Flue gas desulfurization treatment system based on intelligent control

Abstract

The invention discloses a flue gas desulfurization treatment system based on intelligent control, and relates to the technical field of industrial environmental protection and automatic control, and the flue gas desulfurization treatment system comprises a data acquisition layer used for acquiring multi-source data related to a desulfurization process in real time; the data processing and model layer is used for processing, analyzing and modeling the collected data; and the intelligent control layer is used for generating and executing a control strategy based on model output. According to the flue gas desulfurization treatment system based on intelligent control, process state prediction, equipment health assessment and multi-target optimization decision are integrated, and multi-dimensional cooperative control over the flue gas desulfurization process is achieved; the method can dynamically coordinate the desulfurizer feeding strategy inside and outside the furnace according to the real-time working condition, optimizes the balance between desulfurizer consumption and system energy consumption on the premise of ensuring that the emission index meets the requirement, and improves the system energy consumption through the quantitative evaluation and fusion decision of the operation state of key equipment. And the adaptability of the system to load and coal quality fluctuation and the overall operation stability are improved.

Description

Technical Field

[0001] This invention relates to the field of industrial environmental protection and automatic control technology, specifically to a flue gas desulfurization system based on intelligent control. Background Technology

[0002] With increasingly stringent environmental regulations and the continuous expansion of industrial production, the efficient removal of pollutants such as sulfur dioxide from flue gas has become crucial for industrial emission control. Common desulfurization technologies include wet, semi-dry, and dry methods, with the semi-dry method being widely adopted due to its balance of efficiency and economy. However, in actual operation, factors such as coal quality fluctuations, load changes, and operating condition switching cause the desulfurization process to exhibit complex characteristics such as multivariate coupling, high inertia, and nonlinearity. Traditional control methods struggle to achieve stable, efficient, and economical desulfurization results and are highly dependent on personnel experience.

[0003] To improve the automation and intelligence level of desulfurization systems, existing research has proposed corresponding solutions. For example, the patent publication number "CN119847089A" describes an automatic control system for sulfur removal from flue gas. This system constructs an in-furnace / outside SO2 prediction model, a limestone / hydrated lime control model, a cooperative coordination model, and a health status monitoring and early warning model. It then employs algorithms such as LSTM, fuzzy logic, PID neural networks, and multi-objective optimization to achieve real-time prediction, optimized control, and status monitoring of the desulfurization process. This technical solution improves the adaptive adjustment capability and operational stability of the desulfurization system to a certain extent and has good engineering application value.

[0004] However, the existing systems still have the following areas for improvement: First, their intelligent control architecture relies heavily on the cascading and simple coordination of multiple independent models, leaving room for improvement in overall system integration and collaborative optimization capabilities. This is particularly evident when dealing with drastic fluctuations in operating conditions, where the response coordination between modules and the overall control robustness are insufficient. Second, the system's early warning models are mostly threshold or trend alarms, limiting their ability to predict faults proactively and make autonomous decisions, thus failing to achieve truly preventative maintenance and intelligent intervention. Furthermore, in achieving a dynamic balance between multiple objectives such as desulfurization efficiency, operating costs, and equipment reliability, existing methods are mostly based on fixed rules or offline optimization, making it difficult to achieve real-time, global, and adaptive optimal control in complex operating environments. Therefore, how to construct a more integrated, more autonomous, and more proactive and adaptive intelligent desulfurization control system has become a pressing technical problem in the current industrial environmental protection field.

[0005] This invention aims to address the shortcomings of the existing technologies by providing a flue gas desulfurization system based on intelligent control. By deeply integrating advanced artificial intelligence algorithms with industrial process mechanisms, an integrated intelligent control architecture is constructed to achieve full-condition adaptive optimization, intelligent fault early warning, and autonomous decision-making in the desulfurization process, thereby further improving the operating efficiency, economy, and reliability of the desulfurization system. Summary of the Invention

[0006] The purpose of this invention is to provide a flue gas desulfurization system based on intelligent control to solve the problems mentioned in the background art.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a flue gas desulfurization treatment system based on intelligent control, comprising:

[0008] The data acquisition layer is used to collect multi-source data related to the desulfurization process in real time.

[0009] The data processing and modeling layer is used to process, analyze, and model the collected data.

