Thermal station autonomous optimization operation method based on large and small model collaboration and intelligent agent

Abstract

The invention discloses a thermal station autonomous optimization operation method based on large and small model collaboration and an intelligent agent, and the method comprises the steps: carrying out the global empirical reasoning through a global intelligent safety protection large model through a thermal station intelligent agent, and forming a first protection decision; calling a preset global intelligent security protection small model to perform real-time quantitative analysis to form a second protection decision, and performing decision fusion to generate a global intelligent security protection decision; a heating station intelligent agent generates a heating station time-sharing regulation and control target through a trained equipment intelligent regulation and control large model, and a preset equipment intelligent regulation and control small model is called to generate a pump valve equipment regulation and control action instruction according to the task core requirement and the heating station time-sharing regulation and control target; and the heating station intelligent agent calls a preset predictive maintenance small model to predict maintenance equipment related parameters, and after a maintenance global target is generated through the trained predictive maintenance large model according to the task core demand, an equipment maintenance decision is generated in combination with maintenance equipment related parameter fusion.

Description

Technical Field

[0001] This invention belongs to the field of thermal power station operation technology, specifically relating to a method for autonomous optimization operation of thermal power stations based on large and small model collaboration and intelligent agents. Background Technology

[0002] Currently, some heating stations have achieved a certain degree of automation control through PLCs and other equipment, enabling them to complete basic operational adjustments of heating equipment. However, the automation level of many heating stations remains low, resulting in low operating efficiency and control accuracy, and a tendency to malfunction or become unstable. Furthermore, although some heating stations have installed remote monitoring systems that can transmit some operational data to the monitoring center, the remote monitoring functions are still incomplete, failing to comprehensively perform video security monitoring, noise identification and analysis, and abnormal gas identification and analysis, thus affecting the timely understanding and effective handling of the heating station's safe operation.

[0003] Furthermore, during the operation of a heating station, its core maintenance equipment, such as heat exchangers and water pumps, requires predictive maintenance based on changes in operating time and data. This involves predicting relevant core parameters of the equipment in advance to determine whether maintenance is necessary, thus avoiding any impact on heating quality. However, current heating stations, whether in terms of remote safety monitoring, intelligent control, or predictive maintenance, still possess relatively poor autonomous perception and decision-making capabilities. They cannot fully adapt to the different types of core tasks of heating stations and cannot achieve an intelligent upgrade from passive response to proactive prediction and precise execution.

[0004] Based on the aforementioned technical issues, a new method for autonomous optimization operation of thermal power stations based on large-scale model collaboration and intelligent agents needs to be designed. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to overcome the shortcomings of the prior art and provide a method for autonomous optimization operation of heat stations based on the collaboration of large and small models and intelligent agents. By using the collaborative decision-making method of intelligent agents, large models and small models, different operation decision-making modes are adopted to fit different core tasks, so as to realize the autonomous optimization operation of heat stations.

[0006] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: This invention provides a method for the autonomous optimization operation of a heat station based on size model collaboration and intelligent agents, which includes: After training the thermal station operation task database using the constructed thermal station intelligent agent to form a large-scale model for operation task discrimination, the model is used to extract keywords, reason about context, and apply rule-based judgment and case matching to output the results of thermal station operation task type discrimination and core task requirements. The thermal station operation task types include full-domain intelligent safety protection tasks, intelligent equipment control tasks, and predictive maintenance tasks. When the task type of the heating station is identified as a full-domain intelligent safety protection task, it enters the decision fusion mode: the heating station agent uses the trained full-domain intelligent safety protection big model to perform global experience reasoning based on the core requirements of the task to form the first protection decision, and calls the preset full-domain intelligent safety protection small model to perform real-time quantitative analysis based on the core requirements of the task to form the second protection decision. Then, the first protection decision and the second protection decision are fused to generate the full-domain intelligent safety protection decision. When the task type of the heating station is identified as an intelligent equipment control task, it enters the target guidance mode: the heating station agent generates the time-sharing control target of the heating station based on the core requirements of the task through the trained intelligent equipment control model, and calls the preset intelligent equipment control small model to generate the control action instructions of the heating station pump valve equipment based on the core requirements of the task and the time-sharing control target of the heating station. When the task type of the heating station is identified as a predictive maintenance task, it enters a hybrid mode of goal guidance and decision fusion: the heating station agent calls a preset predictive maintenance small model to predict the relevant parameters of the maintenance equipment based on the core requirements of the task, and the trained predictive maintenance large model generates a global maintenance goal based on the core requirements of the task, including maintenance cost, equipment failure impact data and maintenance priority, and then combines the relevant parameters of the maintenance equipment to generate equipment maintenance decisions.

[0007] Furthermore, the comprehensive intelligent security protection tasks include video security monitoring, noise identification and analysis, and gas identification and analysis; the video security monitoring includes fire and smoke identification, leakage alarm, and area intrusion alarm; the noise identification and analysis includes abnormal noise analysis and vibration monitoring and analysis; the gas identification and analysis includes fire odor early warning, electrical equipment fault odor diagnosis, and air quality assessment. The intelligent control task of the equipment is the control of pumps and valves in the heating station. The predictive maintenance tasks include predictive maintenance for scaling in heat exchanger systems and predictive maintenance for pump systems; the predictive maintenance for pump systems includes bearing wear early warning, cavitation monitoring, and mechanical seal failure assessment.

[0008] Furthermore, after training the thermal station operation task database using the constructed thermal station intelligent agent to form a large-scale operation task discrimination model, the model is used for keyword extraction, contextual reasoning, rule-based judgment, and case matching to output the thermal station operation task type discrimination results and core task requirements, including: The constructed intelligent agent for heating stations is used to obtain historical operating data and industry standards for heating station operation, and after data annotation processing, a database of heating station operation tasks is constructed. Using the task database and labeled keywords and context as inputs, and the task type and core requirements as outputs, a task discrimination model is formed by selecting the base model to learn professional knowledge of thermal station operation and fine-tuning it using typical task cases. When receiving unstructured natural language instructions from heating station operators, the heating station agent invokes the task discrimination model to extract core keywords related to the task and assign keyword weights, infer the implicit intent and context of the instructions, and then uses keyword and core requirement rules to determine the type of the task and performs case matching from the heating station's task data. Finally, it outputs structured data containing the task type discrimination results and the core requirements of the task, including: task ID, task type, constraints, core parameters, associated equipment, and task execution time limit.

