Load self-adaptive time sequence control method for coal pulverizing system of thermal power plant

Abstract

The invention discloses a thermal power plant coal pulverizing system load adaptive time sequence control method, and particularly relates to the technical field of thermal power plant intelligent control, and the method comprises the steps: S1, collecting multi-mode operation data of a coal pulverizing system, S2, carrying out the multi-step prediction of numerical time sequence data, generating a key parameter prediction sequence, and carrying out the calculation of the key parameter prediction sequence. The method comprises the steps of S1, predicting data, real-time data and text working condition data, S3, fusing the predicted data, the real-time data and the text working condition data to generate a structured natural language description, S4, inputting a state description into a large language model special for the coal pulverizing system subjected to field fine adjustment, and outputting a time sequence control instruction containing a decision reason, S5, carrying out safety verification on the control instruction, S6, executing control or feedback correction according to a verification result, and S5, outputting a control result. And closed-loop intelligent control is formed. According to the method, through deep fusion of time sequence prediction, natural language processing and physical modeling technologies, cognitive-level adaptive control of the coal pulverizing system from perception to decision is realized, and the control quality and economic safety of the system under complex working conditions are effectively improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent control of thermal power plants, and more particularly, to a load self-adaptive time sequence control method for a pulverizing system of a thermal power plant. BACKGROUND

[0002] The pulverizing system of a thermal power plant is a key auxiliary system of a coal-fired unit, and its running state is directly related to the safety, economy and environmental protection of the unit. The pulverizing system is a typical complex object with multivariable, strong coupling, large inertia and nonlinear characteristics. Its core task is to accurately prepare and deliver coal powder with qualified concentration and fineness according to the load demand of a boiler. Therefore, it has been a goal pursued by those skilled in the art to realize self-adaptive and optimized control of the load of the pulverizing system.

[0003] At present, the existing control technology in this field mainly relies on the following schemes: Traditional PID control: based on classical control theory, the key parameters such as coal supply and primary air volume are adjusted by single loop or cascade regulation through proportional, integral and differential links. However, this method is difficult to cope with the nonlinear and large inertia characteristics of the pulverizing system, and when the coal type changes frequently, the parameter setting is difficult, the control quality is reduced, and true self-adaptation cannot be realized. Advanced process control: such as fuzzy logic control, expert system or model-based predictive control. These methods improve the control performance of the system to some extent. For example, fuzzy control can handle certain uncertainties, but its control rules are heavily dependent on expert experience and are difficult to update dynamically. Traditional predictive control is mostly based on linear or simplified mechanism models. For a complex process such as the pulverizing system, the model mismatch problem is prominent, and especially under conditions such as coal quality fluctuation, the prediction accuracy and robustness are difficult to guarantee. Data-driven intelligent control: in recent years, some researches have attempted to apply machine learning algorithms such as neural networks and support vector machines to modeling and optimization of the pulverizing system. However, most of these methods belong to "shallow" learning, and their perception and decision-making are limited. They can usually only process numerical time series data and cannot understand and utilize text information rich in knowledge such as equipment logs and coal quality parameter specifications. More importantly, the decision-making process of these models lacks explainability, and operating personnel are difficult to understand and trust their control logic. Moreover, the generation of their control instructions lacks deep consideration of physical laws and safety boundaries, and the application risk is high in industrial safety-related scenarios.

[0004] Therefore, in view of the above problems, the present application provides a load self-adaptive time sequence control method for a pulverizing system of a thermal power plant. SUMMARY

[0005] In order to overcome the above-mentioned defects of the prior art, the present application provides a load adaptive time sequence control method for a pulverizing system of a thermal power plant to solve the problems raised in the above background art.

[0006] To achieve the above object, the present application provides the following technical solution: a load adaptive time sequence control method for a pulverizing system of a thermal power plant, specifically comprising the following steps: S1, collecting multi-modal operation data of the pulverizing system of the thermal power plant, wherein the multi-modal operation data includes numerical time sequence data and text type working condition data; S2, inputting the numerical time sequence data into a multi-step prediction model based on an Informer algorithm which has been pre-trained to obtain a prediction sequence of key operation parameters in a future period of time; S3, fusing the prediction sequence, current real-time operation data and text type working condition data to generate a structured natural language state description through a pre-defined state text template; S4, inputting the structured natural language state description into a large language model for the pulverizing system which has been fine-tuned in the field, the large language model being based on a Transformer architecture and outputting a natural language text containing time sequence control instructions and their decision reasons; S5, analyzing the natural language text to extract a specific control instruction sequence and inputting the instruction sequence into a digital twin model based on physical mechanism for safety and feasibility verification; S6, if the verification is passed, the final control instruction sequence is issued to the actual controller of the pulverizing system for execution, and if the verification is not passed, the result of the verification failure is inputted as feedback information to step S3 to guide the generation of a new state description for iterative decision-making.