[0010] And an intelligent control layer, used to generate and execute control strategies based on model output;

[0011] The intelligent control layer includes a global collaborative optimization unit and at least one local execution unit;

[0012] The global collaborative optimization unit is used to simulate and evaluate multiple control strategy combinations based on the state information output by the data processing and model layer, under a preset optimization objective, and generate global collaborative control instructions.

[0013] The local execution unit is used to receive and execute the global collaborative control command to adjust the designated desulfurization process equipment;

[0014] The data processing and modeling layer includes:

[0015] The process status prediction module is used to predict the concentration or amount of substances generated at key desulfurization nodes.

[0016] The equipment health status assessment module is used to assess and output quantitative indicators of the operating status of key equipment based on non-process parameter data of equipment operation;

[0017] And a strategy verification and correction module, used to cross-verify and correct the key prediction results output by the process state prediction module;

[0018] When generating the global collaborative control command, the global collaborative optimization unit simultaneously considers the prediction results of the process state prediction module and the quantitative indicators of the operating state output by the equipment health status assessment module.

[0019] Furthermore, the global collaborative optimization unit includes a digital twin simulation module and a strategy optimization module;

[0020] The digital twin simulation module has a built-in dynamic simulation model of the desulfurization system.

[0021] The strategy optimization module is used in the digital twin simulation module to simulate in parallel multiple future control time domains various control strategies formed by different combinations of control parameters of the local execution unit, and to evaluate the comprehensive benefits of each strategy according to a preset reward function.

[0022] The reward function integrates at least the desulfurization efficiency index, the material consumption index, and the equipment status penalty item related to the output of the equipment health status assessment module.

[0023] The strategy optimization module selects a control strategy that meets the overall benefit requirements based on the evaluation results and generates the global collaborative control command.

[0024] Furthermore, the equipment health status assessment module includes an equipment feature extraction network and a health quantification model;

[0025] The device feature extraction network is used to extract features from the historical and real-time runtime sequence data of key equipment, wherein the time sequence data includes at least one of vibration, current, and pressure data.

[0026] The health measurement model is used to output a health index representing the current health level of the device, as well as the probability of a specific type of failure occurring within a set future time period, based on the extracted features.

[0027] The global collaborative optimization unit adjusts the control parameter constraint range of the corresponding device or its weight in the reward function based on the health index or the failure probability.

[0028] Furthermore, the strategy verification and correction module includes a first prediction unit, a second prediction unit, and a result calibration unit;

[0029] The first prediction unit uses a first type of algorithm model to predict the sulfur dioxide concentration at the inlet of the desulfurization island and outputs a first predicted value.

[0030] The second prediction unit uses a second type of algorithm model, different from the first type of algorithm model, to predict the sulfur dioxide concentration of the same object and outputs a second prediction value.

[0031] The result calibration unit is used to introduce a third type of state information when the deviation between the first predicted value and the second predicted value exceeds a preset threshold, and to perform fusion calibration on the first predicted value and the second predicted value based on the third type of state information, and output a calibrated predicted value with a confidence interval.

[0032] The third type of state information is a qualitative pattern label obtained by evaluating the combustion state through image recognition technology.

[0033] Furthermore, the system also includes a self-learning feedback loop;

[0034] The self-learning feedback loop is used to continuously compare the actual measured value with the calibration prediction value after the control strategy is executed based on the calibration prediction value;

[0035] When the comparison error persists and the confidence interval is lower than the set standard, the offline analysis mechanism is triggered.

[0036] The offline analysis mechanism packages the raw data, intermediate predicted values, calibration process data and final actual values ​​within the time period into special working condition samples, and extracts potential unmodeled features or new rules through anomaly detection and causal analysis algorithms.

[0037] The analysis results output by the offline analysis mechanism are fed back to the process state prediction module or the strategy verification and correction module for parameter updates or structural fine-tuning of the online model.

[0038] Furthermore, the local execution unit includes an in-furnace desulfurizer control module and an external desulfurizer control module;

[0039] The in-furnace desulfurizing agent control module is used to adjust the operating parameters of the equipment related to adding limestone into the furnace according to the global collaborative control command;

[0040] The external desulfurizing agent control module is used to adjust the operating parameters of the equipment related to adding quicklime to the desulfurization tower according to the global collaborative control command;

[0041] The global collaborative optimization unit, through the digital twin simulation module, collaboratively optimizes the control parameters of the in-furnace desulfurizer control module and the out-of-furnace desulfurizer control module to achieve dynamic load distribution and cost balance between the in-furnace and out-of-furnace desulfurization processes.