[0009] Furthermore, the first protection decision, formed by the global intelligent security protection model trained by the thermal station agent based on the core requirements of the task, through global experience reasoning, includes: A comprehensive intelligent safety protection knowledge base for heating stations is constructed by acquiring relevant data and safety specification manuals for the comprehensive intelligent safety protection of heating stations using the constructed intelligent agent of the heating station. Semantic adaptation training and fine-tuning training using specific safety protection tasks are performed on the selected base large model to obtain a comprehensive intelligent safety protection large model. The relevant data for comprehensive intelligent safety protection includes video security monitoring, noise recognition, and gas recognition data. For each subtask under the overall intelligent security protection task, a corresponding protection decision chain is preset to indicate the protection decision steps, and a corresponding structured protection decision prompt template is preset for the protection decision chain of different subtasks. When a comprehensive intelligent safety protection task is received from the heating station operation staff, the comprehensive intelligent safety protection big model is used to select the corresponding sub-task protection decision chain from the preset protection decision chain based on the core requirements of the task, and the RAG retrieval enhancement technology is called to retrieve relevant protection information from the heating station's comprehensive intelligent safety protection knowledge base, including historical similar protection tasks, safety protection identification and analysis, safety protection operations, and core protection parameters. By leveraging the comprehensive intelligent security protection model and guided by structured protection decision prompt templates, relevant protection information and core task requirements are integrated and retrieved. Following the decision-making steps of the protection decision chain, global experience reasoning is performed step by step to form the first protection decision.

[0010] Furthermore, after the preset full-domain intelligent security protection mini-model performs real-time quantitative analysis based on the core requirements of the task to form a second protection decision, the first protection decision and the second protection decision are fused to generate a full-domain intelligent security protection decision, including: The system includes pre-defined intelligent security protection mini-models, such as: fire and smoke recognition mini-models, liquid leak recognition mini-models, and area intrusion alarm mini-models, all established using image recognition algorithms; equipment abnormal noise recognition mini-models and equipment abnormal vibration monitoring mini-models, all established using machine learning algorithms to extract sound features and train models; and fire odor early warning mini-models, electrical equipment fault odor diagnosis mini-models, and air quality assessment mini-models, all established using machine learning algorithms to extract statistical, temporal, and correlation features of different gas signals and, after gas signal feature learning. The constructed intelligent agent of the thermal station, based on the type of the received full-domain intelligent security protection task and the core requirements of the task, calls the corresponding preset sub-task small model to perform real-time quantitative analysis and form a second protection decision. A lightweight XGBoost model is used to train a weight allocation model for the first and second protection decisions. The first and second protection decisions are then dynamically fused with weights to generate a global intelligent security protection decision.

[0011] Furthermore, the intelligent control model of the heating station, trained by the intelligent agent, generates time-segmented control targets for the heating station based on the core requirements of the task, and calls a preset intelligent control model to generate control action instructions for the heating station's pumps and valves based on the core requirements of the task and the time-segmented control targets, including: A knowledge base for the regulation of pumps and valves in a heating station is constructed by acquiring relevant data and manuals on the regulation of pumps and valves. Semantic adaptation training is performed on a selected base model, and fine-tuning training is conducted using pump and valve regulation cases during peak, valley, and flat load periods to obtain a large-scale intelligent regulation model for the equipment. The relevant data on the regulation of pumps and valves in the heating station includes the station number, affiliated branch company, energy-saving type, heat user type, heating area, heating method, secondary water supply temperature regulation target, primary and secondary heat metering parameters, heat load value, and weather data. Using the knowledge base of pump and valve control of the heating station and the core parameters of the current intelligent control task of the equipment as input, and the secondary water supply temperature control target as output, when the intelligent control task of the equipment is received from the heating station operator, the intelligent control model of the equipment is used to generate the secondary water supply temperature control target of the heating station in different time periods according to the core requirements of the task. Taking the time-segmented secondary water supply temperature control target of the heating station and the current real-time operation data of the heating station as input, the preset valve opening PID control small model and water pump control DRL small model are called to run in coordination. Based on meeting the time-segmented secondary water supply temperature control target, the valve and water pump control action commands of the heating station for each time period are output.

[0012] Furthermore, during the decision-making process of the intelligent control task of the equipment, when an interruption event occurs, the valve opening PID control small model and the water pump regulation DRL small model send an interruption request and real-time operating data of the heating station to the intelligent control large model of the equipment. The intelligent control large model of the equipment generates a temporary control target and feeds it back to the valve opening PID control small model and the water pump regulation DRL small model for decision-making. The interruption events include water supply pressure exceeding the safety threshold, abnormal fluctuation of valve opening, daily energy consumption exceeding the standard, room temperature compliance rate being less than the preset value, and heating system malfunction.

[0013] Furthermore, the process of the heating station's intelligent agent calling a preset predictive maintenance mini-model to predict relevant parameters of the maintenance equipment based on the core requirements of the task includes: Using the collected heat exchanger heat exchange efficiency, running time, medium temperature and medium pH value as input, and the heat exchanger fouling thickness as output, after extracting core features and time-series features, the model is trained using the LSTM algorithm to establish a small model for predicting heat exchanger fouling thickness. Using the collected vibration signal of the water pump bearing as input and the bearing wear stage as output, after extracting the frequency domain and time domain features of the vibration signal, the model is trained using the SVM algorithm to establish a small model for predicting the wear stage of the water pump bearing. Using the collected high-frequency vibration signal of the water pump, high-frequency oscillation data of the current, inlet and outlet pressure, real-time flow rate and medium temperature as inputs, and the degree of water pump cavitation as output, after extracting the high-frequency vibration energy characteristics, high-frequency oscillation amplitude of the current and parameter fluctuation characteristics, the LightGBM algorithm is used to train the model and establish a small model for water pump cavitation monitoring. Using the changes in water pump pressure gradient and bearing temperature as inputs, and the risk of water pump bearing seal failure as output, after extracting the characteristics of pressure gradient, temperature change and time series features, the model is trained using the GRU algorithm to establish a small model for predicting the risk of water pump bearing seal failure. When a predictive maintenance task is received from the heating station operators, the heating station's intelligent agent calls the corresponding small model to output the predicted values ​​of relevant parameters for the maintenance equipment based on the task type and core requirements.

[0014] Furthermore, the trained predictive maintenance model generates a global maintenance objective based on the core requirements of the task, including maintenance costs, equipment failure impact data, and maintenance priorities. This objective is then combined with relevant equipment parameters to generate equipment maintenance decisions, including: By using the constructed intelligent agent of the heating station to obtain relevant data and maintenance manuals for the predictive maintenance of the heating station's water pumps and heat exchangers, a predictive maintenance knowledge base for the heating station is built. Semantic adaptation training is performed on the selected base model, and fine-tuning training is carried out using typical predictive maintenance cases to obtain a predictive maintenance model. Using a predictive maintenance big model, a global maintenance objective is generated based on the core requirements of the task, including maintenance cost budget, scope of equipment failure impact, and maintenance priority. This indicates the minimum maintenance cost budget, the minimum number of users affected by equipment failure and the minimum downtime for the current predictive maintenance task, as well as the equipment maintenance priority. By using a predictive maintenance big data model, we integrate the global maintenance goals and the predicted values ​​of relevant parameters of the maintenance equipment, extract and fuse the target features and parameter status features, and generate equipment maintenance decisions including maintenance timing, maintenance methods, resource allocation and cost budget.

