after learning the normal operating mode of the instrument. For the energy consumption analysis subtask, a small energy consumption analysis model group including the ARIMA model is invoked. Appropriate methods are selected based on the characteristics of the energy consumption data for energy consumption prediction, high energy consumption identification, and energy consumption anomaly detection. If the data shows a trend, a nonlinear trend prediction method is used; if it shows seasonality, a seasonal autoregressive model is used; if there is no trend or seasonality, a moving average method is used. The model analysis results are then returned to the larger model. S3 Result Integration and Control Execution Steps: The large model integrates the results of the small model to perform root cause analysis, dig out the root cause of the problem from the perspective of the overall system, generate optimization strategy suggestions based on the root cause analysis, including instrument fault handling suggestions and energy allocation strategies, generate controller instructions according to the optimization strategy suggestions, realize dynamic control of energy equipment and energy management system, and the system can dynamically adjust the control strategy according to equipment aging, production plan changes and energy supply fluctuations. In step S2, the DBSCAN model determines clustering parameters based on data density distribution during cluster analysis, the AutoEncoder model learns normal operation mode by constructing a multi-layer neural network, and the ARIMA model determines model parameters for prediction based on the autocorrelation and seasonality characteristics of energy consumption data. During energy consumption analysis, it combines historical energy consumption and production plan data for comprehensive judgment. In step S3, during the root cause analysis, the large model combines the equipment operating principle, production plan requirements, and historical data experience to analyze the causal relationship between the fault and abnormal energy consumption. When generating optimization strategy suggestions, it comprehensively considers the enterprise's production goals, cost constraints, and energy supply. The controller commands are transmitted to the PLC through a specific communication protocol. The PLC controls the start-up, shutdown, speed, and power parameters of the energy equipment according to the commands. The system evaluates the adjustment effect through real-time monitoring data and feedback mechanisms.

2. The energy management multi-agent cooperative optimization method based on a large model agent architecture according to claim 1, characterized in that: In step S1, the user input task is in natural language form, specifically the energy management and instrument status requirements for a certain workshop or production process. The large model plans and decomposes the task based on the knowledge and capabilities it has acquired through incremental pre-training and instruction fine-tuning. During decomposition, the logical relationships between sub-tasks are fully considered. The sensors in the data source system include temperature sensors and pressure sensors. Smart instruments are used to measure energy consumption data. The enterprise's internal production database stores instrument operating status data and production plan data. Data cleaning uses statistical analysis methods to remove abnormal data. Normalization processing uses a specific algorithm to make the data comparable. Feature extraction uses principal component analysis.

3. A multi-agent cooperative optimization system for energy management based on a large-model agent architecture, characterized in that: The system is used to implement the method of claim 1 or 2 above, comprising: Multi-agent architecture module: Composed of multiple agents, each responsible for a specific energy device or energy node. Built on a large model agent architecture, it has independent perception, decision-making and execution capabilities. It collaborates with other agents through network communication to achieve information sharing, including sharing monitoring data, diagnostic results and optimization strategies. Model processing module: Includes large and small models. The large model is based on Qwen2.5 and incorporates professional knowledge of energy management and instrument maintenance through incremental pre-training. It uses instruction fine-tuning technology to understand and execute task instructions, and uses prompting engineering to decompose complex tasks through chain-of-thought reasoning. It also has the function of calling external tools, including PLC controllers and energy consumption analysis tools. The small models include DBSCAN model for low-dimensional fault diagnosis, AutoEncoder model for high-dimensional data fault detection, and moving average model, seasonal autoregressive model and nonlinear trend prediction model for time series data analysis in energy management, which assist the large model in fine-grained analysis. Data Acquisition and Management Module: Collects instrument operation status data, energy consumption data, and production environment data through sensors and smart meters. Cleans, preprocesses, and stores the collected data. Data cleaning uses advanced algorithms to remove outliers and noise. Preprocessing includes data normalization and feature extraction. The data storage management system uses efficient structures and algorithms to classify, store, and manage historical data. It uses big data technology and cloud computing platforms to process and analyze the data. System Integration and Interaction Module: Enables seamless integration with one or more production data platforms such as MES, APS, MRO, SCADA, and IoT. It interfaces with other systems through standardized interfaces, develops a user interface with visualization capabilities, displays instrument operating status, energy consumption data, analysis results, and optimization strategy information, and facilitates operators in setting parameters, issuing tasks, and monitoring the system. It also has user access control functions to ensure data security and operational compliance. Collaborative optimization control module: Used for information sharing, collaborative cooperation, and dynamic task allocation among intelligent agents. It establishes a negotiation and coordination mechanism to resolve resource conflicts and task competition. The dynamic task allocation mechanism allocates tasks based on task priority, urgency, agent load, and processing capacity. It can also adjust collaborative strategies and optimization objectives based on equipment aging, production plan changes, and energy supply fluctuations. It adopts a performance evaluation index system to monitor and evaluate key indicators such as energy utilization efficiency, instrument operation stability, and task processing timeliness in real time, and automatically adjusts collaborative strategies and optimization parameters based on the results.

4. The energy management multi-agent cooperative optimization system based on a large model agent architecture according to claim 3, characterized in that: The large model constructs a decision tree according to logical steps during chain-of-thinking reasoning, and adjusts the reasoning path and decision-making strategy based on real-time data and feedback. The tool calling function connects to external tools and interacts with data through secure interface technology. Incremental pre-training uses continuous learning algorithms to update the knowledge system to adapt to environmental changes.

5. The energy management multi-agent cooperative optimization system based on a large model agent architecture according to claim 3, characterized in that: The DBSCAN model improves localization accuracy by adjusting clustering parameters based on data distribution during low-dimensional fault diagnosis. The AutoEncoder model enhances the ability to identify abnormal patterns through reasonable encoding and decoding structures. The moving average model dynamically adjusts the time window size based on data. The seasonal autoregressive model optimizes parameters based on seasonal cycles. The nonlinear trend prediction model uses a fitting algorithm to adapt to trend changes.

6. The energy management multi-agent cooperative optimization system based on a large model agent architecture according to claim 3, characterized in that: The sensors and smart meters in the data acquisition and management module have high precision, high reliability and real-time transmission capabilities. The data cleaning algorithm intelligently identifies and processes abnormal data. The data storage management system supports fast retrieval, backup and recovery. Big data technology and cloud computing platform adopt distributed computing architecture and parallel processing algorithms to improve efficiency, and can intelligently cache and optimize storage strategies.

7. The energy management multi-agent cooperative optimization system based on a large model agent architecture according to claim 3, characterized in that: The standardized interfaces in the system integration and interaction module follow industry standards and specifications to ensure compatibility and interoperability. The user interface is visualized using intuitive charts and dynamic displays for easy operation. The user permission management function is comprehensive to ensure data security and operational compliance.

8. The energy management multi-agent cooperative optimization system based on a large model agent architecture according to claim 3, characterized in that, The negotiation and coordination mechanism in the collaborative optimization control module uses a distributed consensus algorithm to ensure decision consistency. The dynamic task allocation mechanism uses an intelligent algorithm to evaluate the status of tasks and agents. The resource conflict and task competition resolution strategy is based on priority and fairness principles. The performance evaluation index system comprehensively monitors key indicators and automatically adjusts strategies and parameters.

Images