[0042] Furthermore, the system also includes an early warning and autonomous fault-tolerant unit;

[0043] The early warning and autonomous fault-tolerant unit is connected to the equipment health status assessment module and the global collaborative optimization unit;

[0044] The early warning and autonomous fault-tolerant unit is pre-set with a rule library, which defines control constraint adjustment strategies and early warning information corresponding to different quantitative indicators of equipment health status and failure probabilities.

[0045] When the indicators output by the device health status assessment module trigger the rule base, the early warning and autonomous fault tolerance unit sends a constraint adjustment signal to the global collaborative optimization unit and generates early warning information containing maintenance suggestions.

[0046] Furthermore, the data acquisition layer is connected to the boiler control system, desulfurizer feeding control system, and flue gas monitoring system through various industrial communication protocols;

[0047] The collected data points include boiler load, total air volume, primary air volume and temperature, secondary air volume and temperature, bed temperature, bed pressure, coal feed rate, flue gas parameters at furnace outlet, desulfurizing agent feed flow rate and frequency, conveying air pressure, and sulfur dioxide concentration at the inlet and outlet of the desulfurization tower.

[0048] Furthermore, the process state prediction module adopts a prediction model based on a long short-term memory network;

[0049] The data processing and model layer also includes a data preprocessing module, which is used to clean, normalize, and perform feature selection operations on the raw data uploaded by the data acquisition layer.

[0050] Furthermore, the system also includes a human-computer interaction layer;

[0051] The human-computer interaction layer is used to display the overall operating status of the system, the prediction results of the process status prediction module, the health index of the equipment health status assessment module, the control strategy generated by the global collaborative optimization unit, and the early warning information generated by the early warning and autonomous fault tolerance unit.

[0052] This invention provides a flue gas desulfurization system based on intelligent control. It has the following beneficial effects:

[0053] This intelligent control-based flue gas desulfurization system effectively integrates process status prediction, equipment health assessment, and multi-objective optimization decision-making through an integrated architecture of "data acquisition - multi-model collaborative analysis - global intelligent optimization - closed-loop execution feedback," achieving multi-dimensional collaborative control of the flue gas desulfurization process. It can dynamically coordinate the desulfurizing agent dosing strategies inside and outside the furnace based on real-time operating conditions, optimizing the balance between desulfurizing agent consumption and system energy consumption while ensuring emission standards are met. Furthermore, through quantitative assessment and integrated decision-making of the operating status of key equipment, it enhances the system's adaptability to load and coal quality fluctuations and improves overall operational stability.

[0054] This intelligent control-based flue gas desulfurization system enhances the reliability of core status perception through embedded cross-validation and self-learning mechanisms. Its early warning and autonomous fault-tolerance functions based on equipment health status shift from post-fault handling to pre-fault early warning and strategy adjustment. This reduces the risk of unplanned downtime and performance degradation caused by potential equipment failures. Furthermore, by providing transparent decision-making basis and status visualization, it alleviates the burden of manual judgment on operators, providing technical support for the long-term economic, stable, and reliable operation of the desulfurization system. Attached Figure Description

[0055] Figure 1 This is a data flow diagram of a flue gas desulfurization treatment system based on intelligent control according to the present invention;

[0056] Figure 2 This is a flowchart illustrating the control strategy optimization process of a flue gas desulfurization system based on intelligent control according to the present invention. Detailed Implementation

[0057] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0058] Please see Figure 1 and Figure 2 This invention provides a technical solution: a flue gas desulfurization treatment system based on intelligent control, comprising:

[0059] The data acquisition layer is used to collect multi-source data related to the desulfurization process in real time.

[0060] The data processing and modeling layer is used to process, analyze, and model the collected data.

[0061] And an intelligent control layer, used to generate and execute control strategies based on model output;

[0062] The intelligent control layer includes a global collaborative optimization unit and at least one local execution unit;

[0063] The global collaborative optimization unit is used to simulate and evaluate multiple control strategy combinations based on data processing and state information output from the model layer, under a preset optimization objective, and generate global collaborative control commands.

