Large model agent control method and system for thermal power generation

Abstract

The invention discloses a thermal power generation-oriented large model agent control method and system, and the method comprises the steps: S1, periodically obtaining the real-time operation data of a thermal power generating unit, reading the function description prompt words of all registered equipment agents, and constructing a large language model input prompt word; s2, sending the input prompt word into a large language model, outputting a device agent needing to be called and an action sequence of the device agent until it is judged that any device agent does not need to be called, and outputting a decision result; s3, issuing the action sequence to an equipment function operation side, performing protocol conversion, executing the action sequence by a corresponding hardware equipment or software algorithm, and feeding back an execution result to a next period; a natural language decision output by the model can be automatically converted into a heterogeneous protocol instruction which can be executed on site without manual translation, and online hot plug and unattended closed-loop control under complex working conditions are realized.

Description

Technical Field

[0001] This invention relates to the field of intelligent dispatching technology for thermal power generation, specifically to a large-scale intelligent agent control method and system for thermal power generation. Background Technology

[0002] For many years, my country's rapid economic development has been inseparable from the support of the power industry, with thermal power being the primary source of electricity. The stable operation of thermal power units is crucial to ensuring the safety and reliability of the power system. With the continuous development of artificial intelligence technology, next-generation coal-fired power technology has placed new demands on the intelligent and unmanned operation of thermal power units. Large language models have seen significant improvements in their reasoning and decision-making capabilities in recent years, making it feasible to use them for intelligent decision-making and then invoke intelligent agents to issue commands. However, thermal power scenarios are extremely complex, making equipment execution difficult, and the differences in management platforms across different equipment hinder unified scheduling.

[0003] For example, although a Chinese patent with announcement number CN106406101B uses offline neural network output to control increments, its instructions are fixed values ​​and still need to be manually written to the DCS point by point. It cannot flexibly schedule multiple devices with natural language, nor does it have online protocol conversion. When expanding capacity, it must stop the machine to download. In essence, it is still a semi-closed-loop solution with "decision-making but no automatic execution", which is difficult to meet the closed-loop requirements of large language models for unified, unmanned, and hot-swappable heterogeneous devices. Summary of the Invention

[0004] To address the challenge of inconsistent control of heterogeneous equipment caused by the disconnect between decision-making and execution in large language models in thermal power scenarios, this invention proposes a large model intelligent agent control method for thermal power generation. This method enables the natural language decisions output by the model to be automatically converted into field-executable heterogeneous protocol instructions without manual translation, achieving online hot-swappable and unattended closed-loop control under complex operating conditions.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a large-scale intelligent agent control method for thermal power generation, comprising the following steps: S1 periodically acquires real-time operating data of thermal power units and reads the functional description prompts of all registered intelligent devices to construct a large language model for input prompts; S2, the input prompt is sent to the large language model, and the device agent to be called and its action sequence are output until it is determined that no device agent needs to be called, and the decision result is output. S3, the action sequence is sent to the device function operation side, where it is converted into a protocol and executed by the corresponding hardware device or software algorithm, and the execution result is fed back to the next cycle.

[0006] In this technical solution, on the decision-making side, the output of the large language model is uniformly constrained to structured semantics of "equipment intelligent agent, action, and parameter", shielding the details of thermal power process; on the execution side, a pluggable protocol conversion plugin library is set up to map the above semantics into Modbus, OPC UA, IEC-104 or DCS private messages of each plant in one go, realizing "semantics-signal" plug-and-play translation; both sides are closed-loop through a single standardized interface, iterating periodically, without the need to write communication drivers or customize control logic for each device, realizing the direct driving of heterogeneous thermal power equipment with natural language decision-making.

[0007] Preferably, step S1 includes: S11, periodically poll the real-time database of thermal power units to obtain real-time operating data; S12, Read the function description prompts of all registered device smart agents; S13, the real-time running data and the function description prompt words are concatenated according to a preset template to generate the input prompt words for the large language model.

