Coal-fired power plant desulfurization control system and method

Abstract

The application relates to the technical field of flue gas desulfurization technology, in particular to a coal-fired power plant desulfurization control system and method, wherein the system comprises: a collection module that fuses operation data, prediction data and machine vision and the like multi-modal information, a large language model that is enhanced by reasoning and fine-tuning is used for thought chain reasoning to generate a desulfurization control strategy with interpretability, and a closed-loop feedback module is combined to autonomously correct the strategy when the actual effect does not reach the expectation. Therefore, the problems of single perception dimension, lack of interpretability in the strategy generation process and lack of self-adaptive correction capability and the like in the related art desulfurization control are solved.

Description

Technical Field

[0001] This application relates to the field of flue gas desulfurization technology, and in particular to a desulfurization control system and method for coal-fired power plants. Background Technology

[0002] Wet flue gas desulfurization (FGD) systems in coal-fired power plants are core environmental protection facilities for controlling sulfur dioxide (SO2) emissions. With the continuous increase in the proportion of new energy installed capacity, the role of thermal power units is shifting from "main power source" to "basic guarantee + system regulation" power source, making deep peak shaving the norm. Unit loads frequently fluctuate significantly between 30% and 100%, easily leading to uncontrolled SO2 concentrations at the outlet. To ensure environmental compliance, excessive amounts of desulfurizing agents (limestone slurry) are often required, resulting in a surge in operating costs; improper control can lead to the risk of exceeding environmental standards.

[0003] Related technologies employ neural network models to dynamically model the desulfurization process and combine them with genetic algorithms to optimize the slurry supply; alternatively, predictive control strategies are introduced to adjust the slurry supply in advance by predicting the changing trend of inlet SO2 concentration. However, significant shortcomings remain: First, the perception dimension is singular, mainly relying on process parameters (such as pH value, slurry density, inlet SO2 concentration, etc.) collected by the distributed control system, lacking the ability to perceive multimodal information such as slurry foam state, crystallization, demister fouling, and equipment operating visual status. Second, the strategy generation process is a "black box" operation, lacking semantic interpretability and difficulty in integrating expert experience with operating rules, resulting in low human-machine trust. Summary of the Invention

[0004] This application provides a desulfurization control system and method for coal-fired power plants to solve the problems of single perception dimension, lack of perception ability of key visual states, lack of interpretability of strategy generation process, and low efficiency of human-computer interaction in related technologies.

[0005] The first aspect of this application provides a desulfurization control system for a coal-fired power plant, comprising: a data acquisition module for acquiring and fusing operational data, prediction data, manual constraints, and machine vision data to obtain multimodal fusion data; a large language processing module for semantically annotating the multimodal fusion data, inputting the annotated multimodal fusion data into a large language model that has undergone reasoning enhancement and fine-tuning for thought chain reasoning, and generating a desulfurization control strategy containing reasoning paths; a parsing module for parsing the desulfurization control strategy into executable control commands and issuing the control commands to the target actuators of the coal-fired power plant desulfurization system; and a closed-loop feedback module for acquiring the actual operating effect after execution by each target actuator, generating a feedback correction signal and transmitting it to the large language processing module when the actual operating effect does not meet the expected effect, and the large language processing module re-performing thought chain reasoning or correcting the desulfurization control strategy based on the feedback correction signal.

[0006] Optionally, the large language processing module is configured to: load a large language model that has been fine-tuned with inference enhancement, wherein the large language model is configured to perform the following steps: receive fused information text containing a description of the current operating scenario and key parameters, and output inference chain text containing scenario feature recognition, causal relationship analysis, control target priority ranking, adjustment means evaluation and strategy generation; the inference chain text embeds expert knowledge in the field of desulfurization, including control logic rules, equipment response characteristics and safety constraints under different scenarios.

[0007] Optionally, the training data used for inference enhancement fine-tuning consists of inference chain samples from multiple typical operating scenarios; the inference chain samples adopt a triplet structure, including: operating scenario description text, corresponding thought chain inference process, and the finally generated desulfurization control strategy.

[0008] Optionally, the inference enhancement fine-tuning also includes scenario-policy mapping relationship learning, which is used to establish the correspondence between concurrent operating condition characteristics and optimal control methods, enabling the large language model to identify multi-objective conflicts and cooperative relationships in the current operating scenario, and output an inference chain containing priority trade-offs and conflict resolution logic.

[0009] Optionally, the large language processing module includes: a semantic annotation unit, an inference chain generation unit, a policy extraction unit, a policy evaluation and ranking unit, and a policy correction unit. The semantic annotation unit converts multimodal fusion data into natural language descriptions to obtain text annotation information. The inference chain generation unit uses the text annotation information as input, loads a large language model that has been fine-tuned through inference enhancement, and performs inference chain inference to generate inference chain text. The policy extraction unit extracts structured desulfurization control policies from the inference chain text. The policy evaluation and ranking unit performs multi-objective evaluation and ranking of the generated multiple control policies. The policy correction unit receives feedback correction signals from the closed-loop feedback module and generates corrected desulfurization control policies based on the feedback correction signals.

[0010] Optionally, the inference chain text includes a reasoning path encompassing scenario diagnosis, causal analysis, control logic, and strategy output. The inference chain text adopts a hierarchical structure, including: a scenario identification layer, used to describe the key characteristics and abnormal patterns of the current operating condition; a causal analysis layer, used to analyze the causal relationship between operating condition characteristics and desulfurization process parameters; a target priority layer, used to list the control targets and constraints that need to be prioritized in the current scenario; a means evaluation layer, used to evaluate the feasibility, effectiveness, and side effects of various adjustment means; and a strategy output layer, used to output specific control strategies and expected effects.

[0011] Optionally, the execution module includes: a strategy parsing unit, an instruction mapping unit, an instruction verification unit, and an instruction timing coordination unit. The strategy parsing unit parses the desulfurization control strategy into structured instructions; the instruction mapping unit maps the structured instructions to control instructions for the target actuators; the instruction verification unit performs security verification on the control instructions for each target actuator; and the instruction timing coordination unit arranges the execution order and interval of multiple control instructions according to a preset timing logic.

[0012] Optionally, the closed-loop feedback module includes: an effect acquisition unit, a deviation detection unit, and a correction triggering unit. The effect acquisition unit is used to acquire the actual operating effect of each target actuator after executing the control command; the deviation detection unit compares the actual operating effect with the expected effect and calculates the deviation of each control target; the correction triggering unit is used to generate a feedback correction signal and transmit it to the large language processing module when the deviation exceeds a preset threshold.

[0013] Optionally, the actuators include at least one of the following: a slurry supply actuator, a circulating pump actuator, a slurry preparation actuator, an oxidation air actuator, a flushing actuator, and a gypsum discharge and dewatering actuator; the operating data includes real-time operating parameters collected by the desulfurization system of the coal-fired power plant, the prediction data includes prediction information generated based on a time-series prediction model, the manual constraints include manually set constraints and operating rules, and the machine vision data includes visual information acquired by image acquisition devices deployed at key locations in the desulfurization system of the coal-fired power plant.

[0014] The second aspect of this application provides a desulfurization control method for a coal-fired power plant. This method is applied to the desulfurization control system of the coal-fired power plant described in the above embodiment and includes the following steps: collecting and fusing operational data, prediction data, manual constraints, and machine vision data to obtain multimodal fused data; semantically annotating the multimodal fused data; inputting the annotated multimodal fused data into a large language model that has undergone reasoning enhancement and fine-tuning for thought chain reasoning to generate a desulfurization control strategy containing reasoning paths; parsing the desulfurization control strategy into executable control instructions and issuing the control instructions to the target actuators of the coal-fired power plant desulfurization system; collecting the actual operating effects after execution by each target actuator; when the actual operating effect does not meet the expected effect, generating a feedback correction signal and transmitting it to the large language processing module; the large language processing module re-performs thought chain reasoning or corrects the desulfurization control strategy based on the feedback correction signal until the expected effect is achieved.

[0015] Therefore, this application has at least the following beneficial effects: This application's embodiments integrate multimodal information such as operational data, predicted data, and machine vision through a data acquisition module. It utilizes a large language model, enhanced and fine-tuned through reasoning, to perform thought chain reasoning to generate an interpretable desulfurization control strategy. Furthermore, a closed-loop feedback module autonomously corrects the strategy when the actual results do not meet expectations. This solves the problems of single perception dimensions, lack of interpretability in the strategy generation process, and lack of adaptive correction capabilities in related desulfurization control technologies, significantly improving the desulfurization control accuracy, operational safety, and automation level of coal-fired power plants under complex operating conditions.

[0016] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0017] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is a schematic diagram of the architecture of a desulfurization control system for a coal-fired power plant according to an embodiment of this application; Figure 2 This is a schematic diagram of the internal structure of the acquisition module provided according to an embodiment of this application; Figure 3 This is a schematic diagram of the internal structure of the large language processing module provided according to an embodiment of this application; Figure 4 This is a schematic diagram of the training data format for inference enhancement fine-tuning provided according to an embodiment of this application; Figure 5 This is a schematic diagram of the internal structure of the execution module provided according to an embodiment of this application; Figure 6 This is a schematic diagram of the internal structure of the closed-loop feedback module provided according to an embodiment of this application; Figure 7 This is a flowchart of a desulfurization control method for a coal-fired power plant according to an embodiment of this application; Figure 8 This is a schematic flowchart of a closed-loop control method according to an embodiment of this application. Detailed Implementation

[0018] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.