[0064] The local execution unit is used to receive and execute global collaborative control commands to regulate the designated desulfurization process equipment;

[0065] The data processing and modeling layer includes:

[0066] The process status prediction module is used to predict the concentration or amount of substances generated at key desulfurization nodes.

[0067] The equipment health status assessment module is used to assess and output quantitative indicators of the operating status of key equipment based on non-process parameter data of equipment operation;

[0068] And a strategy verification and correction module, used to cross-verify and correct the key prediction results output by the process state prediction module;

[0069] When generating global collaborative control commands, the global collaborative optimization unit considers both the prediction results of the process state prediction module and the quantitative indicators of operating status output by the equipment health status assessment module.

[0070] It should be further explained that the system architecture includes a data acquisition layer, a data processing and modeling layer, and an intelligent control layer. The data acquisition layer acquires multi-source operational data in real time, including boiler load, total air volume, primary and secondary air volume and temperature, bed temperature, bed pressure, coal feed rate, flue gas temperature, pressure and composition at the furnace outlet, sulfur dioxide concentration at the desulfurization island inlet and outlet, desulfurizer feed flow rate and frequency, and air pressure in the conveying pipeline, through sensors and instruments deployed on the boiler, flue gas system, desulfurizer feed device, and continuous emission monitoring system. This data is then transmitted to the data processing and modeling layer via industrial network protocols.

[0071] The data processing and modeling layer first performs preprocessing operations on the received raw data, including cleaning, normalization, and feature filtering. Then, it performs in-depth analysis through multiple integrated functional modules. The process state prediction module in this layer uses a model built on a long short-term memory network to predict the sulfur dioxide generation at the desulfurization island inlet at future moments, utilizing preprocessed historical and real-time data sequences. The equipment health status assessment module constructs specialized deep residual network models for key equipment such as limestone feeders and conveyor fans, analyzes their vibration, current, and other operational sequence signals, and outputs a health index characterizing the current wear or blockage trend of the equipment and the probability of failure.

[0072] The strategy verification and correction module establishes a parallel first prediction unit and a second prediction unit, which independently predict the same key parameter (such as sulfur dioxide concentration at the desulfurization island inlet) using a long short-term memory network and a simplified mechanism model based on reaction kinetics, respectively. When the prediction results of the two units deviate from the set range, the module activates the result calibration unit, which introduces the qualitative mode label of combustion state (such as stable, off-center combustion, and fluctuation) obtained by infrared thermal imager and identified by convolutional neural network as an additional feature input to a pre-trained calibration network, performs weighted fusion of the two prediction values, and finally outputs a calibration prediction value with confidence assessment.

[0073] The intelligent control layer, serving as the system's decision-making center, adopts an architecture where a global collaborative optimization unit and a local execution unit work together. The global collaborative optimization unit incorporates a digital twin simulation environment for the entire desulfurization process. Its strategy optimization engine takes verified and corrected process state predictions and equipment health indices as inputs, and simulates multiple combinations of control strategies (such as simultaneously adjusting the limestone feeding frequency inside the furnace and the amount of quicklime injected outside the furnace) in parallel and continuously within the simulation environment. Based on a reward function that comprehensively considers the sulfur dioxide emission compliance rate, total desulfurizing agent consumption, system energy consumption, and equipment health penalties, it evaluates the long-term comprehensive benefits of each strategy, thereby generating global collaborative control commands.

[0074] The local execution unit, including the in-furnace desulfurizer control module and the external desulfurizer control module, receives and executes the aforementioned instructions, precisely adjusting field equipment parameters such as feeder frequency and metering valve opening. The system also features a self-learning feedback loop. When control is executed based on calibration prediction values, and the error between the actual measured value and the predicted value persists with low confidence, all data for the relevant time period will be packaged and sent to the offline analysis module. Unsupervised anomaly detection and causal discovery algorithms will be used to uncover potential new features or rules, which will then be fed back to the online model for automatic parameter fine-tuning and updates.

[0075] In addition, the early warning and autonomous fault tolerance unit queries the predefined rule base in real time based on the output of the equipment health status assessment module. If the health index or failure probability exceeds the threshold, it sends an instruction to the global collaborative optimization unit to dynamically adjust the control constraints, and at the same time generates early warning information containing specific maintenance suggestions to the human-machine interface.