[0008] Preferably, step S2 includes: S21, The input prompt word is sent to the large language model, and the large language model parses the current operating condition; S22, The large language model outputs the device agent to be invoked and its action sequence based on the parsing results; S23, determine whether the action sequence is empty. If it is empty, terminate the iteration and output the decision result. If it is not empty, retain the current decision result and enter S3. In the next cycle, return the execution result as new running data to S1 to form a closed loop.

[0009] Preferably, the real-time operating data includes at least one of the following: unit load, main steam temperature, main steam pressure, coal quantity, air volume, valve opening, and coal mill start / stop status. The function description prompts include at least the equipment name, function name, function description, input signal type, and dimension.

[0010] Preferably, the action sequence is a structured list including a device agent, an action, and parameter values.

[0011] Preferably, step S3 includes: S31, The action sequence is sent to the device function operation side through a unified interface; S32, the device function operation side calls the corresponding protocol conversion plug-in according to the device intelligent agent identifier in the action sequence to convert the action parameters into Modbus, OPC UA, IEC-104 or DCS private protocol messages; S33 drives the corresponding hardware device or software algorithm to execute and collects the execution results; S34, the execution result is fed back through the same unified interface as part of the running data in the next cycle S1.

[0012] This invention also employs the following technical solution: a large-scale intelligent agent control system for thermal power generation, implementing the aforementioned large-scale intelligent agent control method for thermal power generation, comprising: The data acquisition module periodically provides real-time operating data of the thermal power unit; The registration module saves the functional description prompts of the device's intelligent agent; The decision-making module inputs the real-time running data and prompt words into the large language model and outputs an action sequence; The execution module receives and executes the action sequence through a unified interface and provides feedback on the execution results.

[0013] Preferably, the running module has a built-in protocol conversion plugin that supports Modbus, OPC UA, IEC-104 and DCS proprietary protocols, and the protocol conversion plugin corresponds one-to-one with the device's intelligent agent.

[0014] Preferably, the decision-making module, registration module, and operation module are deployed in parallel through a plant-level security isolation network, and the modules interact only through the unified interface.

[0015] Preferably, the unified interface uses the same message format between the decision-making module and the operation module, and the message format includes the device intelligent agent identifier, action name, parameter value, and timestamp.

[0016] The beneficial effects of this invention are: 1) Complete decoupling of semantics and signals: The large language model only needs to output the semantic list of "device agent-action-parameter value", without having to care about the underlying Modbus, OPC UA or DCS private protocol. On-site, it eliminates the need to write table by table for each device, greatly shortening the debugging cycle. 2) Online hot-swappable expansion: Registration module prompts and protocol conversion plugins can be added or deleted online. New equipment such as coal mills and fans can be connected without stopping the machine to download, significantly improving the unit's availability. 3) Cycle-closed-loop self-correction: The results of each round of execution are rolled back in real time as the input for the next cycle. The large model has an automatic correction strategy, resulting in low main steam temperature deviation and small load fluctuation. 4) Independent upgrade at the module level: The three modules of data acquisition, decision-making and operation are coupled only through a unified interface, and the model version or protocol gateway can be replaced in a gray-scale manner to achieve long-term continuous operation and rapid iteration. Attached Figure Description

[0017] Figure 1 This is a flowchart of a large-scale intelligent agent control method for thermal power generation according to the present invention.

[0018] Figure 2 This is a flowchart of Embodiment 1 of the present invention.

[0019] Figure 3 This is a system block diagram of Embodiment 2 of the present invention. Detailed Implementation

[0020] Example 1 Taking the "grinding" scenario of a coal mill in a coal-fired power plant as an example, this embodiment provides a large-scale intelligent agent control method for thermal power generation implemented on a 660 MW subcritical coal-fired unit. Figure 2 As shown, it includes the following steps.