[0019] The following description, with reference to the accompanying drawings, describes a desulfurization control system, method, apparatus, vehicle, and storage medium for coal-fired power plants according to embodiments of this application. Addressing the problems mentioned in the background section, this application provides a desulfurization control system for coal-fired power plants. In this system, a data acquisition module fuses multimodal information such as operational data, predicted data, and machine vision. A large language model, enhanced and fine-tuned through reasoning, is used to perform thought chain reasoning to generate an interpretable desulfurization control strategy. Combined with a closed-loop feedback module 400, the strategy is autonomously corrected when the actual effect does not meet expectations. This solves the problems of single perception dimension, lack of interpretability in the strategy generation process, and lack of adaptive correction capability in related desulfurization control technologies, significantly improving the desulfurization control accuracy, operational safety, and automation level of coal-fired power plants under complex operating conditions.

[0020] Specifically, Figure 1 This is a block diagram of a desulfurization control system for a coal-fired power plant according to an embodiment of this application.

[0021] like Figure 1 As shown, the desulfurization control system 10 of the coal-fired power plant includes: a data acquisition module 100, a large language processing module 200, a parsing module 300, and a closed-loop feedback module 400.

[0022] The system comprises the following modules: a data acquisition module 100, which collects and integrates operational data, prediction data, manual constraints, and machine vision data to obtain multimodal fusion data; a large language processing module 200, which performs semantic annotation on the multimodal fusion data and inputs the annotated multimodal fusion data into a large language model that has been enhanced and fine-tuned by reasoning to perform thought chain reasoning and generate a desulfurization control strategy containing reasoning paths; a parsing module 300, which parses the desulfurization control strategy into executable control commands and issues the control commands to the target actuators of the desulfurization system in the coal-fired power plant; and a closed-loop feedback module 400, which collects the actual operating effects of each target actuator after execution. When the actual operating effect does not meet the expected effect, a feedback correction signal is generated and transmitted to the large language processing module 200. The large language processing module 200 then re-performs thought chain reasoning or corrects the desulfurization control strategy based on the feedback correction signal.

[0023] The actuators include at least one of the following: a slurry supply actuator, a circulating pump actuator, a slurry preparation actuator, an oxidation air actuator, a flushing actuator, and a gypsum discharge and dewatering actuator; the slurry supply actuator includes a slurry pump frequency converter for adjusting the limestone slurry supply volume; The circulating pump actuator includes one or more circulating pump frequency converters, used to adjust the amount of slurry sprayed from the absorption tower; The slurry preparation actuator includes a slurry system feeder, a slurry tank agitator, and a slurry water regulating valve, which are used to control the concentration and amount of limestone slurry prepared. The oxidation air actuator includes an oxidation air fan frequency converter and / or an oxidation air regulating valve, used to regulate the oxidation air volume; The flushing actuators include demister flushing valves and slurry pipeline flushing valves, which are used to control the flushing frequency, flushing duration, and flushing sequence. The gypsum discharge and dewatering actuators include a gypsum discharge pump frequency converter, a gypsum hydrocyclone feed valve, a vacuum belt dewatering machine frequency converter, and a filter cake flushing water valve, which are used to control the discharge volume of gypsum slurry and the operating parameters of the dewatering process.

[0024] Understandably, by enhancing and fine-tuning the large language model through reasoning, the model acquires desulfurization expert-level capabilities in scene recognition, causal analysis, and strategy derivation. Compared to existing intelligent control systems, each control strategy output by the system in this application's embodiment is accompanied by a complete thought chain reasoning process. Operators can understand, review, and intervene in the system's decision-making logic, significantly improving human-machine collaboration efficiency and system reliability. Simultaneously, the reasoning-based decision-making paradigm enables the system to make reasonable decisions through analogical reasoning even when facing boundary conditions not covered by training data, enhancing the system's scenario generalization ability.

[0025] This application embodiment achieves integrated intelligent optimization of the entire desulfurization process through coordinated control of six key links: slurry supply, circulation, slurry preparation, oxidation, rinsing, and paste discharge and dewatering. It can significantly improve the automatic operation rate (target above 98%), and while ensuring that the outlet SO2 concentration is stably up to standard (steady-state fluctuation within ±3mg / Nm³), it effectively reduces limestone consumption and power consumption of circulation pumps and oxidation fans, achieving the dual goals of environmental compliance and cost optimization.

[0026] In one embodiment of this application, the operating data includes real-time operating parameters collected by the desulfurization system of a coal-fired power plant, the prediction data includes prediction information generated based on a time-series prediction model, the manual constraints include manually set constraints and operating rules, and the machine vision data includes visual information acquired by image acquisition devices deployed at key locations in the desulfurization system of a coal-fired power plant.

[0027] The operational data includes real-time operating parameters collected by the DCS (Distributed Control System) of the desulfurization system. These parameters cover all key aspects of the desulfurization process: the pH value of the absorber slurry reflects the chemical environment of the desulfurization reaction; the slurry density reflects the limestone slurry concentration and gypsum crystallization state; the slurry level reflects the liquid holdup in the absorber; the SO2 concentration in the inlet flue gas is the core input for calculating the required slurry supply; the SO2 concentration in the outlet flue gas is the target variable for control; the inlet flue gas flow rate and temperature affect the desulfurization reaction rate; the operating status and frequency of the circulating pump determine the slurry spraying volume; the operating status of the oxidation blower affects the oxidation efficiency of calcium sulfite; the operating status and flow rate of the slurry supply pump determine the amount of limestone slurry added; the differential pressure of the demister reflects the degree of demister blockage; the slurry tank level and slurry density; the gypsum discharge pump current; the gypsum hydrocyclone pressure; the speed of the vacuum belt dewatering machine and the filter cake thickness, etc. These parameters are acquired in real time from the DCS system at different sampling frequencies.

[0028] Predictive data includes forecast information generated based on a time-series prediction model. The desulfurization process exhibits significant time lag (a delay of several minutes to tens of minutes typically exists between adjustments to the slurry supply and the response to the outlet SO2 concentration), making feedback control based solely on real-time data often slow. Therefore, this invention introduces predictive data, including predicted inlet SO2 concentration (based on feedforward information such as boiler load and coal quality analysis), predicted outlet SO2 concentration (based on current operating conditions and the desulfurization reaction model), predicted slurry pH change trends, predicted desulfurization efficiency, predicted slurry demand, predicted oxidation air demand, and predicted gypsum discharge. The predictive data can be generated using algorithms such as Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), Transformer time-series models, or Bayesian variational inference.

[0029] Artificial constraints: These include manually set constraints and operating rules. The operation of a desulfurization system is subject to multiple safety and economic constraints: SO2 emission limits at the outlet are stipulated by national and local environmental regulations (e.g., ultra-low emission standards require an outlet SO2 concentration not exceeding 35 mg / Nm³); the slurry pH value typically needs to be maintained within a safe range of 5.2 to 5.8 (too low a pH value will exacerbate equipment corrosion and reduce desulfurization efficiency, while too high a pH value will easily lead to slurry caking and pipe blockage); the slurry density needs to be controlled within a reasonable range of 1080 to 1150 kg / m³; the circulating pump has a minimum operating frequency limit to prevent bearing overheating; the rate of change in slurry supply is constrained by the conveying capacity of the slurry supply pipeline and the response speed of the chemical reaction within the absorption tower; the target range for slurry concentration (typically 15%-25%); upper and lower limits for oxidation air volume; minimum interval for demister flushing; minimum operating frequency of the gypsum discharge pump; and the maximum processing capacity of the dewatering machine, etc. These artificial constraints are input into the system in the form of rules or parameter constraints, providing safety boundaries for strategy generation.

[0030] Machine vision data includes visual information acquired by image acquisition devices deployed at key locations in the desulfurization system. Traditional desulfurization control systems rely entirely on sensor numerical signals, neglecting the rich state information contained in visual data. This invention deploys industrial cameras or explosion-proof cameras at key locations in the desulfurization system to collect the following visual data in real time: images of the absorber slurry foam layer (foam thickness, foam morphology, foam color, etc., reflect the degree of foaming and defoamer requirements); images of the slurry crystallization state (crystal morphology and particle size distribution reflect the gypsum crystallization quality and dewatering effect); images of scale buildup in the demister (the degree of scale buildup reflects the demister flushing requirements); images of the slurry circulation pump's operating status (visual characteristics of abnormal states such as vibration and leakage); images of the slurry mixing uniformity in the slurry mixing tank; and images of the gypsum filter cake thickness and cracks. These image data are preprocessed (denoising, enhancement, and size normalization) before being input into the acquisition module 100.

[0031] This application's embodiments integrate four types of information: operational data, predictive data, manual constraints, and machine vision, achieving a "comprehensive and multi-dimensional" perception of the desulfurization system's operational status. In particular, the introduction of machine vision data (such as images of the foam layer and crystallization state) compensates for the lack of visual state information that traditional numerical sensors struggle to perceive, enabling the system to see the true state within the desulfurization tower like an experienced operator, thus allowing for more accurate diagnosis and decision-making. Figure 2 The diagram shows the internal architecture of the data acquisition module 100 in the control system of this application. The module contains four parallel processing units: a data acquisition unit responsible for capturing real-time process parameters; a prediction data generation unit responsible for processing time-series prediction information; a manual constraint configuration unit responsible for digitally inputting expert rules and hard constraints; and a machine vision processing unit responsible for extracting image features (such as foam thickness and color). The information processed by all units is integrated into a fused multimodal feature vector, serving as the unified input basis for subsequent large language models to perform thought chain reasoning.

[0032] Preferably, the acquisition module 100 in this embodiment further includes a multi-source data spatiotemporal alignment unit. Since the sampling frequencies of different data sources vary greatly (e.g., DCS data is usually sampled at the second level, while machine vision data is sampled at the second level or in stages), and the time bases of each data source may differ, the multi-source data spatiotemporal alignment unit is responsible for aligning these heterogeneous data with timestamps and spatial coordinates to form unified multimodal fusion information.