[0076] The entire system achieves coordinated optimization and autonomous operation of the flue gas desulfurization process through the closed-loop linkage of the aforementioned data flow and control flow.

[0077] The global collaborative optimization unit includes a digital twin simulation module and a strategy optimization module;

[0078] The digital twin simulation module has a built-in dynamic simulation model of the desulfurization system;

[0079] The strategy optimization module is used in the digital twin simulation module to simulate in parallel multiple future control time domains various control strategies formed by different combinations of control parameters of local execution units, and to evaluate the comprehensive benefits of each strategy according to a preset reward function.

[0080] The reward function integrates at least the desulfurization efficiency index, the material consumption index, and the equipment status penalty items related to the output of the equipment health status assessment module.

[0081] The strategy optimization module selects a control strategy that meets the overall benefit requirements based on the evaluation results and generates global collaborative control instructions.

[0082] It should be further explained that in the global collaborative optimization unit, the digital twin simulation module integrates multi-physics field mechanism equations and data-driven proxy models such as boiler combustion, flue gas flow, heat and mass transfer and desulfurization chemical reaction to construct a dynamic virtual image of the entire process of the desulfurization system from coal feeding to chimney emission; the input of the simulation model receives real-time prediction data and equipment status data from data processing and model layer, and its output can simulate the state evolution of the system in multiple sampling periods under different combinations of control variables.

[0083] The strategy optimization module runs in the virtual environment provided by the digital twin simulation module. Its workflow is as follows: First, based on the current system state and optimization objectives, it generates multiple parameter combination schemes covering multiple control variables such as the limestone feeding frequency in the furnace, the hydrated lime feeding flow rate outside the furnace, and the circulating fan speed, forming a set of candidate control strategies to be evaluated. Then, these strategies are executed in parallel in the simulation environment to quickly deduce the system's operating trajectory within a set time window under each strategy and simultaneously calculate multiple performance indicators.

[0084] These performance indicators are integrated into a preset reward function for quantitative evaluation. The specific components of the reward function include: a desulfurization efficiency item reflecting the deviation of sulfur dioxide emission concentration from the set value; a material cost item characterizing the total consumption of limestone and hydrated lime; an energy consumption item related to the operating power of fans and pumps; and an equipment status penalty item directly linked to the health index or failure probability output by the equipment health status assessment module. This penalty item ensures that when the equipment health deteriorates or the failure risk increases, the evaluation benefits of the relevant control strategy will be reduced accordingly.

[0085] The strategy optimization module compares the comprehensive evaluation values ​​of all candidate strategies calculated by the reward function, selects one or more strategies whose evaluation values ​​meet the preset requirements, and parses the corresponding control parameters into specific global collaborative control instructions that can be issued to each local execution unit, thereby completing one optimization decision cycle.

[0086] The equipment health status assessment module includes an equipment feature extraction network and a health measurement model;

[0087] The equipment feature extraction network is used to extract features from the historical and real-time runtime sequence data of key equipment. The time sequence data includes at least one of vibration, current, and pressure data.

[0088] A health measurement model is used to output a health index representing the current health level of a device, as well as the probability of a specific type of failure occurring within a set future time period, based on extracted features.

[0089] The global collaborative optimization unit adjusts the control parameter constraint range of the corresponding equipment or adjusts its weight in the reward function based on the health index or failure probability.

[0090] It should be further explained that in the specific implementation of the equipment health status assessment module, the equipment feature extraction network adopts a hybrid architecture combining a one-dimensional convolutional neural network and a long short-term memory network. This network receives historical and real-time runtime sequence data streams from key equipment such as limestone feeders and conveying fans. These data include, but are not limited to, three-dimensional vibration acceleration signals, three-phase current signals, and conveying pipeline pressure signals.

[0091] The network first extracts and reduces the dimensionality of the original time-series signal using a one-dimensional convolutional layer. Then, it uses a long short-term memory layer to capture the temporal dependencies of the features, ultimately outputting a multi-dimensional feature vector that comprehensively represents the mechanical and electrical states of the equipment. The health measurement model uses this feature vector as input. It is itself a fully connected neural network. After supervised training on a large amount of historical normal operation data and known fault case data, the model outputs a scalar value between zero and one as the real-time health index of the equipment. The closer the index is to zero, the closer the equipment is to being in good condition; the closer it is to one, the more severe the degradation or abnormality. Simultaneously, the model also outputs a vector representing the probability of the equipment experiencing a predefined specific fault type, such as "bearing wear," "pipe blockage," or "coil overheating," within the next few operating hours.