[0015] Furthermore, the intelligent agent of the heating station includes a memory module, a reasoning module, and a tool invocation module; the memory module is used to store the heating station operation task database, the heating station's full-domain intelligent safety protection knowledge base, the heating station's pump and valve control knowledge base, the heating station's predictive maintenance knowledge base, the heating station's real-time operation data, and cache relevant decisions and control targets generated by large models; the reasoning module is used to train and apply various large models to perform relevant reasoning decisions; the tool invocation module is used to select and invoke relevant small models.

[0016] The beneficial effects of this invention are: (1) The intelligent control task of the equipment of the present invention adopts a target-guided mode of setting the target by a large model and executing by a small model. The large model combines heating load, ambient temperature and historical data to generate precise control targets in different time periods. The small model focuses on real-time data to quickly generate action commands such as pump and valve frequency conversion and flow regulation, thus avoiding the lag and experience bias of traditional manual control. (2) The task identification model of this invention quickly identifies the task type through keyword extraction, case matching, etc. Under different modes, the small model and the large model perform their respective functions: the small model focuses on real-time quantitative analysis, while the large model focuses on global empirical reasoning, avoiding the waste of computing power of a single model having to do both global planning and real-time calculation, improving the overall task processing efficiency, and meeting the rapid response needs of the heat station. (3) This invention targets multiple sub-tasks in the whole-domain intelligent safety protection task. It uses a large model to perform global experience reasoning based on safety specifications and historical accident cases, and a small model to perform quantitative analysis based on real-time data such as video surveillance, gas sensors, and vibration sensors, thereby improving the accuracy of risk identification and ensuring the safe operation of the heating station. The decision fusion mode avoids single logic defects through the dual decision fusion of the global process of the large model and the real-time response of the small model. Even if a single model has a deviation, the other model can fill in and correct it, which greatly improves the robustness of safety protection. (4) This invention targets predictive maintenance tasks. It utilizes a small model to predict parameter trends based on data such as equipment vibration, temperature, and energy consumption, while a large model combines maintenance costs, fault impact range, and maintenance priorities to generate a global objective. After fusion, it formulates a maintenance plan with optimal timing, lowest cost, and least impact. Compared to traditional periodic maintenance, equipment downtime is reduced, over-maintenance rate decreases, and maintenance costs drop. The large model integrates historical maintenance cases, equipment manuals, and cost data to generate global objectives and boundary constraints, while the small model provides data support based on actual equipment status, avoiding the subjectivity of traditional maintenance plans based on experience. Simultaneously, all decision-making processes are traceable, facilitating post-mortem analysis and optimization by maintenance personnel, and reducing management complexity. (5) The architecture of the large model and small model collaboration of the present invention has flexible adaptability: small heat stations can reuse the trained general large model and only need to fine-tune the small model according to the local equipment parameters; large heat stations can optimize the large model based on their own massive operating data, improve the task discrimination and decision-making pertinence, without the need to reconstruct the entire system and reduce the cost of technology implementation. (6) The present invention realizes a closed loop of automatic task identification, automatic decision generation and automatic command execution throughout the entire process. It does not require manual intervention in routine control, daily safety monitoring and maintenance plan formulation, etc., which can reduce repetitive operation and maintenance work, reduce labor costs, and allow operation and maintenance personnel to focus on core tasks such as complex fault handling and system optimization. In addition, all operation data is stored and recorded, and large models can be continuously iterated and trained based on these data. As the operation time increases, indicators such as task identification accuracy, decision optimization effect and energy saving rate will continue to improve, promoting the continuous upgrading of the operation level of the heating station. (7) The present invention solves the problems of inaccurate control, blind spots in safety, and reliance on experience in operation and maintenance of traditional heat stations by the innovative design of precise matching of task types to operation mode, complementary collaboration of large and small models, and overall scheduling of intelligent agents. It achieves multi-dimensional value of energy saving, high efficiency, safety and reliability, economic controllability, flexibility and scalability.

[0017] Other features and advantages will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention are realized and obtained through the structures particularly pointed out in the description and the drawings.

[0018] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0019] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0020] Figure 1 This is a flowchart of a method for autonomous optimization operation of a heat station based on big-small model collaboration and intelligent agents according to the present invention; Figure 2 This is a flowchart of the method for distinguishing the operation tasks of a heating station according to the present invention; Figure 3 The flowchart for the first protection decision method of the full-domain intelligent security protection model of this invention is shown below; Figure 4 The flowchart of the method for generating control action instructions for pump and valve equipment in a heating station according to the present invention is shown below. Figure 5 A flowchart for the equipment maintenance decision-making method of this invention is generated. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] like Figure 1 As shown, this embodiment provides a method for the autonomous optimization operation of a heat station based on big-small model collaboration and intelligent agents, which includes: After training the thermal station operation task database using the constructed thermal station intelligent agent to form a large-scale model for operation task discrimination, the model is used to extract keywords, reason about context, and apply rule-based judgment and case matching to output the results of thermal station operation task type discrimination and core task requirements. The thermal station operation task types include full-domain intelligent safety protection tasks, intelligent equipment control tasks, and predictive maintenance tasks. When the task type of the heating station is identified as a full-domain intelligent safety protection task, it enters the decision fusion mode: the heating station agent uses the trained full-domain intelligent safety protection big model to perform global experience reasoning based on the core requirements of the task to form the first protection decision, and calls the preset full-domain intelligent safety protection small model to perform real-time quantitative analysis based on the core requirements of the task to form the second protection decision. Then, the first protection decision and the second protection decision are fused to generate the full-domain intelligent safety protection decision. When the task type of the heating station is identified as an intelligent equipment control task, it enters the target guidance mode: the heating station agent generates the time-sharing control target of the heating station based on the core requirements of the task through the trained intelligent equipment control model, and calls the preset intelligent equipment control small model to generate the control action instructions of the heating station pump valve equipment based on the core requirements of the task and the time-sharing control target of the heating station. When the task type of the heating station is identified as a predictive maintenance task, it enters a hybrid mode of goal guidance and decision fusion: the heating station agent calls a preset predictive maintenance small model to predict the relevant parameters of the maintenance equipment based on the core requirements of the task, and the trained predictive maintenance large model generates a global maintenance goal based on the core requirements of the task, including maintenance cost, equipment failure impact data and maintenance priority, and then combines the relevant parameters of the maintenance equipment to generate equipment maintenance decisions.

[0023] In practical applications, task types and operating modes are matched to allow the intelligent agent of the heating station to leverage the global planning capabilities of the large model and the real-time execution capabilities of the small model in different task scenarios. Comprehensive intelligent safety protection tasks require rapid response and standardized handling. Relying solely on the global experience of the large model leads to poor real-time performance, while relying solely on the real-time perception of the small model lacks global experience. The small model generates quantitative responses based on collected sensor data, while the large model generates complete protection processes based on safety regulations and historical cases. Decision-making is then fused to dynamically balance the advantages of both approaches.