[0033] In one embodiment of this application, the large language processing module 200 includes: a semantic annotation unit, an inference chain generation unit, a policy extraction unit, a policy evaluation and ranking unit, and a policy correction unit. The semantic annotation unit converts multimodal fusion data into natural language descriptions to obtain text annotation information. The inference chain generation unit uses the text annotation information as input, loads a large language model that has been fine-tuned through inference enhancement, and performs inference chain inference to generate inference chain text. The policy extraction unit extracts structured desulfurization control policies from the inference chain text. The policy evaluation and ranking unit performs multi-objective evaluation and ranking of the generated multiple control policies. The policy correction unit receives feedback correction signals from the closed-loop feedback module 400 and generates corrected desulfurization control policies based on the feedback correction signals.

[0034] like Figure 3 The diagram illustrates the core workflow of the Large Language Processing Module 200, constructing a complete inference chain from semantic understanding to policy generation and closed-loop optimization. The module first performs semantic parsing on the multimodal fusion data through a semantic annotation unit. Then, it uses an inference chain generation unit to load a finely tuned large language model, generating text containing detailed logical deduction processes. Next, a policy extraction unit parses structured control policies from the text, and a policy evaluation and ranking unit scores and optimizes multiple candidate solutions. Notably, this process introduces a feedback-based dynamic optimization mechanism: the policy correction unit receives closed-loop feedback signals and continuously adjusts the inference process and model parameters through the optimization unit, thereby achieving iterative improvements in policy generation accuracy and adaptability.

[0035] Specifically, the semantic annotation unit converts the multimodal fusion information output by the acquisition module 100 into natural language descriptions, forming text annotation information consistent with the training data format. For example, numerical information such as "pH=5.3, inlet SO2=1800mg / Nm³, outlet SO2=32mg / Nm³, foam layer thickness=15cm" is converted into natural language descriptions such as "The current pH value of the absorber slurry is 5.3, which is at the lower limit of the safe range; the inlet SO2 concentration is 1800 mg / m³, which has increased by 12% compared to the previous moment; the outlet SO2 concentration is 32 mg / m³, which is close to the emission limit of 35 mg / m³; the foam layer thickness in the absorber is 15 cm, which has reached the warning threshold."

[0036] The reasoning chain generation unit loads a large language model that has been fine-tuned with reasoning enhancements. Using semantically labeled information as input, it generates a complete reasoning chain text containing scene diagnosis, causal analysis, control logic, and policy output through reasoning. The reasoning chain adopts a hierarchical structure: Scene Recognition Layer → Causal Analysis Layer → Target Priority Layer → Means Evaluation Layer → Policy Output Layer.

[0037] Strategy Extraction Unit: Extracts structured desulfurization control strategies from the "strategy output layer" of the inference chain text and converts them into a format that can be parsed by the Action module.

[0038] Strategy Evaluation and Ranking Unit: When multiple candidate strategies are generated, multi-objective evaluation and ranking are performed. Evaluation indicators include: SO2 emission compliance confidence level, operating cost change, strategy execution risk level, and equipment safety margin. Candidate strategies are ranked by weighted comprehensive scoring of these indicators, and the optimal strategy is passed to the Action module.

[0039] Strategy correction unit: Receives feedback correction signals from closed-loop feedback module 400, integrates deviation cause analysis into the new round of inference chain, and generates a correction strategy that includes analysis of "why the previous strategy did not meet expectations".

[0040] In one embodiment of this application, the large language processing module 200 is configured to: load a large language model that has been fine-tuned by inference enhancement, wherein the large language model is configured to perform the following steps: receive fused information text containing a description of the current operating scenario and key parameters, and output inference chain text containing scenario feature recognition, causal relationship analysis, control target priority ranking, adjustment means evaluation and strategy generation; the inference chain text embeds expert knowledge in the field of desulfurization, including control logic rules, equipment response characteristics and safety constraints under different scenarios.

[0041] It is understandable that the innovation of the large language processing module 200 in this application embodiment lies in the use of a large language model that has undergone reasoning enhancement and fine-tuning. By introducing CoT (Chain-of-Thought) reasoning during the fine-tuning process, the large language model learns to perform step-by-step reasoning like an experienced desulfurization expert, thereby generating an interpretable control strategy with a complete logical chain. In this application embodiment, the reasoning enhancement and fine-tuning of the large language model adopts the chain-of-thought fine-tuning method, and the format of the training data samples is as follows: Input: Integrated information text, including a description of the current working condition and key parameters; Output: Inference chain text, including scene feature recognition, causal relationship analysis, control target priority ranking, regulation method evaluation, and policy generation; The inference chain text embeds expert knowledge in the field of desulfurization, including control logic rules, equipment response characteristics and safety constraints under different scenarios.

[0042] In one embodiment of this application, the training data used for inference enhancement fine-tuning consists of inference chain samples from multiple typical operating scenarios. The inference chain samples employ a triplet structure, including: a description of the operating scenario, the corresponding thought chain inference process, and the finally generated desulfurization control strategy. The corresponding thought chain inference process serves as a bridge connecting the scenario and the strategy, explicitly recording the expert's complete thought process from observation to decision-making.

[0043] The training data for inference enhancement fine-tuning in this application includes inference chain samples from the following typical scenarios: (1) Situation of sudden increase in SO2 concentration at the inlet: The reasoning chain includes “identifying sudden increase in sulfur content at the inlet → analyzing the cause (increased load / change in coal quality) → assessing the current pH value and SO2 safety margin at the outlet → determining whether to prioritize increasing the slurry supply or circulation volume → considering the slurry supply response lag characteristics → generating slurry pump frequency increase command and rate limit”; (2) Scenario of rapid load increase of unit: The reasoning chain includes "monitoring the load change rate → predicting the change of inlet sulfur → predicting the desulfurization load demand → increasing the output of the circulating pump in advance → preparing the pulping system simultaneously to increase the slurry concentration → executing the feedforward control strategy"; (3) Foaming scenario of absorber slurry: The reasoning chain includes “identifying abnormal foam layer area / thickness → analyzing the cause of foaming (accumulation of organic matter / fine particulate matter / limestone quality) → determining whether it affects liquid level measurement → prioritizing adjustment of demister flushing → assessing whether defoamer needs to be added → avoiding excessive increase in circulation volume to aggravate foaming”; (4) Scenario of low pH value of slurry: The reasoning chain includes "judging the rate of pH drop → analyzing the cause (insufficient slurry supply / increased inlet sulfur / insufficient circulation volume) → assessing whether it is close to the safety lower limit → determining the increase in slurry supply → considering the matching of oxidation air volume → monitoring gypsum discharge to prevent abnormal density"; (5) High slurry density scenario: The reasoning chain includes “identifying the density increase trend → analyzing the cause (excessive slurry supply / insufficient gypsum discharge) → prioritizing the increase of gypsum discharge → assessing the processing capacity of the dewatering system → adjusting the slurry concentration to reduce the water content of the slurry supply → preventing the risk of pipeline blockage”.

[0044] In another embodiment, the inference enhancement fine-tuning also includes scenario-policy mapping relationship learning, used to establish the correspondence between concurrent operating condition characteristics and optimal control methods, enabling the large language model to identify multi-objective conflicts and cooperative relationships in the current operating scenario, and output an inference chain containing priority trade-offs and conflict resolution logic. In actual execution, the large language model of this application embodiment can identify feature combinations of multiple concurrent scenarios (such as "sudden increase in inlet sulfur + low pH value + excessively thick foam layer" occurring simultaneously); determine the priority ranking of control objectives in different scenarios (such as environmental compliance taking precedence over economic operation, and safety boundary taking precedence over efficiency optimization); identify conflicts and cooperative relationships between scenarios (such as increasing the circulation volume can improve desulfurization efficiency but may exacerbate foam problems); and output a comprehensive inference chain containing priority trade-offs and conflict resolution logic.

[0045] The following example uses the scenario of "sudden increase in SO2 concentration at the inlet". Figure 4 The image shows a complete training data format diagram: Input (description of working conditions): Current operating conditions: The unit load increased from 280MW to 420MW within 60 seconds (rate of change 2.33MW / s), and the inlet SO2 concentration jumped from 1050mg / Nm³ to 1850mg / Nm³. The current outlet SO2 concentration is 28mg / Nm³ (limit 35mg / Nm³), with a safety margin of 7mg / Nm³. The absorber slurry pH is 5.35 (safe range 5.20-5.80), which is at a low to medium level. The slurry density is 1090kg / m³ (normal range 1080-1150). #1 circulation pump frequency is 45Hz, #2 circulation pump frequency is 48Hz. The slurry supply pump frequency is 38Hz, and the slurry concentration is 17%. The foam layer area is 12% (normal). The demister differential pressure is 150Pa (normal).

[0046] Output (inference chain + strategy).

[0047] [Output Reasoning Chain] such as Figure 4 As shown, the chain text in this embodiment includes a reasoning path of scenario diagnosis, causal analysis, control logic, and strategy output. The reasoning chain text adopts a hierarchical structure, including: a scenario identification layer, used to describe the key characteristics and abnormal patterns of the current operating condition; a causal analysis layer, used to analyze the causal relationship between the operating condition characteristics and the desulfurization process parameters; a target priority layer, used to list the control targets and constraints that need to be prioritized in the current scenario; a means evaluation layer, used to evaluate the feasibility, effectiveness, and side effects of various adjustment means; and a strategy output layer, used to output specific control strategies and expected effects.