[0092] During operation, the global collaborative optimization unit reads the aforementioned health index and failure probability in real time. This unit has pre-set mapping rules; for example, when the health index of a device falls below a first threshold or a specific failure probability exceeds a second threshold, the adjustable range of its control variables will be narrowed during the optimization process, such as limiting the maximum operating frequency of the feeder in a digital twin simulation. Simultaneously, in the reward function of the strategy optimization module, terms positively correlated with the device's operating intensity are assigned a dynamic penalty coefficient. This coefficient increases as the health index decreases or the failure probability increases, thereby guiding the optimization algorithm to automatically avoid control strategies that may exacerbate equipment wear or induce failures, achieving deep collaboration between production control and equipment maintenance.

[0093] The strategy verification and correction module includes a first prediction unit, a second prediction unit, and a result calibration unit;

[0094] The first prediction unit uses a first type of algorithm model to predict the sulfur dioxide concentration at the inlet of the desulfurization island and outputs the first predicted value.

[0095] The second prediction unit uses a second type of algorithm model, different from the first type of algorithm model, to predict the sulfur dioxide concentration of the same object and outputs a second prediction value.

[0096] The result calibration unit is used to introduce third-class state information when the deviation between the first predicted value and the second predicted value exceeds a preset threshold, and to perform fusion calibration on the first predicted value and the second predicted value based on the third-class state information, and output a calibration predicted value with confidence interval.

[0097] The third type of status information is a qualitative pattern label obtained by evaluating the combustion status through image recognition technology.

[0098] It should be further explained that the specific implementation of the strategy verification and correction module is as follows: The first prediction unit adopts a data-driven model based on a long short-term memory network. This model takes pre-processed real-time and historical time-series data such as boiler load, air volume, and coal feed as input, and outputs the first predicted value of sulfur dioxide concentration at the desulfurization island inlet.

[0099] The second prediction unit adopts a simplified mechanism model based on mass balance and chemical reaction kinetics. This model takes the same real-time operating parameters (such as coal feed rate, total air volume, and oxygen content) and design parameters such as coal sulfur content and calcium-sulfur ratio as inputs. By solving the preset algebraic and differential equations, it outputs a second predicted value of sulfur dioxide concentration for the same object.

[0100] The system continuously calculates the absolute or relative deviation between the first and second predicted values. When this deviation exceeds a threshold set based on historical prediction error statistics, the result calibration unit is activated. This unit first calls an independent combustion state analysis submodule, which receives a sequence of flame images acquired by an infrared thermal imager installed at the furnace observation port. It then extracts and classifies the image features using a pre-trained convolutional neural network, outputting a qualitative mode label characterizing the current combustion stability, such as "stable combustion," "flame deflection," or "violent pulsation."

[0101] Subsequently, the result calibration unit inputs the first predicted value, the second predicted value, and the aforementioned combustion mode label into a calibration network composed of a fully connected neural network. This calibration network has been trained using historical data containing various operating conditions and their corresponding real measurements. Its output is a corrected sulfur dioxide concentration calibration prediction value. At the same time, the network also outputs a confidence interval characterizing the uncertainty of this prediction.

[0102] This confidence interval is generated through an uncertainty quantification layer within the calibration network, reflecting the model's self-assessment of its prediction accuracy based on the current combination of input features. Ultimately, this calibrated prediction and confidence interval are provided to the global collaborative optimization unit as key state inputs for policy optimization.

[0103] The system also includes a self-learning feedback loop;

[0104] The self-learning feedback loop is used to continuously compare the actual measured value with the calibration prediction value after the control strategy is executed based on the calibration prediction value;

[0105] When the comparison error persists and the confidence interval is lower than the set standard, the offline analysis mechanism is triggered.

[0106] The offline analysis mechanism packages the raw data, intermediate predicted values, calibration process data and final actual values ​​within the time period into special working condition samples, and extracts potential unmodeled features or new rules through anomaly detection and causal analysis algorithms.

[0107] The analysis results output by the offline analysis mechanism are fed back to the process state prediction module or the strategy verification and correction module for parameter updates or structural fine-tuning of the online model.