[0007] Preferably, in step S1, the numerical time sequence data includes coal supply amount, primary air volume, primary air temperature, mill current, mill inlet and outlet pressure difference, separator speed and load instruction. The text type working condition data includes equipment operation log, maintenance record and technical parameter specification of the coal being burned.

[0008] Preferably, the multi-step prediction model based on the Informer algorithm in step S2 uses a ProbSparse self-attention mechanism to process long sequence input, and the model is trained by historical numerical time sequence data to predict mill outlet temperature, required primary air volume and coal fineness at multiple time steps in the future.

[0009] Preferably, the state text template in step S3 is a framework that integrates numerical values, predictions and text information into coherent paragraphs, and its content includes: current time and load, current operating state of equipment, future working condition trend analysis based on prediction sequence, and risk prompts and operating constraints extracted from text type working condition data.

[0010] Preferably, the training process of the large language model dedicated to the pulverizing system in step S4 includes: First stage: field-adapted pre-training of the general pre-trained large language model using the field corpus composed of power plant technical documents, operation procedures and academic papers; Second stage: instruction fine-tuning of the model trained in the first stage using the data set composed of system state description-expert control strategy pairing data, so that it learns to generate control instruction sequences that meet control objectives and safety specifications.

[0011] Preferably, in the natural language text output by the large language model, the control instruction sequence is given in the form of explicit time intervals and operation steps, and each operation step is accompanied by an explanation of the decision-making reason, which is associated with specific working condition information in the input state description.

[0012] Preferably, the digital twin model based on physical mechanism in step S5 is a lightweight calculation model based on mass conservation, energy conservation and momentum conservation equations, which is used to simulate the state evolution of the pulverizing system after the execution of the control instruction sequence, and to judge whether the key parameters exceed the preset safety boundary.

[0013] Preferably, the feedback information that fails the verification in step S6 is formatted into a specific natural language description and integrated into the new state description generated in step S3 to inform the large language model why the previous round of decision instructions was rejected due to physical constraints, thereby guiding it to make corrections.

[0014] The present application also provides an electronic device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the above method when executing the program.

[0015] The technical effects and advantages of the present application are: 1、The present application introduces a state text module and a special large language model for the pulverizing system, converts traditional numerical signal processing into semantic understanding and reasoning based on natural language, fuses multi-dimensional time series data, prediction results and text alarms into structured state descriptions, so that the large language model can read and understand the overall operation context of the system, make strategic planning based on its internalized massive domain knowledge, output sequential control instructions with clear reasons, break through the limitations of traditional control algorithms that can only perform local and instantaneous optimization, achieve deep understanding of the running state context and high-level decision-making, and improve the intelligent processing capability of the system for complex working conditions; 2、The present application adopts Informer algorithm as the time series prediction core, and the ProbSparse self-attention mechanism inside can efficiently capture the long-term dynamic dependency relationship in the large inertia and nonlinear system of the pulverizing system, realize more accurate and longer multi-step prediction of key operating parameters, provide forward-looking information for subsequent decision-making, and combine the sequence generation capability of the large language model, the system can formulate a coherent control strategy spanning multiple time steps, thereby avoiding frequent adjustment and overshoot caused by lag in traditional PID control, realizing more stable load tracking, directly optimizing economic indicators in the strategy, and reducing equipment wear caused by severe fluctuations; 3、The present application combines the generative large language model with the digital twin based on physical mechanism, the large model is responsible for generating innovative optimization strategies, and the digital twin serves as a safety fence that follows strict physical laws to simulate and verify all instructions in advance, ensuring that the final executed instructions have the potential for intelligent optimization and are 100% consistent with physical laws and safety boundaries, in addition, the large model provides decision-making reasons in natural language, converts the AI decision-making process into professional analysis that can be directly understood by human experts, greatly improving the transparency of the system, enabling operation personnel to understand, trust and effectively supervise the AI decision-making, and achieving efficient human-machine collaboration. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 The overall method flowchart of the present application. DETAILED DESCRIPTION

[0017] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0018] The method of this invention requires the construction of a load adaptive timing control system for the pulverizing system of a thermal power plant. This system mainly consists of six modules, each connected in series with the control link through a data interface. The specific modules include: Multimodal data acquisition module: responsible for collecting real-time numerical data and text-based operating condition data during the operation of the pulverizing system; Informer-based multi-step prediction module: analyzes historical numerical data to predict future trends in key operating parameters; State textification module: It integrates numerical data, prediction results and text information, and transforms them into a natural language description that can be understood by the large language model; A dedicated large language model module for pulverizing systems: Based on domain knowledge and control logic, it generates timing control instructions that conform to safety standards. Physical Mechanism Digital Twin Verification Module: Simulates the execution effect of control commands through physical equations to determine whether safety constraints are met; Command execution and closed-loop control module: responsible for issuing commands that have passed verification and feeding back verification failure information to the front end.