[0048] Step 1: Scene Feature Recognition Key anomaly: The inlet SO2 concentration increased by 76% (1050→1850) within 60 seconds, at a rate of approximately 13.3 mg / Nm³ / s, which is classified as a "rapid increase".

[0049] Load change characteristics: The unit load increased rapidly and synchronously, indicating that the sudden increase in inlet sulfur was due to the increase in boiler combustion intensity, rather than a sudden change in coal quality (sudden changes in coal quality are usually not accompanied by synchronous load changes).

[0050] Safety margin assessment: The current SO2 output is 28 mg / Nm³, only 7 mg / Nm³ away from the limit of 35 mg / Nm³. Based on experience, for every 100 mg / Nm³ increase in inlet concentration, the output increases by approximately 0.8-1.2 mg / Nm³ (depending on circulation volume and pH). It is estimated that an inlet concentration of 1850 mg / Nm³ will cause the output to increase by approximately 6-9 mg / Nm³. Without intervention, the output will rise to 34-37 mg / Nm³, posing a risk of exceeding the limit.

[0051] Step 2: Causal Relationship Analysis A sudden increase in inlet sulfur concentration leads to an increase in SO2 absorption load within the absorption tower, a faster rate of decrease in slurry pH, a decrease in desulfurization efficiency, and an increase in outlet SO2.

[0052] Key time constants: slurry supply adjustment (approximately 5-10 seconds from command to flow rate change) → pH response (30-90 seconds) → outlet response (60-120 seconds). Significant lag exists, requiring earlier intervention.

[0053] Circulation pump adjustment: Increasing the circulation volume can immediately improve the gas-liquid contact efficiency (fast response), but it consumes a lot of energy and has limited effect on pH improvement.

[0054] Slurry supply adjustment: Increasing the slurry supply can fundamentally replenish alkalinity (slow response but fundamental solution), but excessive slurry supply will lead to pH overshoot and increase slurry density.

[0055] Step 3: Prioritize Control Targets 1. Highest priority: Prevent SO2 emissions from exceeding limits (environmental red line).

[0056] 2. Second priority: Maintain pH value not lower than 5.20 (to ensure equipment corrosion prevention and desulfurization efficiency).

[0057] 3. Third priority: Control costs and avoid oversupply of slurry and energy waste.

[0058] Step 4: Evaluation of Adjustment Measures Increase slurry supply (slurry pump frequency 38→48Hz): The pH value is expected to start rising after 60 seconds, and the outlet SO2 value is expected to decrease after 90 seconds. Risk: The risk of pH overshoot is low, but density monitoring is required.

[0059] Increasing the circulation rate (#2 circulation pump 48→50Hz): Immediate effect, but marginal desulfurization efficiency decreases (limited improvement from 48→50Hz). Risk: Increased energy consumption.

[0060] Combined strategy: plasma supply as the main method (addressing the root cause) + circulation as a supplement (rapid response) + oxidative air matching (preventing sulfite accumulation).

[0061] Step 5: Strategy Generation The frequency of the grout pump was increased from 38Hz to 48Hz, with an increase rate of 1Hz / s (to prevent pressure shock in the grout supply pipeline).

[0062] The circulation pump frequency was increased from 48Hz to 50Hz (full frequency).

[0063] The frequency of the oxidation fan was increased by 3Hz (from 42Hz to 45Hz) to match the increased oxidation demand.

[0064] Expected results: SO2 at the outlet will drop to 22-25 mg / Nm³ within 90 seconds, and the pH value will drop to a minimum of 5.28 before recovering.

[0065] Risk warning: If the inlet sulfur continues to rise above 2200 mg / Nm³, a second-stage slurry supply boost will be required.

[0066] Output Strategy {Slurry pump:{target:48Hz,ramp_rate:1Hz / s},Circulation pump_2:{target:50Hz},Oxidation blower:{target:45Hz}} By fine-tuning a large number of inference chain samples covering various working conditions, the large language model learns a general inference pattern from scene features to control decisions, rather than a simple statistical mapping.

[0067] Furthermore, the inference enhancement fine-tuning training data in this application covers the following core scenarios and their variations: Scenario A: Sudden increase in inlet sulfur, ΔSO2_in / Δt>10mg / Nm³ / s, core logic of the inference chain: assess safety margin → prioritize slurry supply (fundamental solution) + circulation assistance (rapid response) → match oxidation wind; Scenario A-1: ​​Sudden increase in inlet sulfur + low pH, pH < 5.25 and sudden increase in inlet sulfur, the core logic of the inference chain is: identify dual pressure → the slurry supply needs to be increased (e.g., +12Hz instead of +8Hz) → the circulation volume needs to be increased in advance → monitor pH to prevent it from falling below the lower limit; Scenario A-2: Sudden increase in inlet sulfur + excessive foam thickness, foam > 30% and sudden increase in inlet sulfur, core logic of the inference chain: identify collaborative constraints → avoid excessive increase in circulation volume (exacerbating foam) → prioritize slurry supply + demister flushing → prepare defoamer; Scenario B: Load rapidly increases, load change rate > 1.5MW / min. Core logic of the inference chain: predict the inlet sulfur to rise → execute feedforward control → increase circulation pump output in advance → synchronously adjust pulping concentration. Scenario C: Slurry foaming, foam area > 25% or thickness > 15cm, core logic of the reasoning chain: identify the degree of foaming → analyze possible causes (organic matter / limestone / fine particles) → prioritize rinsing the demister → assess the need for defoamer → avoid a surge in circulation volume; Scenario D: pH value is low, pH < 5.25 and the rate of decrease is > 0.02 / min. The core logic of the reasoning chain is: determine the cause of the decrease (insufficient slurry supply / inlet sulfur rise / insufficient circulation) → determine the slurry supply increment → oxidation air matching → slurry discharge adjustment; Scenario E: The pulp density is too high, with a density > 1140 kg / m³. The core logic of the reasoning chain is: determine the cause of the density increase (excessive pulp supply / insufficient paste discharge) → prioritize increasing paste discharge → assess the dewatering machine capacity → adjust the pulp concentration. Scenario F: Multiple scenarios occurring concurrently, with a sudden increase in inlet sulfur, low pH, and excessively thick foam. The core logic of the reasoning chain is: priority ranking → environmental compliance first → safety boundary second → cost last → conflict resolution (supply of slurry is the main focus, circulation is limited).

[0068] It should be noted that the embodiments of this application adopt the following fine-tuning technical solutions, which can also be limited according to the actual situation, without specific restrictions, as follows: Base model selection: Select open-source large language models that support long context (such as DeepSeek-V3, Qwen2-72B, etc.), with a parameter count between 7B and 72B, balancing inference capabilities and local deployment feasibility.

[0069] Fine-tuning method: Employ LoRA (Low-Rank Adaptation) or QLoRA parameter fine-tuning methods to train only a subset of parameters, thereby reducing computational resource requirements.

[0070] Training data scale: Construct no less than 5,000 high-quality inference chain samples, covering the above-mentioned scenarios and their combined variations, with each sample containing detailed inference chain annotations.

[0071] Loss function design: A weighted combination of inference chain loss and policy loss is used to ensure that the model not only outputs the correct policy, but also the correct inference process.

[0072] Reasoning chain length control: The length of the reasoning chain is constrained to between 150 and 500 tokens by using prompt words to ensure the sufficiency of reasoning while controlling reasoning delay.

[0073] In one embodiment of this application, the parsing module 300 includes: a strategy parsing unit, an instruction mapping unit, an instruction verification unit, and an instruction timing coordination unit. The strategy parsing unit is used to parse the desulfurization control strategy into structured instructions; the instruction mapping unit is used to map the structured instructions into control instructions for the target actuators; the instruction verification unit is used to perform security verification on the control instructions for each target actuator; and the instruction timing coordination unit is used to arrange the execution order and interval of multiple control instructions according to a preset timing logic.

[0074] Specifically, such as Figure 5 As shown, the parsing module 300 is the execution layer of the entire system, connected to the large language processing module 200. It parses the control strategy generated by the large language processing module 200 into executable control instructions, and then distributes these instructions to the slurry supply actuator, circulating pump actuator, slurry preparation actuator, oxidation air actuator, flushing actuator, and gypsum discharge and dewatering actuator. The parsing module 300 includes four core units: (1) Strategy parsing unit: Parses the control strategy in natural language form output by the large language processing module 200 into structured instructions. The structured instructions include fields such as action type, target device, target value, execution rate, and execution sequence. For example, the parsed instruction set may include one or more of the following: slurry supply instruction, circulating pump instruction, slurry preparation instruction, oxidation air instruction, flushing instruction, gypsum discharge and dewatering instruction.

[0075] (2) Instruction mapping unit: maps structured instructions into specific control signals for six types of actuators: Control commands for the slurry supply actuator: mapped to the slurry pump frequency converter (4-20mA signal or Modbus command), used to adjust the limestone slurry supply rate. The supply rate can be expressed as frequency (Hz) or flow rate setpoint (m³ / h).

[0076] Control commands for circulating pump actuators: Mapped to one or more circulating pump frequency converters to adjust the slurry spray volume from the absorber tower. These commands may include pump start / stop, frequency setting, and switching sequence.

[0077] The control commands for the slurry preparation actuator are mapped to the slurry system feeder (controlling the limestone feed rate), the slurry tank agitator (start / stop), and the slurry water regulating valve (opening degree) to control the concentration (typically 15%-25%) and quantity of the limestone slurry. These commands may include parameters such as feed rate, slurry water flow rate, and mixing time.

[0078] Oxidation air actuator control commands: mapped to the oxidation fan frequency converter and / or oxidation air regulating valve, used to adjust the oxidation air volume. These commands may include parameters such as fan frequency, damper opening, and oxidation air pressure.