[0108] It should be further explained that the specific implementation process of the self-learning feedback loop is as follows: After the system generates and executes the control strategy based on the calibration prediction value output by the strategy verification and correction module, the loop continuously compares the actual sulfur dioxide concentration measurement value obtained by the online monitoring instrument at the desulfurization island inlet or chimney outlet with the calibration prediction value previously used for decision-making.

[0109] The system is equipped with an error accumulation counter and a confidence threshold. When the actual measured value continuously deviates from the confidence interval of the calibrated predicted value within multiple consecutive sampling periods, and the width of the interval itself (characterizing the prediction uncertainty) is lower than the preset reliability standard, it is determined that the current prediction model may have insufficiently modeled factors when facing this specific working condition, and the offline analysis mechanism is automatically triggered.

[0110] The mechanism first creates a data packet called a special operating condition sample, which contains all the raw sensor readings from the data acquisition layer during that period, the intermediate output values ​​of the first and second prediction units in the strategy verification and correction module, the processing log of the result calibration unit, the final calibration prediction value and its confidence interval, and the corresponding actual measurement value.

[0111] This data packet is transmitted to an offline analysis module independent of the real-time control loop. This module first uses an unsupervised anomaly detection algorithm based on isolated forests or local anomalies to scan the multi-dimensional operating parameters in the data packet and identify feature combinations that show statistical differences from historical normal operating conditions.

[0112] Subsequently, constraint- or rating-based causal discovery algorithms, such as PC or FCI algorithms, are applied to infer the potential causal or correlation relationships between these anomalous features and the final prediction error, aiming to extract new feature variables or implicit physical rules that have not been considered by existing online prediction models.

[0113] The analysis results output by the offline analysis module, such as identified key anomaly parameters and their correlation weights with errors, or derived new correction rules, are encapsulated as structured update information. The system is designed with a secure model update interface, through which this update information is fed back to the data processing and model layers. Specifically, it can be used to: supplement the input feature vectors of the Long Short-Term Memory network in the process state prediction module; or to fine-tune the internal weights of the result calibration network in the strategy verification and correction module; and to expand the fault feature library of the equipment health status assessment module. This process enables the system to autonomously mine new knowledge from the data and iteratively update the online model when encountering persistent unknown disturbances, thereby gradually expanding its adaptability.

[0114] The local execution unit includes an in-furnace desulfurizer control module and an external desulfurizer control module;

[0115] The in-furnace desulfurizing agent control module is used to adjust the operating parameters of the equipment related to adding limestone into the furnace according to the global coordinated control instructions;

[0116] The external desulfurizing agent control module is used to adjust the operating parameters of the equipment related to adding quicklime to the desulfurization tower according to the global coordinated control instructions;

[0117] The global collaborative optimization unit uses a digital twin simulation module to collaboratively optimize the control parameters of the in-furnace desulfurizer control module and the external desulfurizer control module, so as to achieve dynamic load distribution and cost balance between the in-furnace and external desulfurization processes.

[0118] It should be further explained that, in the local execution unit, the in-furnace desulfurizer control module and the external desulfurizer control module, as the final execution terminals of the system, are connected to the limestone powder silo feeding system and the slaked lime slurry preparation and injection system on site through actuators such as frequency converters and regulating valves, respectively. The specific process of the global collaborative optimization unit in co-optimizing these two modules through its digital twin simulation module and strategy optimization module is as follows: In each optimization decision loop, the strategy optimization module constructs a set of candidate control strategies in the digital twin environment, whose control variables simultaneously include the target feeding frequency (or rotation speed) of the in-furnace limestone feeding system and the target slurry flow rate (or injection pressure) of the external slaked lime system.

[0119] The digital twin simulation module simulates and executes each strategy, composed of combinations of in-furnace and external control parameters, in parallel based on boundary conditions such as current boiler load, flue gas parameters, and equipment health status, and then extrapolates its effects. In the reward function of the strategy optimization module, the material cost term is designed as the sum of the product of in-furnace limestone consumption and unit price, plus the product of external hydrated lime consumption and unit price; the desulfurization efficiency term evaluates the degree to which the final outlet concentration of the simulated flue gas, after undergoing simulated in-furnace and external desulfurization, complies with environmental limits.