[0021] This method is deployed on a horizontal isolation platform between the unit's DCS and Safety Zone III. Utilizing the existing DCS real-time database, OPC server, and newly added edge computing nodes, it enables direct scheduling of equipment such as coal mills, primary air fans, and regulating valves using a large language model without altering the original control logic. The implementation process is described in detail below, following the natural sequence of the method flow.

[0022] Step S1: Periodic data collection and prompt word construction.

[0023] In this step, real-time operating data of the thermal power unit is acquired periodically, and the functional description prompts of all registered intelligent devices are read to construct a large language model for input prompts.

[0024] Step S1 includes the following sub-steps.

[0025] Step S11: Periodically poll the real-time database of thermal power units to obtain real-time operating data.

[0026] The system automatically reads parameters such as unit load, main steam temperature, main steam pressure, coal quantity, air volume, valve opening degree, and coal mill start-up and shutdown status from the DCS real-time database in a 2-second cycle to form the operating data for the current cycle.

[0027] Step S12: Read the function description prompts for all registered device smart agents.

[0028] The system accesses the registration database located in Security Zone III to obtain the function description prompts for all registered device agents. These prompts include the device name, function name, function description, input signal type, and dimension. For example, "Device agent: Mill_A; Function name: Reduce coal feed rate; Function description: Reduce coal feed rate in a resolution of 0.1 t / h within the range of 0-60 t / h; Input signal type and dimension: coal_feed(t / h)".

[0029] S13, the real-time running data and the function description prompt words are concatenated according to a preset template to generate the input prompt words for the large language model.

[0030] After the runtime data and prompt words are concatenated according to a preset template, the input prompt words for the large language model are generated. The template adopts a hybrid approach of natural language and structured fields, first giving the current operating condition, then listing the adjustable capabilities, and finally giving the control target, ensuring that the large language model can understand the global information in a single inference.

[0031] Step S2: Large language model reasoning and action sequence output.

[0032] In this step, the input prompts are fed into the large language model, which outputs the device agents to be invoked and their action sequences, until it is determined that no further invocation of any device agents is needed, and then outputs the decision result.

[0033] Step S2 includes the following sub-steps.

[0034] Step S21: The input prompt word is sent to the large language model, which then analyzes the current operating condition.

[0035] Input prompts are fed into a large language model deployed on edge nodes.

[0036] Step S22: The large language model outputs the device agent to be invoked and its action sequence based on the parsing results.

[0037] The model completes semantic parsing in about 600 ms and outputs the device agent to be invoked and its action sequence.

[0038] Step S23: Determine whether the action sequence is empty. If it is empty, terminate the iteration and output the decision result. If it is not empty, retain the current decision result and enter S3. In the next cycle, return the execution result as new running data to S1 to form a closed loop.

[0039] The action sequence is returned as a structured list of "device agent - action - parameter value", for example, [{device:Mill_A, action: decrease coal feed, param:-2}, {device:Mill_B, action: increase coal feed, param:+2}]. If the model determines that no further adjustment is needed, an empty list is returned with a "switch complete" marker. The system performs an empty check on the action sequence: if it is not empty, the system proceeds with the issuance process; if it is empty, the iteration terminates and the final decision result is output.

[0040] Step S3: Unified interface distribution and protocol conversion, hardware or software algorithm execution and result feedback.

[0041] In this step, the action sequence is sent to the device's functional operation side, where it undergoes protocol conversion and is executed by the corresponding hardware device or software algorithm. The execution result is then fed back to the next cycle.

[0042] Step S3 includes the following sub-steps.

[0043] Step S31: The action sequence is sent to the device function operation side through a single unified interface.

[0044] In step S32, the device function operation side calls the corresponding protocol conversion plugin based on the device intelligent agent identifier in the action sequence to convert the action parameters into Modbus, OPC UA, IEC-104 or DCS private protocol messages.

[0045] The runtime side has a built-in protocol conversion plugin library. Each plugin corresponds to a device intelligent agent and supports four protocols: Modbus-TCP, OPCUA, IEC-104, and the manufacturer's proprietary DCS-NET.