[0079] Flushing actuator control commands: mapped to the demister flushing valve and slurry pipeline flushing valve, used to control flushing frequency, flushing duration, and flushing sequence (e.g., stratified flushing, pulse flushing). May include flushing intervals, single flushing time, flushing valve sequence numbers, etc.

[0080] Control commands for the gypsum discharge and dewatering actuators are mapped to the gypsum discharge pump frequency converter (adjusting discharge flow rate), the gypsum hydrocyclone feed valve (opening degree), the vacuum belt dewatering machine frequency converter (adjusting belt speed), and the filter cake washing water valve (opening degree). These commands control the discharge volume of the gypsum slurry and the operating parameters of the dewatering process. They may include discharge pump frequency, hydrocyclone feed pressure, dewatering machine belt speed, target filter cake thickness, and target filter cake moisture content.

[0081] (3) Command verification unit: performs security verification before the control command is issued, including: Parameter range verification: Check whether the target value of each instruction exceeds the safe operating range set in the manual limit (such as the slurry pump frequency not exceeding the upper limit, the slurry concentration not exceeding the target range, the oxidation air volume not exceeding the fan capacity, etc.).

[0082] Change rate verification: Check whether the change rate of the command exceeds the allowable range of the equipment (e.g., sudden changes in the slurry supply may cause drastic fluctuations in pH value, so the change slope needs to be limited).

[0083] Logical conflict check: Check whether there are timing or logical conflicts between multiple instructions (such as significantly increasing the slurry supply when the circulating pump is completely stopped, which may cause the slurry to be too thick in some areas; starting the dewatering machine when the gypsum discharge pump is not started may cause it to run dry).

[0084] Interlock condition verification: Check whether the command meets the system interlock conditions (e.g., prohibit large-flow gypsum discharge when the absorber level is too low).

[0085] Only instructions that pass verification can be issued to the execution mechanism. If verification fails, the instruction verification unit will return conflict information to the large language processing module 200 and request the strategy to be regenerated.

[0086] (4) Command timing coordination unit: used to arrange the execution order and interval of multiple control commands according to preset timing logic. For example, start the circulating pump first and then increase the slurry supply, or adjust the slurry concentration first and then adjust the slurry supply, so as to avoid system disturbance.

[0087] In one embodiment of this application, the closed-loop feedback module 400 includes: an effect acquisition unit, a deviation detection unit, and a correction triggering unit. The effect acquisition unit is used to acquire the actual operating effect of each target actuator after executing control commands; the deviation detection unit compares the actual operating effect with the expected effect and calculates the deviation of each control target; the correction triggering unit is used to generate a feedback correction signal and transmit it to the large language processing module 200 when the deviation exceeds a preset threshold.

[0088] Specifically, such as Figure 6 As shown, the closed-loop feedback module 400 is the core component for implementing adaptive control in this invention, and it is connected to the acquisition module 100 and the large language processing module 200. This module includes three core units: (1) Effect Acquisition Unit: Used to collect the actual operating effect after the acquisition module 100 is executed. The collected data includes, but is not limited to: the actual change curve of SO2 concentration at the outlet (time series, sampling frequency not less than 1Hz), the actual response curve of slurry pH value, the actual consumption of slurry supply (cumulative flow rate and instantaneous flow rate), the actual power consumption of equipment such as circulating pump, oxidation fan, slurry supply pump, gypsum discharge pump, and dewatering machine, the actual change of slurry concentration, oxidation air volume (measured by flow meter), the differential pressure change of demister (reflecting the flushing effect), the gypsum discharge and the moisture content of gypsum after dewatering, and the changes in machine vision data (such as the change of foam layer area, the change of filter cake state, etc.).

[0089] (2) Deviation Detection Unit: This unit compares the actual operating effect with the expected effect generated by the large language processing module 200 when generating the strategy, and calculates the deviation of each control target. The expected effect can be extracted from the "expected effect" description output by the large language model when generating the strategy, or calculated by the simplified prediction model built into the system. Key deviation indicators include: outlet SO2 concentration overshoot: the difference between the actual minimum value and the expected minimum value (negative value indicates overshoot), outlet SO2 concentration undershoot: the difference between the actual maximum value and the expected maximum value (positive value indicates undershoot), response time deviation: the difference between the actual time to reach the target value and the expected time, steady-state error: the deviation between the actual value and the target value after stabilization, slurry supply overshoot coefficient: the ratio of the actual slurry supply increment to the expected slurry supply increment, and a threshold is preset in the deviation detection unit. When the deviation exceeds the threshold, correction is triggered. Example thresholds: Absolute deviation of SO2 concentration at outlet ≥ 3 mg / Nm³; absolute deviation of pH value ≥ 0.1; deviation of response time relative to expected time ≥ 20%; over-adjustment coefficient of slurry supply ≥ 1.15 or ≤ 0.85.

[0090] (3) Correction triggering unit: When the deviation detection unit determines that the deviation exceeds the preset threshold, it generates a feedback correction signal and transmits it to the strategy correction unit of the large language processing module 200. The feedback correction signal adopts a structured format and includes the following fields: deviation type (under-adjustment, over-adjustment, response delay, steady-state deviation, combined deviation), deviation amplitude (value and ratio relative to the threshold), occurrence time (time window from the issuance of the instruction to the occurrence of the deviation), associated control instruction (pointing to the specific actuator and target value), and suggested correction direction (optional, initially determined by the deviation detection unit based on the deviation characteristics, such as "need to continue to increase slurry supply" or "need to reduce the frequency of the circulating pump").

[0091] Based on the above embodiments, the closed-loop feedback module 400 of this application transmits the actual operating effect back to the large language processing module 200, enabling the system to continuously learn and self-optimize, and constantly improve control accuracy and adaptability. This closed-loop mechanism allows the system to adapt to long-term drift factors such as unit aging, coal quality changes, and ambient temperature changes, maintaining long-term stable operation. The closed-loop feedback module 400 achieves real-time monitoring and deviation detection of the execution effect. When the control effect does not meet expectations or over-adjustment occurs, the system can automatically trigger the large language processing module 200 to correct the strategy. This "execution-feedback-correction" mechanism enables the system to cope with uncertainties such as model prediction errors, nonlinear response of the actuator, and sudden changes in operating conditions, significantly improving the robustness and adaptability of the control system and avoiding the shortcomings of traditional open-loop intelligent control, which is characterized by "one-time decision, no correction."

[0092] The desulfurization control system for coal-fired power plants proposed in this application integrates multimodal information such as operational data, predicted data, and machine vision through a data acquisition module 100. It then uses a large language model, enhanced and fine-tuned through reasoning, to perform thought chain reasoning to generate an interpretable desulfurization control strategy. Furthermore, a closed-loop feedback module 400 autonomously corrects the strategy when the actual effect does not meet expectations. This solves the problems of single perception dimension, lack of interpretability in the strategy generation process, and lack of adaptive correction capability in related desulfurization control technologies, significantly improving the desulfurization control accuracy, operational safety, and automation level of coal-fired power plants under complex operating conditions.

[0093] Next, referring to the accompanying drawings, a desulfurization control method for coal-fired power plants according to an embodiment of this application is described. This method is applied to the desulfurization control system of the coal-fired power plant described in the above embodiment.

[0094] Figure 7 This is a schematic flowchart of a desulfurization control method for a coal-fired power plant provided in an embodiment of this application.

[0095] like Figure 7 As shown, the desulfurization control method for this coal-fired power plant includes the following steps: In step S101, running data, prediction data, manual constraints, and machine vision data are collected and fused to obtain multimodal fusion data.

[0096] In step S102, the multimodal fusion data is semantically labeled, and the labeled multimodal fusion data is input into a large language model that has been fine-tuned by reasoning enhancement to perform thought chain reasoning and generate a desulfurization control strategy containing reasoning paths.

[0097] In step S103, the desulfurization control strategy is parsed into executable control commands, and the control commands are sent to the target actuator of the desulfurization system of the coal-fired power plant.

[0098] In step S104, the actual operating effect of each target actuator after execution is collected. If the actual operating effect does not meet the expected effect, a feedback correction signal is generated and transmitted to the large language processing module. The large language processing module re-performs the thought chain reasoning or corrects the desulfurization control strategy based on the feedback correction signal until the expected effect is achieved.

[0099] Specifically, such as Figure 8 The diagram clearly depicts the complete logical chain from data processing to closed-loop control, as follows: Step S1: Collect and fuse operational data, prediction data, manual constraints, and machine vision data.

[0100] Step S2: Semantically label the fused information, perform thought chain reasoning through a large language model that has been enhanced and fine-tuned by reasoning, generate a complete reasoning chain that includes scenario diagnosis, causal analysis and control logic, and extract the desulfurization control strategy (first-round strategy) from the reasoning chain.

[0101] Step S3: Parse the control strategy into executable control commands, and send the control commands to the slurry supply actuator, circulation pump actuator, slurry preparation actuator, oxidation air actuator, flushing actuator, and gypsum discharge and dewatering actuator respectively.

[0102] Step S4: Collect the actual operating effect after the execution mechanism is executed, compare the actual operating effect with the expected effect, and generate a feedback correction signal when the execution effect is not as expected or over-adjustment occurs.

[0103] Step S5: Based on the feedback correction signal and the current operating condition fusion information, regenerate the inference chain including deviation cause analysis and correct the strategy. Repeat steps S3 to S4 until the control effect meets the preset requirements.

[0104] In the above steps, steps S4 and S5 constitute a feedback correction loop, which can be iterated multiple times until the deviation converges. To prevent infinite loops, the system sets a maximum number of corrections (e.g., 3 times) and a convergence criterion (the deviation drops to less than half of the threshold).