[0024] Intelligent equipment control requires precise execution within a clearly defined target domain, without the need for complex decision fusion. The key is to ensure the coupling between the secondary water supply temperature target and the execution of pump and valve equipment. A large model can generate time-specific secondary water supply temperature control targets based on global historical operating experience values, meeting the heat load demands of different heating periods while adhering to equipment operating safety thresholds. A smaller model can generate specific action execution instructions based on the control targets and real-time operating data of the heating station. This logic of using targets to determine the execution instructions of pump and valve equipment reduces decision redundancy, resulting in faster and more precise control response, thus meeting the time-specific dynamic control needs of the heating station.

[0025] Predictive maintenance tasks require the ability to predict maintenance parameters of equipment during operation and generate optimized maintenance plans. Relying solely on the global maintenance goals of a large model lacks support from the actual state of the equipment. Conversely, relying solely on parameter predictions from a small model lacks global goals and constraints. By using a small predictive maintenance model to predict parameter trends based on equipment operation data, data support is provided for maintenance decisions, avoiding over-maintenance or untimely maintenance. The global maintenance goals generated by the large model define the objectives and constraint boundaries for decision-making. Combining the parameter prediction results from the small model, the final maintenance decision is generated with optimal timing, lowest cost, and least impact, extending equipment lifespan and further reducing operating costs.

[0026] The intelligent agent at the heating station acts as a data and model connection hub, integrating the heating station's operational task database. It provides high-quality training data for the large-scale model that distinguishes operational tasks, ensuring the model can accurately learn task type features and core requirement extraction rules. Simultaneously, it provides training data support for subsequent large-scale models for various specialized tasks, ensuring the model adapts to the specific operational scenarios of the heating station. Furthermore, the agent manages all preset small models and various trained large models, automatically switching the corresponding operating mode based on task type and dynamically calling the appropriate model. The intelligent agent transforms the final decisions generated for each type of task into executable operational signals for the heating station, and continuously monitors the effectiveness of decision execution: for example, after the issuance of control commands, it collects equipment operating data in real time to verify whether the control target has been achieved; after the execution of safety protection decisions, it tracks whether the risks have been eliminated; after the execution of maintenance decisions, it monitors changes in equipment status, forming a closed loop of decision-making, execution, and feedback.

[0027] In this embodiment, the comprehensive intelligent security protection task includes video security monitoring, noise identification and analysis, and gas identification and analysis; the video security monitoring includes fire and smoke identification, leakage alarm, and area intrusion alarm; the noise identification and analysis includes abnormal noise analysis and vibration monitoring and analysis; the gas identification and analysis includes fire odor early warning, electrical equipment fault odor diagnosis, and air quality assessment. The intelligent control task of the equipment is the control of pumps and valves in the heating station. The predictive maintenance tasks include predictive maintenance for scaling in heat exchanger systems and predictive maintenance for pump systems; the predictive maintenance for pump systems includes bearing wear early warning, cavitation monitoring, and mechanical seal failure assessment.

[0028] In practical applications, fire and smoke recognition: using flame and smoke image recognition algorithms, fires and smoke are identified, and alarm information is generated once a fire or smoke event is detected; leakage alarm: intelligent analysis of monitoring images is performed to accurately identify liquid leakage characteristics, such as color, shape, and dynamic changes, effectively distinguishing between normal and abnormal situations, and automatically detecting and alarming liquid leakage in the monitoring scene in real time; area intrusion alarm: a monitoring area is defined, and an alarm is triggered when a target enters the preset monitoring area or reaches the trigger time within the area.

[0029] Abnormal noise analysis: Identify and analyze the operating noise of water pumps, heat exchangers, valves, piping systems, motors, etc., to determine whether the equipment is operating normally; Vibration monitoring analysis: Collect vibration signals and analyze data to monitor whether the vibration is within the normal range.

[0030] Fire Odor Warning: Before an open flame is detected by visual monitoring, a fire warning signal is issued in advance by detecting odors and smoke; Electrical Equipment Fault Odor Diagnosis: Monitoring electrical equipment faults will produce special odors, and timely detection of potential electrical equipment faults; Air Quality Assessment: Continuously monitoring air changes in the heating station and assessing the air quality.

[0031] Predictive maintenance of heat exchanger system structure: Monitor heat exchange efficiency, operating time, medium temperature, and medium pH value; establish a model to predict the scale thickness of the heat exchanger; generate cleaning and maintenance suggestions based on the scale thickness. Bearing wear early warning: Collect bearing vibration frequency and other data in real time; extract abnormal vibration peak values; report bearing wear risk when the peak value exceeds the normal threshold. Cavitation monitoring: Water pump cavitation refers to the phenomenon where a large number of bubbles are generated in the water flow when the pressure at the water pump inlet drops to the saturated vapor pressure of the medium at the corresponding temperature. The bubbles burst rapidly after entering the high-pressure area with the water flow, causing impact and damage. Mechanical seal failure judgment: Water pump bearing seals are key components to ensure the normal operation of the water pump. Their core function is to prevent leakage of the medium inside the pump and prevent external air and impurities from entering the pump body. Seal failure refers to the decline in sealing performance of the sealing components due to wear, aging, improper installation, or external factors.

[0032] like Figure 2 As shown, in this embodiment, after training the thermal station operation task database using the constructed thermal station intelligent agent to form an operation task discrimination model, the model is used for keyword extraction, contextual reasoning, rule-based judgment, and case matching to output the thermal station operation task type discrimination result and core task requirements, including: The constructed intelligent agent for heating stations is used to obtain historical operating data and industry standards for heating station operation, and after data annotation processing, a database of heating station operation tasks is constructed. Using the task database and labeled keywords and context as inputs, and the task type and core requirements as outputs, a task discrimination model is formed by selecting the base model to learn professional knowledge of thermal station operation and fine-tuning it using typical task cases. When receiving unstructured natural language instructions from heating station operators, the heating station agent invokes the task discrimination model to extract core keywords related to the task and assign keyword weights, infer the implicit intent and context of the instructions, and then uses keyword and core requirement rules to determine the type of the task and performs case matching from the heating station's task data. Finally, it outputs structured data containing the task type discrimination results and the core requirements of the task, including: task ID, task type, constraints, core parameters, associated equipment, and task execution time limit.

[0033] In this application, lightweight, open-source large models that support semantic understanding are preferred for the selection of base models to adapt to the edge deployment scenario of heat stations, such as ChatGLM3-6B and Llama2 (7B / 13B).