[0046] Taking coal mill A as an example, the operating side queries the mapping table based on the equipment intelligent agent identifier "Mill_A" to determine the use of the Modbus-TCP plugin; the plugin converts the coal feed change of -2 t / h into a register value of -20 (resolution 0.1 t / h) and generates a write-multiple hold register message, which is sent to the DCS controller via the reverse isolation device.

[0047] The same process applies to OPC UA nodes or IEC-104 remote control devices, ensuring that the same sequence of actions can be executed without modification in different protocol environments.

[0048] Step S33: Drive the corresponding hardware device or software algorithm to execute, and collect the execution results.

[0049] After receiving the message, the DCS controller adds a -2 t / h bias to the output of the coal feed control loop and gradually shuts down the speed of Mill_A coal feeder through the original PID calculation; at the same time, the corresponding loop of Mill_B receives a +2 t / h bias, and the speed steadily increases.

[0050] After execution, the operation side immediately reads the actual control results, including the actual coal feed, air intake, and completion timestamp, to form the execution result.

[0051] Step S34: The execution result is fed back through the same unified interface as part of the running data in the next cycle S1.

[0052] After the execution results are collected, they are sent back to the data acquisition side through the same unified interface, becoming part of the real-time running data for the next cycle, thus completing the closed loop.

[0053] The entire switching process took about 300 cycles (10 minutes). Mill_A's coal feed and air volume both dropped to zero, while Mill_B rose to the target value. The unit load fluctuation remained within ±0.8%, and the main steam temperature and main steam pressure were within the allowable range of the regulations.

[0054] As can be seen from the above implementation process, the present invention can achieve seamless scheduling between large language models and heterogeneous thermal power equipment simply by using a "unified interface + protocol conversion plug-in". There is no need to develop separate drivers for each device or modify the original control strategy of DCS. At the same time, the registration side prompt words and plug-in mapping table can be added and deleted online, realizing hot-swapping and reuse of device intelligent agents, which significantly improves the feasibility and maintainability of intelligent and unmanned operation of thermal power units.

[0055] Specifically, the beneficial effects of applying this invention to thermal power plants are reflected in three aspects: operation, maintenance, and upgrading.

[0056] First, since the large language model only needs to output the semantic list of "device agent-action-parameter value", it does not need to care about whether the underlying layer is a Modbus register or a DCS-NET private message. Field engineers no longer need to write point tables, mapping blocks or communication drivers for each new device. The debugging cycle is shortened from "weeks" to "hours", and a lot of manpower can be saved during the unit commissioning phase.

[0057] Secondly, both the protocol conversion plugin and the device intelligent agent prompt words support online addition and deletion. When a power plant expands the capacity of the coal mill or upgrades the frequency conversion of the primary air fan, it is only necessary to add a prompt word on the registration side and deploy the corresponding plugin to complete the "hot-swap" without affecting the operation of the original equipment. This avoids the risks of traditional DCS expansion requiring shutdown, download, and restart, and significantly improves the availability of the unit.

[0058] Furthermore, this invention adopts a cycle-closed-loop structure, where each execution result is returned as real-time data for the next round. The large language model can automatically correct its strategy based on this, achieving "self-correction." During multiple switching processes, the main steam temperature fluctuation is controlled within ±2℃, and the load deviation is less than 0.8%, which is better than the historical average of manual operation, verifying its dual value in stability and economy.

[0059] Finally, since all modules are coupled only through a unified interface, the factory can upgrade the model version on the edge side or replace it with a higher-performance protocol gateway without affecting on-site operation. This enables the independent evolution of "core decision-making" and "end-level execution," providing a replicable and scalable technical path for the thermal power industry to connect more types of intelligent devices and move towards true unmanned operation.

[0060] Example 2 This embodiment provides a large-scale intelligent agent control system for thermal power generation, referencing... Figure 3 It includes a data acquisition module, a registration module, a decision-making module, and an operation module.