[0105] It should be noted that the explanation of the aforementioned embodiment of the desulfurization control method for coal-fired power plants can refer to the workflow of the desulfurization control system of the coal-fired power plant in the above embodiment. To avoid redundancy, it will not be repeated here.

[0106] The desulfurization control system for coal-fired power plants of this application will be described in detail below with reference to specific embodiments. The control system of this application embodiment can be deployed in a limestone-gypsum wet desulfurization system for a 2×600MW coal-fired power generating unit. The hardware deployment of the system includes: deploying a high-performance industrial server in the desulfurization control room to run the core algorithms of the acquisition module, large language processing module, and execution module; deploying explosion-proof industrial cameras on the top and sides of the absorption tower to acquire images of the slurry foam layer and slurry status; deploying ordinary industrial cameras at the demister outlet and circulating pump area to acquire images of the demister scale accumulation status and pump operation status; the server is connected to the OPC server of the DCS system via industrial Ethernet to read DCS operation data in real time; the server is connected to the actuators (slurry pump frequency converter, circulating pump frequency converter, oxidation fan frequency converter, demister flushing valve, slurry system actuator, gypsum discharge and dewatering system actuator, etc.) via 4-20mA signal line and / or Modbus TCP / IP protocol to issue control commands.

[0107] The large language model uses a locally deployed version of DeepSeek-V3, fine-tuned with inference enhancements (LoRA tuning, training data of 5000 inference chain samples). The model has approximately 671 bytes of parameters and is deployed on two servers equipped with NVIDIA A100 GPUs (80GB VRAM), using half-precision (FP16) loading to reduce VRAM usage. To ensure real-time performance, the model inference uses a streaming generation mode, with the latency of a single policy generation controlled within 2-5 seconds.

[0108] Example 1: The workflow of the data acquisition module is as follows: The first step is data acquisition. Using a 1-second sampling period, the following operational data is read from the DCS system: pH value of the absorber slurry (three-point average, range 5.0-6.0), slurry density (kg / m³), slurry level (m), inlet flue gas SO₂ concentration (mg / Nm³), outlet flue gas SO₂ concentration (mg / Nm³), inlet flue gas flow rate (10,000 Nm³ / h), inlet flue gas temperature (°C), frequency (Hz) and current (A) of circulation pumps #1 to #4, oxidation fan frequency (Hz), slurry supply pump frequency (Hz) and flow rate (m³ / h), demister differential pressure (Pa), slurry tank level and slurry density, gypsum discharge pump current, gypsum hydrocyclone pressure, and vacuum belt dewatering machine speed.

[0109] The second step is forecast data generation. An LSTM-based time-series forecast model is used, taking the inlet SO2 concentration, boiler load, and coal sulfur content data from the past 10 minutes as input, to predict the trend of inlet SO2 concentration changes over the next 5 minutes (outputting a forecast value every 15 seconds). Simultaneously, a simplified desulfurization reaction kinetic model is used, based on the current pH value, slurry density, and spray rate, to predict the trend of outlet SO2 concentration changes over the next 3 minutes.

[0110] The third step is machine vision data processing. Four industrial cameras acquire images at a frequency of 2 frames per second. After image preprocessing (denoising, enhancement, and size normalization), the images are fed into targeted vision models for processing: foam layer images are processed using a semantic segmentation model to extract foam regions, calculate the proportion of foam coverage area, average foam size, and foam morphological features (roundness and roughness); crystallization state images are processed using a classification model to determine the crystallization quality level (excellent / good / medium / poor); and demister fouling images are processed using a target detection model to identify fouling regions and calculate the fouling area ratio.

[0111] The fourth step is multi-source data spatiotemporal alignment. The multi-source data spatiotemporal alignment unit unifies all data to the same time reference (based on the system clock and synchronized by the BeiDou time synchronization equipment), and aligns data with different sampling frequencies to a 1-second time grid through interpolation or nearest neighbor sampling to form a multi-dimensional feature vector.

[0112] The fifth step involves packaging the multidimensional feature vectors together with pre-defined artificial constraints (SO2 emission limit of 35 mg / Nm³, pH safety range of 5.2-5.8, slurry density safety range of 1080-1150 kg / m³, slurry concentration target range of 15%-25%, etc.) to form structured multimodal fusion information, which is then transmitted to the large language processing module.

[0113] Example 2: Construction of Training Data and Model Fine-tuning for Inference Enhancement. This example details how to construct inference chain training data for fine-tuning of a large language model, as follows: Step 1: Collect historical operational data and expert decision records Historical desulfurization operation data from the past three years were extracted from the target power plant's DCS system, including continuous records of parameters such as inlet SO2 concentration, outlet SO2 concentration, pH value, slurry density, slurry supply rate, circulating pump frequency, and load. Simultaneously, operation records (including operation time, operation content, and parameter changes before and after the operation) were collected from operators during changes in critical operating conditions.

[0114] Step 2: Mark decision points and reasoning chains Key decision points in historical data were annotated by experts in the field of desulfurization, and a reasoning chain was written for each decision point. The annotation format follows a five-layer structure of "scenario identification → causal analysis → prioritization → means evaluation → strategy output".

[0115] Inference chain annotation for the scenario of "low pH value of slurry": Scene identification: The pH value was detected to have decreased from 5.55 to 5.28 within 15 minutes (at a rate of 0.018 mg / min), approaching the safety lower limit of 5.20. Simultaneously, the inlet SO2 concentration stabilized between 1050-1150 mg / Nm³ from 1100 mg / Nm³, without any sudden increase. The slurry pump frequency remained stable at 36 Hz, and the slurry flow rate showed no significant change. The circulation pump frequency remained stable.

[0116] Cause-and-effect analysis: The pH value decreased, but the inlet sulfur content did not increase, ruling out the cause of "increased consumption due to increased inlet sulfur." ​​The slurry supply did not decrease, ruling out "insufficient slurry supply" as a direct cause. Other possible causes were analyzed: a) Low limestone slurry concentration: The slurry concentration from the pulping system decreased from 18% to 15%, resulting in a decrease of approximately 17% in effective alkalinity per unit slurry supply. b) Changes in limestone quality: A recent change in limestone batches resulted in a decrease in CaCO3 content from 92% to 88%. c) Insufficient oxidation airflow: The oxidation fan frequency was 40Hz, but the slurry density increased from 1080 to 1120, indicating insufficient oxidation of calcium sulfite and a falsely low pH. Inspection of the pulping system: The pulping water valve opening increased from 45% to 55% (abnormal), indicating that the increased pulping water flow diluted the slurry. Overall judgment: The main cause is insufficient effective alkalinity in the slurry supply due to low pulping concentration.

[0117] Priority targets: 1. Highest priority: Prevent pH from dropping below 5.20 (safety threshold). 2. Second priority: Restore pH to the normal operating range of 5.35-5.55. 3. Third priority: Find and eliminate the root cause (pulping concentration issue).

[0118] Assessment of Methods: Option A (Short-term): Increase the pulp supply pump frequency (36→44Hz) to quickly replenish alkalinity. This is fast but only treats the symptoms, and may lead to a rapid increase in density. Option B (Fundamental): Adjust the pulping system to restore the pulp concentration to 18%. This takes 30-60 minutes to take effect, during which time Option A must be maintained. Option C (Combined): Immediately implement Option A (pulp supply frequency 42Hz), while simultaneously adjusting the pulp concentration. Gradually reduce the pulp supply frequency after the concentration has recovered.

[0119] Strategy Output: Increase the slurry pump frequency from 36Hz to 42Hz (instantaneous increase, rate 2Hz / s). Reduce the pulping water valve opening from 55% to 42% (target concentration 18%). Increase the limestone feed rate by 5% (to compensate for decreased quality). Increase the oxidation blower frequency from 40Hz to 44Hz. Expected Results: The pH value will recover to above 5.35 within 30 minutes; after the pulp concentration recovers, the slurry supply frequency can be reduced to 38Hz.

[0120] Step 3: Construct the training dataset The labeled inference chain data is organized according to the fine-tuned format specified in the instructions. The input is a description of the operating conditions, and the output is a structured text containing the inference chain and policy. The dataset covers no fewer than 5000 samples and is divided into training, validation, and test sets in an 8:1:1 ratio.

[0121] Step 4: Model fine-tuning training The LoRA method was used to fine-tune the basic large language model. Key hyperparameter settings: LoRA rank r=64, α=128; learning rate 2e-4; 3 training epochs; batch size 4 (8 gradient accumulation steps). The cross-entropy loss function was used, assigning different weights to the inference chain and policy components (inference chain weight 0.6, policy weight 0.4) to ensure the model learns both the inference process and the final decision simultaneously.

[0122] Step 5: Model Evaluation and Deployment The model's output was evaluated on the validation set for similarity to the labeled inference chain (using ROUGE-L and BLEU metrics) and policy accuracy (policy parameter deviation <10% was considered correct). After evaluation, the fine-tuned model was deployed to the power plant's local server.

[0123] Example 3: Reasoning and decision-making in a scenario of sudden increase in inlet sulfur and low pH. This example demonstrates the reasoning chain generation process of the large language processing module when handling concurrent abnormal scenarios, as detailed below: Input scenario: The unit load is stable at 400MW, but the sulfur content of the coal suddenly increases (laboratory data shows sulfur content rising from 0.8% to 1.2%), and the inlet SO2 concentration rises from 1200mg / Nm³ to 2100mg / Nm³ within 90 seconds. The current outlet SO2 concentration has risen from 18mg / Nm³ to 29mg / Nm³, the pH value has decreased from 5.52 to 5.26 (decreasing at a rate of 0.017 / min), and the foam layer area has increased from 8% to 22%. The current frequency of the slurry pump is 42Hz (approaching 84% of the upper limit of 50Hz), while the #1 and #2 circulation pumps are at 45Hz and 48Hz respectively. The slurry concentration is 16% (relatively low).