[0019] Based on the system established by the above six modules, the method of the present invention specifically includes the following steps: S1. Real-time collection of multimodal operation data through the existing distributed control system, sensor network and production information management system of the pulverizing system of the thermal power plant; Among them, the numerical time-series data includes, but is not limited to, coal feed rate, primary air volume, primary air temperature, coal mill current, coal mill inlet and outlet pressure difference, separator speed and load commands issued by the superior. These data are collected at a preset frequency through the database interface. Text-based operating data includes, but is not limited to, equipment operation logs, maintenance records, and technical parameter specifications for the current type of coal read from the equipment management system.

[0020] S2. The historical numerical time series data collected in step S1 is cleaned and standardized, and then input into a pre-trained multi-step prediction model based on the Informer algorithm. The Informer model internally employs a ProbSparse self-attention mechanism to process long sequence inputs, effectively reducing computational complexity. The model is trained using massive amounts of historical numerical time-series data, and its output is a predicted sequence of key operating parameters for multiple future time steps (e.g., the next 30 minutes). These key operating parameters include at least the coal mill outlet temperature, the required primary air volume of the system, and the fineness of the coal powder obtained through a soft sensor model. The prediction accuracy of this model, especially in long sequence prediction tasks, is superior to traditional ARIMA or LSTM models. S3. Using a predefined state textualization template, the predicted sequence obtained in step S2, the current real-time running data obtained in step S1, and the textual operating condition data are fused to generate a structured natural language state description. The Status Text Template is a framework that integrates numerical, predictive, and textual information into coherent paragraphs. Its specific components include: Current time and host load command; The current operating status of the equipment includes coal feed rate, primary air volume, and coal mill outlet temperature; Analysis of future operating conditions based on predicted sequences, such as the changing trend of required pulverized coal flow and potential risks; Risk warnings and operational constraints extracted from text-based operating data, such as equipment anomaly records in historical logs or the high volatile matter characteristics of the current coal type.

[0021] This template transforms cold, numerical data into contextual reports rich in semantic information that can be understood by large language models.

[0022] S4. Input the structured natural language state description generated in step S3 into a large language model for a flour milling system that has been fine-tuned for the domain. The large language model is based on the Transformer architecture. The training process of a large language model specifically for flour milling systems includes two stages: The first stage is domain adaptation pre-training: using a domain corpus consisting of thermal power plant technical documents, operating procedures, academic papers, etc., the general pre-trained large language model is further pre-trained to enable it to deeply master the professional knowledge of thermal power and pulverizing systems. The second stage, instruction fine-tuning: Using a dataset consisting of system state description-expert control strategy pairing data, the model trained in the first stage is fine-tuned to learn to generate control instruction sequences that conform to control objectives and safety specifications. After receiving the state description, the large language model outputs a natural language text containing timing control instructions and their decision reasons. The sequence of control instructions is given in the form of clear time intervals and operation steps, such as "In the next 3 minutes, increase the primary air temperature by X℃". The decision reasons are associated with the specific working condition information in the input state description, such as "Due to the high volatile matter content of the current coal, increase the air temperature to reduce the risk of deflagration", thus providing interpretability of the decision.

[0023] S5. Analyze the natural language text output by the large language model in step S4, extract the specific control instruction sequence, and then input the instruction sequence into a digital twin model based on physical mechanism for security and feasibility verification. A digital twin model is a lightweight computational model built on the equations of mass conservation, energy conservation, and momentum conservation. It is used to simulate the state evolution of a pulverizing system over a period of time after the execution of a sequence of control commands. By comparing the simulation results with preset safety boundaries (such as the upper limit of the coal mill outlet temperature and the upper limit of the pressure), it is determined whether the key parameters exceed the limits, thereby completing the verification.

[0024] S6. If step S5 passes the verification, the final control command sequence will be sent to the actual controller of the pulverizing system for execution via the industrial communication protocol. If the verification fails, the result of the failure (e.g., "Instruction rejected: Simulation predicts that the outlet temperature will exceed the limit to Y℃") is formatted into a specific natural language description. This feedback information is then re-inputted into step S3 and integrated into the newly generated state description (e.g., adding a prompt at the beginning of the description: "The previous decision was rejected because it may cause the outlet temperature to exceed the limit"). This informs the large language model of the physical constraints that led to the rejection of the previous decision, thereby guiding it to make corrections, generate new and safer control strategies, and form an iterative optimization cognitive loop.