[0034] Implicit Intent Reasoning: Based on keyword association and professional knowledge, the underlying needs of instructions are uncovered. For example, if an instruction states that a water pump is recently making a lot of noise and is prone to problems, the keywords "noise" and "problem" are extracted, and the implicit intent is inferred to be "predictive maintenance (abnormal noise analysis)." The core need is to monitor water pump noise data and determine if there are any anomalies. Contextual scenario matching is performed by combining a contextual scenario library to determine the instruction scenario.

[0035] Keyword and Core Requirement Rule Determination: Establish mapping rules between keyword combinations and task type + core requirement. For example, if the keywords include "cavitation, vibration, wear, and seal failure," then the task type is predictive maintenance, and the core requirements are specific equipment, equipment operating data, anomaly monitoring, and maintenance plan generation. Runtime Task Database Case Matching: Retrieve similar historical cases from the runtime task database to assist in verifying the determination results. The specific logic is to calculate the matching degree between the input command and the historical case library based on keyword vector similarity.

[0036] like Figure 3 As shown, in this embodiment, the first protection decision is formed by the thermal station agent through a trained global intelligent security protection model based on the core requirements of the task, using global experience reasoning. This includes: A comprehensive intelligent safety protection knowledge base for heating stations is constructed by acquiring relevant data and safety specification manuals for the comprehensive intelligent safety protection of heating stations using the constructed intelligent agent of the heating station. Semantic adaptation training and fine-tuning training using specific safety protection tasks are performed on the selected base large model to obtain a comprehensive intelligent safety protection large model. The relevant data for comprehensive intelligent safety protection includes video security monitoring, noise recognition, and gas recognition data. For each subtask under the overall intelligent security protection task, a corresponding protection decision chain is preset to indicate the protection decision steps, and a corresponding structured protection decision prompt template is preset for the protection decision chain of different subtasks. When a comprehensive intelligent safety protection task is received from the heating station operation staff, the comprehensive intelligent safety protection big model is used to select the corresponding sub-task protection decision chain from the preset protection decision chain based on the core requirements of the task, and the RAG retrieval enhancement technology is called to retrieve relevant protection information from the heating station's comprehensive intelligent safety protection knowledge base, including historical similar protection tasks, safety protection identification and analysis, safety protection operations, and core protection parameters. By leveraging the comprehensive intelligent security protection model and guided by structured protection decision prompt templates, relevant protection information and core task requirements are integrated and retrieved. Following the decision-making steps of the protection decision chain, global experience reasoning is performed step by step to form the first protection decision.

[0037] In practical applications, semantic adaptation for heating station safety protection enables large models to understand the professional terminology, scenario logic, and data associations related to heating station safety protection; fine-tuning training for safety protection tasks enables large models to master the logic of task requirements, decision chain selection, information integration, and decision generation.

[0038] Taking video security monitoring as an example, a pre-defined protection decision chain is established: 1) Confirm the authenticity of the alarm (video feature matching); 2) Locate the alarm area (associate with electronic map); 3) Retrieve safety regulations (fire emergency response procedures); 4) Call up related data (fire monitoring data); 5) Generate a response plan (activate fire extinguishing devices, evacuate personnel, and shut down relevant equipment).

[0039] For each sub-task decision chain, a prompt template is designed to guide the large model to integrate information and generate decisions step by step. The template includes a task description, decision chain guidance, information retrieval and integration requirements, and output format.

[0040] Matching the corresponding protection decision chain: Based on the task type, a unique corresponding chain is selected from the preset protection decision chains. RAG retrieval enhancement technology application and information retrieval: The large model transforms core requirements into search keywords, quickly matches security specification manuals and historical cases based on keyword indexes, and supplements relevant knowledge and filters search results based on vector indexes for semantic retrieval.

[0041] In this embodiment, after the preset full-domain intelligent security protection mini-model performs real-time quantitative analysis based on the core requirements of the task to form a second protection decision, the first protection decision and the second protection decision are fused to generate a full-domain intelligent security protection decision, including: The system includes pre-defined intelligent security protection mini-models, such as: fire and smoke recognition mini-models, liquid leak recognition mini-models, and area intrusion alarm mini-models, all established using image recognition algorithms; equipment abnormal noise recognition mini-models and equipment abnormal vibration monitoring mini-models, all established using machine learning algorithms to extract sound features and train models; and fire odor early warning mini-models, electrical equipment fault odor diagnosis mini-models, and air quality assessment mini-models, all established using machine learning algorithms to extract statistical, temporal, and correlation features of different gas signals and, after gas signal feature learning. The constructed intelligent agent of the thermal station, based on the type of the received full-domain intelligent security protection task and the core requirements of the task, calls the corresponding preset sub-task small model to perform real-time quantitative analysis and form a second protection decision. A lightweight XGBoost model is used to train a weight allocation model for the first and second protection decisions. The first and second protection decisions are then dynamically fused with weights to generate a global intelligent security protection decision.

[0042] In practical applications, YOLOv8 can be used as an image recognition algorithm. When there is a fire, extract the color, shape, and diffusion features of the flames; when there is smoke, extract the grayscale features, texture features, and spatiotemporal features of smoke diffusion.

[0043] Sound features were extracted using machine learning algorithms, and models were trained to establish small models for identifying abnormal noise in equipment and monitoring abnormal vibration in equipment. Feature extraction employed MFCC Mel-frequency cepstral coefficients and spectral analysis, while training utilized the GRU algorithm.

[0044] The statistical, temporal, and correlation features of different gas signals were extracted using machine learning algorithms. After learning the gas signal features, small models for fire odor early warning, electrical equipment fault odor diagnosis, and air quality assessment were established. Multilayer perceptron (MLP) was used for feature extraction, and LightGBM was used for model training.

[0045] A lightweight XGBoost model is used to train a weight allocation model for the first and second protection decisions. The first and second protection decisions are then dynamically fused using weights to generate a comprehensive intelligent security protection decision, including: 1) Historical decision-making cases: Data on the execution effects of the first protection decision, the second protection decision, and the integrated decision in the historical safety protection tasks of the heating station; Labeling rules: Label the success rate of risk handling after the implementation of the integrated decision as a label, assign the optimal weight, and form a training set; 2) Use the training set to train a lightweight XGBoost model, obtain the optimal weights of the first protection decision and the second protection decision, and perform weighted fusion to generate a global intelligent security protection decision.