[0061] The data acquisition module is used to periodically provide real-time operating data of thermal power units; the registration module is used to save functional description prompts for the equipment's intelligent agents; the decision-making module is used to input the real-time operating data and prompts into a large language model and output an action sequence; the operation module is used to receive and execute the action sequence through a unified interface and provide feedback on the execution results.

[0062] The modules interact only through a unified interface, enabling semantic-signal mapping between the large language model and heterogeneous thermal power equipment, completing the decision-making-execution closed loop, supporting online hot-swapping and independent module-level upgrades, and providing a safe, real-time, and scalable hardware platform and software framework for the unmanned and intelligent operation of thermal power units.

[0063] In this embodiment, the system includes three modules: a decision-making main process module based on a large model, a device function registration module, and a device function operation module.

[0064] The main decision-making process module runs periodically, receives real-time operating data of the thermal power unit, and receives registration information of equipment function modules. Based on the current operating data, it makes real-time operating decisions and determines whether it is necessary to call equipment function modules.

[0065] The device function registration module is used to determine all device functions that can be invoked by the large language model. All functions need to be described in a message and form a function description prompt that can be received by the large language model.

[0066] Finally, there is the device function operation module, which is a collection of device functions. It can be either a hardware device or a software algorithm. This module provides a unified calling interface for large language models to call. Internally, the module converts and executes the functions according to the industrial protocols used by the actual hardware or software, and then feeds back the final execution results.

[0067] The detailed steps for system operation are described below.

[0068] The main decision-making process module, as the overall operating module of the system, is responsible for the decision-making and invocation of all device functions. Each of its operating cycles can be broken down into the following steps. Start-up Phase: In this phase, the main decision-making process module will integrate the entire thermal power unit's operating data, including unit load and steam temperature, with information from equipment functional modules to form a large language model input prompt words.

[0069] Decision-making phase: This phase mainly involves planning decision-making tasks based on the unit's operating data and operational requirements input into the large language model. The large language model outputs the actions to be executed and sends the corresponding actions to the functional execution module for execution. This phase repeatedly calls the large language model until the model no longer needs to call any tools and outputs all results.

[0070] Final stage: This stage mainly involves post-processing of the large language model output, outputting the overall unit equipment operation status after processing by the large language model.

[0071] As can be seen from the above steps, the decision-making method in this embodiment is decoupled from the large language model, and the decision-making capability does not depend on a specific large language model. In the actual implementation process, it can be replaced with a suitable large language model according to the specific business scenario requirements.

[0072] The device function registration module is primarily used for registering device functions that can be called by large language models. Device function descriptions must follow a specific format, and their main information must include: device name, device description, function name, function description, and function execution input signal type. The information provided by this module is provided periodically with the main decision-making process module; that is, each time the main decision-making process module runs, it needs to obtain the device function agent registration information from the device function registration module.

[0073] The equipment function operation module primarily manages the operation of functional equipment. This can be either hardware devices, such as automatic valves connected to a DCS system, or software algorithms, such as fault detection algorithms. This module mainly executes the decision results from the main decision-making process module, and its detailed operation includes the following steps.

[0074] Step a, Device Function Activation: When the main decision-making process module generates a decision result, it invokes the corresponding device function for execution based on the information provided by the device function registration module. Since different devices may use different industrial protocols, protocol conversion will be performed at this stage according to the specific device type.

[0075] Step b, Equipment function execution: The specific execution of equipment functions can be provided by specific equipment. For hardware equipment, such as automatic valves, which are already connected to the DCS system of the thermal power plant, valve opening can be controlled by sending valve opening commands through the DCS.

[0076] Step c, Equipment function execution feedback: The results of the equipment function execution need to be fed back to the large model to help you make the next decision.

[0077] As can be seen from the above steps, the industrial protocol conversion process of the underlying hardware device is decoupled from the main decision-making process. The main process is not aware of the underlying protocol of the hardware device. In this way, the operation of the device functional intelligence agent does not depend on the main decision-making process, and the device functional intelligence agent can be easily reused.