[0025] Example 2 This embodiment also provides an electronic device for implementing the above method, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the method described in steps S1 to S6 of Embodiment 1. Specifically, the processor is configured to: Control the data acquisition interface and execute step S1; Load and run the Informer model, and proceed to step S2; The execution status text template is populated and generated, completing step S3; Load and infer the large language model specifically for the flour milling system, and execute step S4; Drive the digital twin model of physical knowledge to perform simulation verification, and execute step S5; The control command issuance and closed-loop iteration logic are executed in step S6.

[0026] The electronic device can be a dedicated industrial server, integrated in the power plant's central control room, to interact with the power plant's existing control system.

[0027] In summary, this invention creatively combines modern time-series prediction models and natural language processing technology with traditional industrial control by constructing a core technical solution that includes multimodal data acquisition, multi-step prediction based on the Informer algorithm, system state textification, large language model decision-making for pulverizing systems, digital twin verification of physical knowledge, and closed-loop execution. This achieves full-process intelligentization of the pulverizing system from perception and cognition to decision-making, effectively improving the system's adaptive control capability, economy, and safety under complex operating conditions.

[0028] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A load adaptive timing control method for a pulverizing system in a thermal power plant, characterized in that: Specifically, the following steps are included: S1. Collect multimodal operating data of the pulverizing system of thermal power plant, including numerical time series data and text-based operating condition data. S2. Input the numerical time series data into the pre-trained multi-step prediction model based on the Informer algorithm to obtain the prediction sequence of key operating parameters in the future period. S3. The predicted sequence, current real-time operating data, and text-based operating condition data are fused together using a predefined state text template to generate a structured natural language state description. S4. Input the structured natural language state description into a domain-fine-tuned large language model specifically for the flour milling system. The large language model is based on the Transformer architecture and outputs a natural language text containing timing control instructions and their decision reasons. S5. Parse the natural language text, extract the specific control command sequence, and input the command sequence into a digital twin model based on physical mechanisms for security and feasibility verification; S6. If the verification passes, the final control command sequence is sent to the actual controller of the pulverizing system for execution. If the verification fails, the result of the failure is used as feedback information and re-inputted into step S3 to guide the generation of a new state description for iterative decision-making.

2. The adaptive timing control method for load control of a pulverizing system in a thermal power plant according to claim 1, characterized in that: In step S1, the numerical time-series data includes coal feed rate, primary air volume, primary air temperature, coal mill current, coal mill inlet and outlet pressure difference, separator speed, and load command; The text-based operating data includes equipment operation logs, maintenance records, and technical parameter specifications for the type of coal used.

3. The adaptive timing control method for load control of a pulverizing system in a thermal power plant according to claim 1, characterized in that: In step S2, the multi-step prediction model based on the Informer algorithm uses the ProbSparse self-attention mechanism to process long sequence inputs. The model is trained using historical numerical time series data to predict the coal mill outlet temperature, required primary air volume, and coal powder fineness for multiple future time steps.

4. The adaptive timing control method for load of a pulverizing system in a thermal power plant according to claim 1, characterized in that: The status text template in step S3 is a framework that integrates numerical, forecast, and text information into coherent paragraphs. Its content includes: current time and load, current operating status of the equipment, future operating condition trend analysis based on the forecast sequence, and risk warnings and operating constraints extracted from text-based operating condition data.

5. The adaptive timing control method for load of a pulverizing system in a thermal power plant according to claim 1, characterized in that: The training process of the large language model for the powder-making system in step S4 includes: Phase 1: Use a domain corpus consisting of technical documents, operating procedures and academic papers from thermal power plants to perform domain-adaptive pre-training on a general pre-trained large language model. The second stage involves using a dataset consisting of system state descriptions and expert control strategies paired with data to fine-tune the model trained in the first stage, enabling it to learn to generate control command sequences that conform to control objectives and safety specifications.

6. The adaptive timing control method for load of a pulverizing system in a thermal power plant according to claim 5, characterized in that: In the natural language text output by the large language model, the sequence of control instructions is given in the form of clear time intervals and operation steps, and an explanation of the decision-making reason for each operation step is provided. The decision-making reason is related to the specific working condition information in the input state description.

7. The adaptive timing control method for load of a pulverizing system in a thermal power plant according to claim 1, characterized in that: The digital twin model based on physical mechanisms in step S5 is a lightweight computational model built on the equations of mass conservation, energy conservation, and momentum conservation. It is used to simulate the state evolution of the pulverizing system after the execution of the control command sequence and to determine whether the key parameters exceed the preset safety boundaries.

8. The adaptive timing control method for load of a pulverizing system in a thermal power plant according to claim 1, characterized in that: The feedback information that fails verification in step S6 is formatted into a specific natural language description and incorporated into the newly generated state description in step S3. This informs the large language model of the physical constraints that led to the rejection of the previous round of decision instructions, thereby guiding it to make corrections.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, it implements the steps of the method as described in any one of claims 1-8.

Images