[0046] like Figure 4 As shown, in this embodiment, the intelligent agent of the heating station generates time-sharing control targets for the heating station based on the core requirements of the task using a trained large-scale intelligent equipment control model. It then calls a preset small-scale intelligent equipment control model to generate control action instructions for the heating station's pumps and valves based on the core requirements of the task and the time-sharing control targets. These instructions include: A knowledge base for the regulation of pumps and valves in a heating station is constructed by acquiring relevant data and manuals on the regulation of pumps and valves. Semantic adaptation training is performed on a selected base model, and fine-tuning training is conducted using pump and valve regulation cases during peak, valley, and flat load periods to obtain a large-scale intelligent regulation model for the equipment. The relevant data on the regulation of pumps and valves in the heating station includes the station number, affiliated branch company, energy-saving type, heat user type, heating area, heating method, secondary water supply temperature regulation target, primary and secondary heat metering parameters, heat load value, and weather data. Using the knowledge base of pump and valve control of the heating station and the core parameters of the current intelligent control task of the equipment as input, and the secondary water supply temperature control target as output, when the intelligent control task of the equipment is received from the heating station operator, the intelligent control model of the equipment is used to generate the secondary water supply temperature control target of the heating station in different time periods according to the core requirements of the task. Taking the time-segmented secondary water supply temperature control target of the heating station and the current real-time operation data of the heating station as input, the preset valve opening PID control small model and water pump control DRL small model are called to run in coordination. Based on meeting the time-segmented secondary water supply temperature control target, the valve and water pump control action commands of the heating station for each time period are output.

[0047] It should be noted that the establishment of the small model for PID control of valve opening includes: 1) Control objective: To stabilize the secondary water supply temperature within the time-period targets given by the large model; Input data: Time-segmented temperature target, real-time secondary water supply temperature, real-time valve opening, real-time load change; Output data: Valve opening adjustment amount; 2) Construct a basic PID controller; 3) The basic PID controller parameters are fixed and difficult to adapt to the load fluctuation scenario of the heating station. It is necessary to optimize the PID parameters by self-tuning. The historical valve opening, temperature deviation and load data are used as training data for model training to learn the mapping relationship between load and PID parameters and realize real-time parameter self-tuning.

[0048] The establishment of the small-scale model for pump regulation DRL includes: 1) Framework definition of DRL small model Environmental status space: secondary water supply temperature deviation, current water pump frequency, real-time heat load, energy consumption; Action range: Pump frequency adjustment range; Reward function: Achieving the target temperature while maximizing energy consumption; 2) Environmental simulation: Construct a simulation environment for the operation of water pumps in the heating station to simulate the relationship between pump frequency, temperature, and energy consumption under different loads; 3) Offline training Initialize the policy network and value network of the DDPG algorithm; Interacting in the simulation environment: The policy network outputs pump frequency adjustment actions, the value network evaluates the value of the actions, and the network parameters are updated through gradient descent; Training termination condition: Continue training for multiple rounds until the temperature reaches the target and the energy consumption is lower than the preset value.

[0049] In this embodiment, during the decision-making process of the intelligent control task of the equipment, when an interruption event occurs, the valve opening PID control small model and the water pump regulation DRL small model send an interruption request and real-time operating data of the heating station to the intelligent control large model of the equipment. The intelligent control large model of the equipment generates a temporary control target and feeds it back to the valve opening PID control small model and the water pump regulation DRL small model for decision-making. The interruption events include water supply pressure exceeding the safety threshold, abnormal fluctuation of valve opening, daily energy consumption exceeding the standard, room temperature compliance rate being less than the preset value, and heating system failure.

[0050] In this embodiment, the step of the intelligent agent of the heating station calling a preset predictive maintenance mini-model to predict relevant parameters of the maintenance equipment based on the core requirements of the task includes: Using the collected heat exchanger heat exchange efficiency, running time, medium temperature and medium pH value as input, and the heat exchanger fouling thickness as output, after extracting core features and time-series features, the model is trained using the LSTM algorithm to establish a small model for predicting heat exchanger fouling thickness. Using the collected vibration signal of the water pump bearing as input and the bearing wear stage as output, after extracting the frequency domain and time domain features of the vibration signal, the model is trained using the SVM algorithm to establish a small model for predicting the wear stage of the water pump bearing. Using the collected high-frequency vibration signal of the water pump, high-frequency oscillation data of the current, inlet and outlet pressure, real-time flow rate and medium temperature as inputs, and the degree of water pump cavitation as output, after extracting the high-frequency vibration energy characteristics, high-frequency oscillation amplitude of the current and parameter fluctuation characteristics, the LightGBM algorithm is used to train the model and establish a small model for water pump cavitation monitoring. Using the changes in water pump pressure gradient and bearing temperature as inputs, and the risk of water pump bearing seal failure as output, after extracting the characteristics of pressure gradient, temperature change and time series features, the model is trained using the GRU algorithm to establish a small model for predicting the risk of water pump bearing seal failure. When a predictive maintenance task is received from the heating station operators, the heating station's intelligent agent calls the corresponding small model to output the predicted values ​​of relevant parameters for the maintenance equipment based on the task type and core requirements.

[0051] like Figure 5 As shown, in this embodiment, the trained predictive maintenance model generates a global maintenance objective based on the core requirements of the task, including maintenance costs, equipment failure impact data, and maintenance priorities. This objective is then combined with relevant equipment parameters to generate equipment maintenance decisions, including: By using the constructed intelligent agent of the heating station to obtain relevant data and maintenance manuals for the predictive maintenance of the heating station's water pumps and heat exchangers, a predictive maintenance knowledge base for the heating station is built. Semantic adaptation training is performed on the selected base model, and fine-tuning training is carried out using typical predictive maintenance cases to obtain a predictive maintenance model. Using a predictive maintenance big model, a global maintenance objective is generated based on the core requirements of the task, including maintenance cost budget, scope of equipment failure impact, and maintenance priority. This indicates the minimum maintenance cost budget, the minimum number of users affected by equipment failure and the minimum downtime for the current predictive maintenance task, as well as the equipment maintenance priority. By using a predictive maintenance big data model, we integrate the global maintenance goals and the predicted values ​​of relevant parameters of the maintenance equipment, extract and fuse the target features and parameter status features, and generate equipment maintenance decisions including maintenance timing, maintenance methods, resource allocation and cost budget.

[0052] In practical applications, the inputs and core criteria for global target generation include: The inputs to the maintenance cost budget are equipment type, abnormal fault type, and severity. The core basis is the spare parts price list, personnel hour costs, and historical maintenance cost cases. The quantitative logic is to match the historical maintenance costs of similar abnormal faults, combine them with the current spare parts inventory and personnel scheduling, and output the minimum cost budget. The inputs for the impact of equipment failure are equipment location, equipment function, and heating area correlation. The core basis is the heating station network topology map, equipment-user correlation table, and historical failure impact records. The quantification logic is as follows: the heating area served by the equipment is determined based on the network topology map, the number of affected users is counted by combining the user correlation table, and the minimum heating outage duration is predicted by referring to historical cases. The inputs for maintenance priority are the severity of the fault, the scope of impact, and the heating period. The core basis is the fault level classification standard. The quantification logic is to set priority determination rules based on the severity of the fault, the number of affected users, and the peak / off-peak heating periods.