[0124] Output of the inference chain generation unit in the large language processing module: inference chain Scene Recognition Layer: The following abnormal characteristics were detected: 1. Inlet SO2 concentration: 1200→2100mg / Nm³, a 75% increase within 90 seconds, with an average rate of 10mg / Nm³ / s → "Rapid Increase" 2. pH value: 5.52→5.26, a decrease rate of 0.017 / min, approaching the safety lower limit of 5.20 → "pH Too Low and Decreasing Rapidly" 3. Foam layer: 8%→22%, exceeding the warning threshold of 20% → "Foaming Trend" 4. Slurry Pump: 42Hz (current), upper limit 50Hz, remaining adjustment space only 8Hz (approximately 19%) → "Limited Slurry Adjustment Margin" 5. Slurry Concentration: 16%, lower than the target of 18% → "Insufficient Effective Alkalinity per Unit Slurry Supply" 6. Outlet SO2: 18→29mg / Nm³, safety margin decreased from 17mg to 6mg → "High Environmental Pressure" Cause-and-effect analysis: Sudden increase in inlet sulfur (main cause) → Rapid increase in SO2 absorption load in the absorption tower → Accelerated consumption of slurry alkalinity → Rapid decrease in pH value → Decrease in desulfurization efficiency → Increase in outlet SO2. The pH value is already low (5.26). If it continues to drop below 5.20: the risk of equipment corrosion increases; desulfurization efficiency further decreases; and an alarm may be triggered. Analysis of the reasons for the rise in foam layer: Combustion of high-sulfur coal produces more fine particulate matter and organic matter; accumulation of organic matter in the slurry leads to changes in surface tension; the decrease in pH value also affects foam stability. The low slurry concentration (16%) means that the actual effective alkalinity of the current 42Hz slurry supply is only equivalent to the 37.3Hz slurry supply at the normal concentration (18%), exacerbating the insufficient alkalinity. Key constraints: Limited remaining adjustment space for the slurry pump (only 8Hz); the circulation pump is close to full frequency (only 2Hz space from 48Hz to 50Hz); the excessively thick foam layer limits the surge of the circulation pump (to avoid exacerbating foaming).

[0125] Priority Targets: P0 (Highest Priority): Prevent SO2 from exceeding the export limit (currently 29 → limit 35, safety margin 6mg, expected to exceed the limit if inbound sulfur continues to rise) P1 (Second Priority): Prevent pH from falling below 5.20 (currently 5.26, rate of decrease 0.017 / min, expected to reach the lower limit in 10-15 minutes) P2 (Third Priority): Control the growth of the foam layer (currently 22%, target ≤30%) P3 (Fourth Priority): Economic operation (control limestone and power consumption).

[0126] The evaluation layer outputs expected effects, response time, side effects / risks, and feasibility, as shown in Table 1.

[0127] Table 1

[0128] Conflict Resolution and Strategy Combination: The core contradiction is that a sudden increase in inlet sulfur necessitates a significant increase in circulation and slurry supply, but the excessively thick foam layer limits the surge in circulation. Decision Logic: Adopt a combined strategy of "slurry supply as the main focus + limited increase in circulation + increased slurry concentration + foam suppression". Slurry Supply Pump: Utilize the remaining 8Hz adjustment space, increase from 42Hz to 50Hz (full frequency) to prioritize alkalinity supply. Circulation Pump: Increase #2 circulation pump from 48Hz to 50Hz (full frequency), while #1 remains at 45Hz (do not start #3). Slurry System: Immediately adjust the slurry concentration from 16% to 18% (close the slurry water valve slightly), effective after 30-60 minutes. Flushing / Defoaming: Immediately adjust the demister flushing frequency from 4 hours / time to 2 hours / time to suppress foam growth; if foam continues to rise to 30%, trigger the addition of defoamer. Oxidation Air: Simultaneously increase by 3Hz (42→45Hz) to match oxidation requirements. Gypsum Discharge: Monitor density; if the upward trend is significant, increase the discharge pump frequency.

[0129] Strategy output layer: Outputs the specific methods of each implementing agency.

[0130] Expected Results: Outlet SO2: Expected to stop rising within 120 seconds, begin to decrease within 180 seconds, and drop below 25 mg / Nm³ within 300 seconds. pH: Expected to drop to a minimum of 5.22, and rebound to above 5.35 after the pulp concentration increase takes effect. Foam Layer: Expected to keep the foam area below 25% after increasing the rinsing frequency, without triggering the defoamer. Risk: If the inlet sulfur continues to rise above 2500 mg / Nm³ under full-frequency operation of the pulp pump, adjustment methods will be exhausted, requiring manual intervention.

[0131] Example 4: Multi-actuator collaborative control of the Action module. This example describes in detail how the Action module coordinates six types of actuators to execute the strategy generated by the large language processing module under the condition of a sudden increase in SO2 concentration at the inlet.

[0132] The strategy text generated by the large language processing module is (based on the output of Example 3): Grouting: Increase the grouting pump frequency from 42Hz to 50Hz, with an increase rate of 1Hz / s.

[0133] Circulation pump: Increase the frequency of circulation pump #2 from 48Hz to 50Hz, keep circulation pump #1 unchanged at 45Hz, and keep circulation pump #3 in standby mode.

[0134] Pulping solution: Increase the concentration of pulping solution from 16% to 18% and reduce the pulping water valve by 8%.

[0135] Oxidation air: Increase the frequency of the oxidation fan from 42Hz to 45Hz.

[0136] Flushing: The flushing interval for the demister has been adjusted from 4 hours to 2 hours, lasting for 180 minutes.

[0137] Gypsum discharge: Monitor the density; if it exceeds 1130 kg / m³, increase the discharge pump frequency by 5 Hz. The strategy parsing unit parses the strategy into six sets of structured instructions: slurry supply instruction, circulating pump instruction, slurry preparation instruction, oxidation air instruction, flushing instruction, and gypsum discharge instruction.

[0138] The instruction mapping unit converts the above instructions into specific control signals: Slurry pump frequency converter: 4-20mA output corresponds to 50Hz (15.0mA), speed limit 1Hz / s.

[0139] #2 Circulating Pump Inverter: Directly set to 50Hz.

[0140] Pulping water regulating valve: Actuator output -8% opening.

[0141] Oxidation fan frequency converter: frequency ramp up to 45Hz.

[0142] Demister flushing valve: Modify the PLC timer task table.

[0143] Gypsum discharge pump frequency converter: set condition monitoring.

[0144] The instruction verification unit verifies each instruction one by one: The slurry pump operates at 50Hz, which is equal to the upper limit of 50Hz.

[0145] The circulating pump operates at 50Hz, which is equal to the upper limit of 50Hz, and is successful.

[0146] After the pulping water valve is closed by 8%, the pulping concentration is expected to rise from 16% to 18.5%, which is within the target range (15%-25%).

[0147] The oxidation fan operates at 45Hz, which is below the upper limit of 50Hz, and passes the test.

[0148] Interlock check: The current liquid level in the absorption tower is 8.2m (safe range 6-9m), pass.

[0149] All instructions were verified and issued for execution in sequence: first, the pulping water valve was adjusted (to prepare high-concentration slurry in advance), and at the same time, the frequency of the slurry pump and oxidation fan was increased; the demister flushing schedule was updated.

[0150] Example 5: Closed-Loop Feedback Detection and Strategy Correction This embodiment demonstrates how the closed-loop feedback module detects deviations and triggers the large language processing module to correct the strategy when the effect after strategy execution does not meet expectations.

[0151] Scenario: After the strategy in Example 3 was implemented, the actual effect deviated from the expectation, as shown in Table 2.

[0152] Table 2

[0153] The closed-loop feedback module detects that the deviation exceeds the preset threshold (outlet SO2 deviation > 3 mg, pH drops below 5.20) and generates a feedback correction signal. The large language processing module's strategy correction unit receives the feedback correction signal and, combined with the latest operating conditions (inlet SO2 = 2250 mg / Nm³, outlet SO2 = 33 mg / Nm³, pH = 5.19, slurry pump = 50 Hz (reached its upper limit), slurry concentration is still 16.3% (not yet 18%)), generates a correction inference chain that includes deviation analysis. Step 1: Deviation Identification and Confirmation: Outlet SO2 at 300 seconds was 29 mg / Nm³, higher than the expected 25 mg, a shortfall of 4 mg. The pH value dropped below 5.20 to 5.19, violating safety constraints. The inlet sulfur was actually 2250 mg / Nm³, slightly higher than the original prediction of 2100 mg, but the deviation was only 7%, insufficient to fully explain the outlet shortfall. The pulp concentration only increased from 16% to 16.3%, far below the target of 18%, which is the key source of deviation.

[0154] Step 2: Root Cause Identification: The slurry pump was operating at full frequency of 50Hz, but the slurry concentration did not increase → insufficient effective alkalinity per unit slurry supply → the actual supplemented alkalinity was only equivalent to the normal concentration at a 44Hz slurry supply level. Analysis of slurry concentration adjustment command execution: The slurry water valve was closed by 8%, but the feeder response was delayed (limestone feeding was lagging), resulting in a slow increase in concentration. After the circulating pump reached full frequency of 50Hz, the marginal desulfurization efficiency was lower than expected (possibly related to partial blockage of the slurry circulation nozzles, but there was no direct sensor). pH value dropped below 5.20: Insufficient effective alkalinity in the slurry supply caused the pH to continue to decrease, approaching the lower limit.