[0078] This example illustrates the agent-based decision-making process using a coal mill in a coal-fired power plant as an example of a "grinding" scenario. During operation, the coal mill is responsible for grinding coal into pulverized coal, which is then supplied to the boiler for combustion. Typically, one boiler is equipped with 2-4 coal mills. In daily operation, it is frequently necessary to shut down coal mill A and start coal mill B. The entire "grinding" process involves the shutdown and pre-start of different coal mills. This example simplifies the coal mill operation to two operations: adjusting the pulverized coal feed rate and adjusting the air inlet flow rate. The start-up and shutdown process of the coal mills is controlled by controlling the pulverized coal feed rate and the air inlet flow rate.

[0079] Before the entire decision-making process can begin, coal mill A and coal mill B need to be registered as intelligent agents. The registration information is as follows: "Coal mill A intelligent agent:" Description: Coal mill A provides pulverized coal for boiler combustion. ability: 1. Increase the coal powder feed rate; 2. Reduce the amount of pulverized coal fed; 3. Increase the air intake volume; 4. Reduce the air intake volume;

[0080] After the agent registration is completed, it enters the main decision-making process module. In this module, the large language model receives real-time information, including the current combustion status of the boiler and the coal feed rate and air intake of coal mills A and B. It then outputs the current decision, determining whether to invoke the capabilities of coal mills A and B. An example of the prompt words received by the large language model is shown below: "The current main steam temperature of the boiler is xx degrees, the main steam pressure is xx Pa, the current coal feed rate of coal mill A is xx, the air intake rate is xx, the coal feed rate of coal mill B is xx, and the air intake rate is xx. You possess the following abilities:" #### Coal Mill A Intelligent Agent: Description: Coal mill A provides pulverized coal for boiler combustion. ability: 1. Increase the coal powder feed rate; 2. Reduce the amount of pulverized coal fed; 3. Increase the air intake volume; 4. Reduce the air intake volume; #### Coal Mill B Intelligent Agent: Description: Coal mill B provides pulverized coal for boiler combustion. ability: 1. Increase the coal powder feed rate; 2. Reduce the amount of pulverized coal fed; 3. Increase the air intake volume; 4. Reduce the air intake volume;

[0081] The current objective is to switch coal mill A to coal mill B. Please determine whether to perform a coal mill operation. If so, output the coal mill object and the operation to be performed; otherwise, output "Switching complete".

[0082] The large model will determine the next action based on the provided operating status of the boiler and coal mill, and output the structured operation to be performed by the coal mill, as shown in the following example: Execute the operation: Coal mill A: {Reduce pulverized coal feed rate; reduce air intake rate} Coal mill B: {Increase the coal powder feed rate; increase the air intake rate}.

[0083] Once the large model provides the decision result, the intelligent agent needs to execute the action. The intelligent agent executes the action through a unified interface call. That is, when coal mills A and B register, they attach the action capability to the execution interface. Through the unified interface, the agent sends a signal to reduce the coal feed rate and the air intake rate of coal mill A, and at the same time sends a signal to increase the coal feed rate and the air intake rate of coal mill B.

[0084] After completing the above steps, the entire decision-making process enters the next cycle until the coal feed rate and air intake rate of the coal mill A are finally 0, the large model decision outputs "Switch Complete", and the entire milling decision-making process is completed.

[0085] The beneficial effects of this embodiment will be described below.

[0086] First, autonomous decision-making for the operation of thermal power equipment based on a large language model.

[0087] This embodiment proposes a general framework for autonomous decision-making in equipment operation using a large language model. This framework enables thermal power equipment operation decisions to be made with the help of a large language model, providing a solution for further realizing unmanned equipment operation.

[0088] Second, it allows for easy integration of large models.

[0089] In this embodiment, the autonomous decision-making framework is decoupled from the large language model, does not depend on a specific architecture of the large language model, and reserves an interface for subsequent upgrades of the decision core, facilitating subsequent upgrades of the large language model.