[0053] In this embodiment, the intelligent agent of the heating station includes a memory module, a reasoning module, and a tool invocation module. The memory module is used to store the heating station operation task database, the heating station's comprehensive intelligent safety protection knowledge base, the heating station's pump and valve control knowledge base, the heating station's predictive maintenance knowledge base, the heating station's real-time operation data, and cache relevant decisions and control targets generated by large models. The reasoning module is used to train and apply various large models to perform relevant reasoning decisions. The tool invocation module is used to select and invoke relevant small models.

[0054] It should be noted that the memory module is responsible for data storage, indexing, updating, and caching, providing the thinking and reasoning module with full, accurate, and real-time data support, while also retaining the agent's decision history and operational status. Based on the data source of the memory module, the thinking and reasoning module trains and deploys various large models to achieve core reasoning capabilities such as task type discrimination, full-domain intelligent security protection, intelligent pump and valve control, and predictive maintenance. The tool invocation module serves as a bridge between the thinking and reasoning module and external tools (various small models, hardware control interfaces, and third-party systems), realizing a closed loop of on-demand selection, automatic invocation, and result feedback, thereby expanding the agent's execution capabilities.

[0055] In the several embodiments provided in this application, it should be understood that the disclosed systems and methods can also be implemented in other ways. The system embodiments described above are merely illustrative; for example, the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code, which contains one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram and / or flowchart, and combinations of blocks in block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0056] Furthermore, the functional modules in the various embodiments of this invention can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part. If the function is implemented as a software functional module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory, random access memory, magnetic disks, or optical disks.

[0057] Based on the above-described preferred embodiments of the present invention, and through the foregoing description, those skilled in the art can make various changes and modifications without departing from the inventive concept. The technical scope of this invention is not limited to the contents of the specification, but must be determined according to the scope of the claims.

Claims

1. A method for autonomous optimization operation of a heat station based on large-scale model collaboration and intelligent agents, characterized in that, It includes: After training the thermal station operation task database using the constructed thermal station intelligent agent to form a large model for operation task discrimination, the large model is used to extract keywords, reason about context, and make judgments based on rules and case matching to output the results of thermal station operation task type discrimination and core task requirements. The types of tasks performed at the heating station include comprehensive intelligent safety protection tasks, intelligent equipment control tasks, and predictive maintenance tasks. When the task type of the heating station is identified as a full-domain intelligent safety protection task, it enters the decision fusion mode: the heating station agent uses the trained full-domain intelligent safety protection big model to perform global experience reasoning based on the core requirements of the task to form the first protection decision, and calls the preset full-domain intelligent safety protection small model to perform real-time quantitative analysis based on the core requirements of the task to form the second protection decision. Then, the first protection decision and the second protection decision are fused to generate the full-domain intelligent safety protection decision. When the task type of the heating station is identified as an intelligent equipment control task, it enters the target guidance mode: the heating station agent generates the time-sharing control target of the heating station based on the core requirements of the task through the trained intelligent equipment control model, and calls the preset intelligent equipment control small model to generate the control action instructions of the heating station pump valve equipment based on the core requirements of the task and the time-sharing control target of the heating station. When the task type of the heating station is identified as a predictive maintenance task, it enters a hybrid mode of goal guidance and decision fusion: the heating station agent calls a preset predictive maintenance small model to predict the relevant parameters of the maintenance equipment based on the core requirements of the task, and the trained predictive maintenance large model generates a global maintenance goal based on the core requirements of the task, including maintenance cost, equipment failure impact data and maintenance priority, and then combines the relevant parameters of the maintenance equipment to generate equipment maintenance decisions.

2. The method for autonomous optimization operation of a heating station according to claim 1, characterized in that, The comprehensive intelligent security protection tasks include video security monitoring, noise identification and analysis, and gas identification and analysis; the video security monitoring includes fire and smoke identification, leakage alarm, and area intrusion alarm; the noise identification and analysis includes abnormal noise analysis and vibration monitoring and analysis; the gas identification and analysis includes fire odor early warning, electrical equipment fault odor diagnosis, and air quality assessment. The intelligent control task of the equipment is the control of pumps and valves in the heating station. The predictive maintenance tasks include predictive maintenance for scaling in heat exchanger systems and predictive maintenance for pump systems; the predictive maintenance for pump systems includes bearing wear early warning, cavitation monitoring, and mechanical seal failure assessment.

3. The method for autonomous optimization operation of a heating station according to claim 2, characterized in that, After training the constructed heating station intelligent agent on the heating station operation task database to form a large-scale model for task differentiation, the model is used for keyword extraction, contextual reasoning, rule-based judgment, and case matching to output the heating station operation task type differentiation results and core task requirements, including: The constructed intelligent agent for heating stations is used to obtain historical operating data and industry standards for heating station operation, and after data annotation processing, a database of heating station operation tasks is constructed. Using the task database and labeled keywords and context as inputs, and the task type and core requirements as outputs, a task discrimination model is formed by selecting the base model to learn professional knowledge of thermal station operation and fine-tuning it using typical task cases. When receiving unstructured natural language instructions from heating station operators, the heating station agent invokes the task discrimination model to extract core keywords related to the task and assign keyword weights, infer the implicit intent and context of the instructions, and then uses keyword and core requirement rules to determine the type of the task and performs case matching from the heating station's task data. Finally, it outputs structured data containing the task type discrimination results and the core requirements of the task, including: task ID, task type, constraints, core parameters, associated equipment, and task execution time limit.

4. The method for autonomous optimization operation of a heating station according to claim 1, characterized in that, The first protection decision, formed by the global intelligent security protection model trained by the thermal station agent based on the core requirements of the task, through global experience-based reasoning, includes: A comprehensive intelligent safety protection knowledge base for heating stations is constructed by acquiring relevant data and safety specification manuals for the comprehensive intelligent safety protection of heating stations using the constructed intelligent agent of the heating station. Semantic adaptation training and fine-tuning training using specific safety protection tasks are performed on the selected base large model to obtain a comprehensive intelligent safety protection large model. The relevant data for comprehensive intelligent safety protection includes video security monitoring, noise recognition, and gas recognition data. For each subtask under the overall intelligent security protection task, a corresponding protection decision chain is preset to indicate the protection decision steps, and a corresponding structured protection decision prompt template is preset for the protection decision chain of different subtasks. When a comprehensive intelligent safety protection task is received from the heating station operation staff, the comprehensive intelligent safety protection big model is used to select the corresponding sub-task protection decision chain from the preset protection decision chain based on the core requirements of the task, and the RAG retrieval enhancement technology is called to retrieve relevant protection information from the heating station's comprehensive intelligent safety protection knowledge base, including historical similar protection tasks, safety protection identification and analysis, safety protection operations, and core protection parameters. By leveraging the comprehensive intelligent security protection model and guided by structured protection decision prompt templates, relevant protection information and core task requirements are integrated and retrieved. Following the decision-making steps of the protection decision chain, global experience reasoning is performed step by step to form the first protection decision.