[0155] Step 3: Correcting the Strategy and Generating Current Constraints: The slurry supply pump has reached its upper limit of 50Hz and cannot increase the slurry supply rate further. The circulation pump has reached its upper limit of 50Hz and cannot increase the circulation rate further. The slurry concentration is increasing slowly and will take more than 30 minutes to reach the target. Remaining available measures: 1. Start the #3 circulation pump (45Hz) → quickly improve desulfurization efficiency, but will exacerbate foaming (currently 28% → expected to be over 35%), requiring simultaneous use of defoamer. 2. Start the defoamer → suppress foaming, creating conditions for increasing the circulation rate. 3. Temporarily switch to the standby high-concentration slurry tank (concentration 22%). 4. Emergency addition of limestone powder (emergency measure). Recommended Correction Strategy: Immediately switch to the standby slurry tank (concentration 22%), equivalent to increasing the effective alkalinity of the slurry supply by 37.5%. Simultaneously start the #3 circulation pump (frequency 45Hz) to increase the spraying rate. Simultaneously start the defoamer (80L / h, continuously for 30 minutes) to suppress foaming. Expected results: SO2 levels at the outlet will drop below 28 mg within 90 seconds and below 23 mg within 180 seconds; pH levels will stop falling and start rising within 60 seconds.

[0156] Step 4: Learning from Experience. The root cause of this deviation is the slow response to pulp concentration adjustment, revealing that the response characteristics of the pulping system were not incorporated into the original inference model. Subsequent inference chains should include explicit evaluation of the pulping system's response time.

[0157] After the modified strategy was implemented, the SO2 at the outlet dropped to 24 mg / Nm³ within 120 seconds, the pH value rose back to 5.28, and the foam layer was controlled to within 30% under the action of the defoamer, which was a successful control.

[0158] Example 6: Multiple Iterative Corrections Until Convergence This example demonstrates the system's ability to support multiple rounds of feedback correction. The maximum number of corrections is set to 3, and the convergence condition is that the absolute value of the outlet SO2 concentration deviation is ≤2mg / Nm³.

[0159] After the first round of strategy execution, the deviation increased by 5.5 mg / Nm³. After the first round of correction, the deviation increased by 2.8 mg / Nm³ (still exceeding the threshold). After the second round of correction, the deviation increased by 1.2 mg / Nm³ (convergence condition met). The system then stopped correction and maintained the current control parameters. The system records the instructions and effects of each round of correction, forming a correction log for operators to review and for offline model training.

[0160] In summary, the desulfurization control system and method for coal-fired power plants described in this application achieve comprehensive intelligent and adaptive control of the desulfurization system through a large language model for enhanced fine-tuning based on inference, collaborative control of six types of actuators, and a closed-loop feedback correction mechanism. The system hardware can utilize mature industrial-grade equipment and servers, and the large language model can be fine-tuned using mainstream models such as DeepSeek and Qwen that support localized deployment, requiring no dedicated chips or customized hardware, thus exhibiting good scalability and economic efficiency. Implementation of this application can achieve an automatic operation rate of over 98% for the desulfurization system, significantly reducing operating costs while ensuring ultra-low emissions, resulting in significant economic, environmental, and social benefits.

[0161] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0162] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "N" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0163] Any process or method described in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more N executable instructions for implementing custom logic functions or processes, and the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as should be understood by those skilled in the art to which embodiments of this application pertain.

[0164] It should be understood that the various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, the N steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (FPGAs), field-programmable gate arrays (FPGAs), etc.

[0165] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0166] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A desulfurization control system for a coal-fired power plant, characterized in that, include: The acquisition module is used to collect and fuse operational data, prediction data, manual constraints, and machine vision data to obtain multimodal fusion data; The large language processing module is used to perform semantic annotation on the multimodal fusion data, input the annotated multimodal fusion data into the large language model that has been fine-tuned by reasoning enhancement, perform thought chain reasoning, and generate a desulfurization control strategy containing reasoning paths. The parsing module is used to parse the desulfurization control strategy into executable control instructions and send the control instructions to the target actuator of the desulfurization system of the coal-fired power plant. The closed-loop feedback module is used to collect the actual operating effect after each target actuator is executed. When the actual operating effect does not meet the expected effect, a feedback correction signal is generated and transmitted to the large language processing module. The large language processing module re-performs the thought chain reasoning or corrects the desulfurization control strategy based on the feedback correction signal.

2. The desulfurization control system for coal-fired power plants according to claim 1, characterized in that, The large language processing module is configured to: load a large language model that has been fine-tuned through inference enhancement, wherein the large language model is configured to perform the following steps: It receives fused information text containing a description of the current working condition scenario and key parameters, and outputs inference chain text containing scenario feature recognition, causal relationship analysis, control target priority ranking, adjustment means evaluation and strategy generation. The inference chain text embeds expert knowledge in the field of desulfurization, which includes control logic rules, equipment response characteristics, and safety constraints under different scenarios.

3. The desulfurization control system for coal-fired power plants according to claim 1, characterized in that, The training data used for the enhanced fine-tuning of inference consists of inference chain samples from multiple typical operating scenarios. The inference chain samples adopt a triplet structure, including: the operating scenario description text, the corresponding thought chain inference process, and the finally generated desulfurization control strategy.

4. The desulfurization control system for coal-fired power plants according to claim 1, characterized in that, The inference enhancement fine-tuning also includes scenario-policy mapping relationship learning, which is used to establish the correspondence between concurrent operating condition characteristics and optimal control methods, so that the large language model can identify multi-objective conflicts and cooperative relationships in the current operating scenario, and output an inference chain containing priority trade-offs and conflict resolution logic.

5. The desulfurization control system for coal-fired power plants according to any one of claims 1-4, characterized in that, The large language processing module includes: a semantic annotation unit, an inference chain generation unit, a policy extraction unit, a policy evaluation and ranking unit, and a policy correction unit. The semantic annotation unit is used to convert the multimodal fusion data into natural language descriptions to obtain text annotation information; The reasoning chain generation unit is used to take the text annotation information as input, load a large language model that has been fine-tuned by reasoning enhancement, perform reasoning chain reasoning, and generate reasoning chain text. The strategy extraction unit is used to extract structured desulfurization control strategies from the inference chain text. The strategy evaluation and ranking unit is used to perform multi-objective evaluation and ranking of the generated multiple control strategies. The strategy correction unit is used to receive the feedback correction signal from the closed-loop feedback module and generate a corrected desulfurization control strategy based on the feedback correction signal.

6. The desulfurization control system for coal-fired power plants according to claim 5, characterized in that, The inference chain text contains inference paths for scene diagnosis, causal analysis, control logic, and policy output. The inference chain text adopts a hierarchical structure, including: The scene recognition layer is used to describe the key features and abnormal patterns of the current working condition; The causal analysis layer is used to analyze the causal relationship between operating condition characteristics and desulfurization process parameters. The target priority layer is used to list the control objectives and constraints that need to be prioritized in the current scenario. The means evaluation layer is used to evaluate the feasibility, effectiveness, and side effects of various regulatory means; The strategy output layer is used to output specific control strategies and expected effects.

7. The desulfurization control system for coal-fired power plants according to claim 1, characterized in that, The execution module includes: a policy parsing unit, an instruction mapping unit, an instruction verification unit, and an instruction timing coordination unit, wherein, The strategy parsing unit is used to parse the desulfurization control strategy into structured instructions; The instruction mapping unit is used to map the structured instructions into control instructions for the target execution mechanism. The instruction verification unit is used to perform security verification on the control instructions of each target actuator; The instruction timing coordination unit is used to arrange the execution order and interval of multiple control instructions according to a preset timing logic.

8. The desulfurization control system for coal-fired power plants according to claim 1, characterized in that, The closed-loop feedback module includes: an effect acquisition unit, a deviation detection unit, and a correction trigger unit, wherein... The effect acquisition unit is used to acquire the actual operating effect of each target actuator after executing control commands; The deviation detection unit compares the actual operating effect with the expected effect and calculates the deviation of each control target. The correction triggering unit is used to generate a feedback correction signal and transmit it to the large language processing module when the deviation exceeds a preset threshold.

9. The desulfurization control system for coal-fired power plants according to any one of claims 1-8, characterized in that, The actuator includes at least one of the following: a slurry supply actuator, a circulating pump actuator, a slurry preparation actuator, an oxidation air actuator, a flushing actuator, and a gypsum discharge and dewatering actuator; The operational data includes real-time operational parameters collected by the desulfurization system of the coal-fired power plant; the prediction data includes prediction information generated based on a time-series prediction model; the manual constraints include manually set constraints and operational rules; and the machine vision data includes visual information acquired by image acquisition devices deployed at key locations in the desulfurization system of the coal-fired power plant.

10. A desulfurization control method for a coal-fired power plant, characterized in that, The method, applied to the desulfurization control system of a coal-fired power plant as described in any one of claims 1 to 9, comprises the following steps: Collect and fuse operational data, prediction data, manual constraints, and machine vision data to obtain multimodal fusion data; The multimodal fusion data is semantically annotated, and the annotated multimodal fusion data is input into a large language model that has been fine-tuned by reasoning enhancement to perform thought chain reasoning and generate a desulfurization control strategy containing reasoning paths. The desulfurization control strategy is parsed into executable control commands, and the control commands are sent to the target actuators of the desulfurization system in the coal-fired power plant. The actual operating effect of each target actuator is collected. When the actual operating effect does not meet the expected effect, a feedback correction signal is generated and transmitted to the large language processing module. The large language processing module re-performs the thought chain reasoning or corrects the desulfurization control strategy based on the feedback correction signal until the expected effect is achieved.

Images