[0090] Third, the intelligent agents of the device can be reused.

[0091] The device agent created in this embodiment can be reused by different decision-making systems or by different large models, which can easily support subsequent upgrades of the decision-making system.

Claims

1. A large-scale intelligent agent control method for thermal power generation, characterized in that, Includes the following steps: S1 periodically acquires real-time operating data of thermal power units and reads the functional description prompts of all registered intelligent devices to construct a large language model for input prompts; S2, the input prompt is sent to the large language model, and the device agent to be called and its action sequence are output until it is determined that no device agent needs to be called, and the decision result is output. S3, the action sequence is sent to the device function operation side, where it is converted into a protocol and executed by the corresponding hardware device or software algorithm, and the execution result is fed back to the next cycle.

2. The large-scale intelligent agent control method for thermal power generation according to claim 1, characterized in that, Step S1 includes: S11, periodically poll the real-time database of thermal power units to obtain real-time operating data; S12, Read the function description prompts of all registered device smart agents; S13, the real-time running data and the function description prompt words are concatenated according to a preset template to generate the input prompt words for the large language model.

3. The large-scale intelligent agent control method for thermal power generation according to claim 1, characterized in that, Step S2 includes: S21, The input prompt word is sent to the large language model, and the large language model parses the current operating condition; S22, The large language model outputs the device agent to be invoked and its action sequence based on the parsing results; S23, determine whether the action sequence is empty. If it is empty, terminate the iteration and output the decision result. If it is not empty, retain the current decision result and enter S3. In the next cycle, return the execution result as new running data to S1 to form a closed loop.

4. A large-scale intelligent agent control method for thermal power generation according to claim 1 or 2, characterized in that, The real-time operating data includes at least one of the following: unit load, main steam temperature, main steam pressure, coal quantity, air volume, valve opening, and coal mill start / stop status. The function description prompts include at least the equipment name, function name, function description, input signal type, and dimension.

5. A large-scale intelligent agent control method for thermal power generation according to claim 1 or 3, characterized in that, The action sequence is a structured list that includes a device agent, an action, and parameter values.

6. The large-scale intelligent agent control method for thermal power generation according to claim 1, characterized in that, Step S3 includes: S31, The action sequence is sent to the device function operation side through a unified interface; S32, the device function operation side calls the corresponding protocol conversion plug-in according to the device intelligent agent identifier in the action sequence to convert the action parameters into Modbus, OPC UA, IEC-104 or DCS private protocol messages; S33 drives the corresponding hardware device or software algorithm to execute and collects the execution results; S34, the execution result is fed back through the same unified interface as part of the running data in the next cycle S1.

7. A large-scale intelligent agent control system for thermal power generation, implementing the large-scale intelligent agent control method for thermal power generation as described in any one of claims 1-6, characterized in that, include: The data acquisition module periodically provides real-time operating data of the thermal power unit; The registration module saves the functional description prompts of the device's intelligent agent; The decision-making module inputs the real-time running data and prompt words into the large language model and outputs an action sequence; The execution module receives and executes the action sequence through a unified interface and provides feedback on the execution results.

8. A large-scale intelligent agent control system for thermal power generation according to claim 7, characterized in that, The operating module has a built-in protocol conversion plugin that supports Modbus, OPC UA, IEC-104 and DCS private protocols. The protocol conversion plugin corresponds one-to-one with the device's intelligent agent.

9. A large-scale intelligent agent control system for thermal power generation according to claim 7 or 8, characterized in that, The decision-making module, registration module, and operation module are deployed in parallel through a plant-level security isolation network, and the modules interact only through the unified interface.

10. A large-scale intelligent agent control system for thermal power generation according to claim 7 or 8, characterized in that, The unified interface uses the same message format between the decision-making module and the operation module. The message format includes the device agent identifier, action name, parameter value, and timestamp.

Images