5. The method for autonomous optimization operation of a heating station according to claim 1, characterized in that, After the preset full-domain intelligent security protection mini-model performs real-time quantitative analysis based on the core requirements of the task to form a second protection decision, the first protection decision and the second protection decision are fused to generate a full-domain intelligent security protection decision, including: The system includes pre-defined intelligent security protection mini-models, such as: fire and smoke recognition mini-models, liquid leak recognition mini-models, and area intrusion alarm mini-models, all established using image recognition algorithms; equipment abnormal noise recognition mini-models and equipment abnormal vibration monitoring mini-models, all established using machine learning algorithms to extract sound features and train models; and fire odor early warning mini-models, electrical equipment fault odor diagnosis mini-models, and air quality assessment mini-models, all established using machine learning algorithms to extract statistical, temporal, and correlation features of different gas signals and, after gas signal feature learning. The constructed intelligent agent of the thermal station, based on the type of the received full-domain intelligent security protection task and the core requirements of the task, calls the corresponding preset sub-task small model to perform real-time quantitative analysis and form a second protection decision. A lightweight XGBoost model is used to train a weight allocation model for the first and second protection decisions. The first and second protection decisions are then dynamically fused with weights to generate a global intelligent security protection decision.

6. The method for autonomous optimization operation of a heating station according to claim 1, characterized in that, The intelligent control model of the heating station, trained by the intelligent agent, generates time-segmented control targets for the heating station based on the core requirements of the task. It then calls a preset intelligent control model to generate control action commands for the heating station's pumps and valves based on the core requirements of the task and the time-segmented control targets. These commands include: A knowledge base for the regulation of pumps and valves in a heating station is constructed by acquiring relevant data and manuals on the regulation of pumps and valves. Semantic adaptation training is performed on a selected base model, and fine-tuning training is conducted using pump and valve regulation cases during peak, valley, and flat load periods to obtain a large-scale intelligent regulation model for the equipment. The relevant data on the regulation of pumps and valves in the heating station includes the station number, affiliated branch company, energy-saving type, heat user type, heating area, heating method, secondary water supply temperature regulation target, primary and secondary heat metering parameters, heat load value, and weather data. Using the knowledge base of pump and valve control of the heating station and the core parameters of the current intelligent control task of the equipment as input, and the secondary water supply temperature control target as output, when the intelligent control task of the equipment is received from the heating station operator, the intelligent control model of the equipment is used to generate the secondary water supply temperature control target of the heating station in different time periods according to the core requirements of the task. Taking the time-segmented secondary water supply temperature control target of the heating station and the current real-time operation data of the heating station as input, the preset valve opening PID control small model and water pump control DRL small model are called to run in coordination. Based on meeting the time-segmented secondary water supply temperature control target, the valve and water pump control action commands of the heating station for each time period are output.

7. The method for autonomous optimization operation of a heating station according to claim 6, characterized in that, In the decision-making process of intelligent equipment control tasks, when an interruption event occurs, the valve opening PID control mini-model and the water pump regulation DRL mini-model send an interruption request and real-time operating data of the heating station to the intelligent equipment control big model. The intelligent equipment control big model then generates a temporary control target and feeds it back to the valve opening PID control mini-model and the water pump regulation DRL mini-model for decision-making. The interruption events include water supply pressure exceeding the safety threshold, abnormal fluctuation of valve opening, daily energy consumption exceeding the standard, room temperature compliance rate being less than the preset value, and heating system malfunction.

8. The method for autonomous optimization operation of a heating station according to claim 1, characterized in that, The process of the intelligent agent at the heating station calling a pre-set predictive maintenance mini-model to predict relevant parameters of the maintenance equipment based on the core requirements of the task includes: Using the collected heat exchanger heat exchange efficiency, running time, medium temperature and medium pH value as input, and the heat exchanger fouling thickness as output, after extracting core features and time-series features, the model is trained using the LSTM algorithm to establish a small model for predicting heat exchanger fouling thickness. Using the collected vibration signal of the water pump bearing as input and the bearing wear stage as output, after extracting the frequency domain and time domain features of the vibration signal, the model is trained using the SVM algorithm to establish a small model for predicting the wear stage of the water pump bearing. Using the collected high-frequency vibration signal of the water pump, high-frequency oscillation data of the current, inlet and outlet pressure, real-time flow rate and medium temperature as inputs, and the degree of water pump cavitation as output, after extracting the high-frequency vibration energy characteristics, high-frequency oscillation amplitude of the current and parameter fluctuation characteristics, the LightGBM algorithm is used to train the model and establish a small model for water pump cavitation monitoring. Using the changes in water pump pressure gradient and bearing temperature as inputs, and the risk of water pump bearing seal failure as output, after extracting the characteristics of pressure gradient, temperature change and time series features, the model is trained using the GRU algorithm to establish a small model for predicting the risk of water pump bearing seal failure. When a predictive maintenance task is received from the heating station operators, the heating station's intelligent agent calls the corresponding small model to output the predicted values ​​of relevant parameters for the maintenance equipment based on the task type and core requirements.

9. The method for autonomous optimization operation of a heating station according to claim 1, characterized in that, The trained predictive maintenance model generates a global maintenance objective based on the core requirements of the task, including maintenance costs, equipment failure impact data, and maintenance priorities. This objective is then combined with relevant equipment parameters to generate equipment maintenance decisions, including: By using the constructed intelligent agent of the heating station to obtain relevant data and maintenance manuals for the predictive maintenance of the heating station's water pumps and heat exchangers, a predictive maintenance knowledge base for the heating station is built. Semantic adaptation training is performed on the selected base model, and fine-tuning training is carried out using typical predictive maintenance cases to obtain a predictive maintenance model. Using a predictive maintenance big model, a global maintenance objective is generated based on the core requirements of the task, including maintenance cost budget, scope of equipment failure impact, and maintenance priority. This indicates the minimum maintenance cost budget, the minimum number of users affected by equipment failure and the minimum downtime for the current predictive maintenance task, as well as the equipment maintenance priority. By using a predictive maintenance big data model, we integrate the global maintenance goals and the predicted values ​​of relevant parameters of the maintenance equipment, extract and fuse the target features and parameter status features, and generate equipment maintenance decisions including maintenance timing, maintenance methods, resource allocation and cost budget.

10. The method for autonomous optimization operation of a heating station according to claim 1, characterized in that, The intelligent agent of the heating station includes a memory module, a reasoning module, and a tool invocation module. The memory module is used to store the heating station's operation task database, the heating station's comprehensive intelligent safety protection knowledge base, the heating station's pump and valve control knowledge base, the heating station's predictive maintenance knowledge base, the heating station's real-time operation data, and cache relevant decisions and control targets generated by large models. The reasoning module is used to train and apply various large models to perform relevant reasoning and decision-making. The tool invocation module is used to select and invoke relevant small models.

Images