[0120] Through this integrated simulation and evaluation, the strategy optimization module can solve for the optimal combination of in-furnace and out-of-furnace desulfurizing agent dosages that maximizes total material cost, system energy consumption, and equipment wear while meeting desulfurization efficiency requirements, thus achieving dynamic load allocation.

[0121] The optimal solution is translated into specific control commands: for example, if simulation results show that slightly increasing the in-furnace desulfurization ratio is more economical under the current operating conditions, a command is generated to appropriately increase the limestone feeding frequency and possibly fine-tune the external lime slaked flow rate; conversely, the opposite adjustment is made. This process is carried out continuously, thus enabling real-time and dynamic balancing of the contribution and cost of the in-furnace and external desulfurization processes based on changes in coal sulfur content, load commands, and desulfurizer market prices.

[0122] The system also includes an early warning and autonomous fault-tolerant unit;

[0123] The early warning and autonomous fault-tolerant unit is connected to the equipment health status assessment module and the global collaborative optimization unit;

[0124] The early warning and autonomous fault-tolerant unit has a pre-built rule base, which defines the control constraint adjustment strategies and early warning information corresponding to different quantitative indicators of equipment health status and failure probabilities.

[0125] When the indicators output by the equipment health status assessment module trigger the rule base, the early warning and autonomous fault tolerance unit sends a constraint adjustment signal to the global collaborative optimization unit and generates early warning information containing maintenance suggestions.

[0126] It should be further explained that the specific implementation of the early warning and autonomous fault-tolerant unit relies on a structured rule base. This rule base is stored in the form of knowledge representation, explicitly mapping the numerical ranges of the health index and various fault probabilities output by the equipment health status assessment module to specific control constraint adjustment instructions and early warning content. For example, the rule base can define: when the health index of the limestone feeding conveyor fan is continuously lower than threshold A, or the probability of "pipeline blockage" exceeds threshold B, a constraint adjustment signal is generated and sent to the global collaborative optimization unit. The content of this signal is to require the optimizer to set the upper limit of the fan's drive frequency to 70% of the rated value in subsequent strategy optimization, and to apply additional penalty weights to the relevant energy consumption items of the fan in the reward function.

[0127] Simultaneously, the rule base generates an early warning message, which includes not only the alarm level, triggering device name and time, but also specific recommended maintenance actions, such as "check for material accumulation in the conveying pipeline and recommend preventative purging." This unit monitors the data stream from the equipment health status assessment module in real time. Once any indicator meets the triggering condition of any rule entry, the aforementioned mapping operation is immediately executed: constraint adjustment signals are sent in real time to the strategy optimization module of the global collaborative optimization unit via the internal communication interface, dynamically modifying the boundary conditions of its optimization problem; the early warning message is formatted and sent to the human-machine interface layer for audible and visual prompts and log recording. This mechanism allows potential equipment degradation to be directly and automatically transformed into preventative adjustments to control strategies and clear operational guidance, achieving a deep closed-loop integration of equipment health management and process operation control.

[0128] The data acquisition layer connects to the boiler control system, desulfurizer feeding control system, and flue gas monitoring system through various industrial communication protocols.

[0129] The collected data points include boiler load, total air volume, primary air volume and temperature, secondary air volume and temperature, bed temperature, bed pressure, coal feed rate, flue gas parameters at furnace outlet, desulfurizing agent feed flow rate and frequency, conveying air pressure, and sulfur dioxide concentration at the inlet and outlet of the desulfurization tower.

[0130] It should be further explained that the specific construction method of the data acquisition layer is as follows: distributed data acquisition gateways supporting multiple communication protocols are deployed in the industrial field. These gateways are directly connected to the input / output modules of the boiler control system, the desulfurizing agent feeding control system, and the data output ports of the continuous emission monitoring system via hardwiring or fieldbus. The choice of communication protocol is adapted according to the actual situation of the field equipment and system. For example, data exchange with the distributed control system uses the OPC protocol, data reading with the programmable logic controller and smart instruments uses the Modbus protocol, and remote data transmission can be based on the TCP / IP protocol.

[0131] The collected data points comprehensively cover the key state and operational variables involved in the desulfurization process. From the boiler side, data includes instantaneous boiler load, total air volume, primary air volume and temperature, secondary air volume and temperature, bed temperature, bed pressure, and total coal feed. From the flue gas side, data includes actual and converted values ​​of flue gas flow rate, temperature, pressure, oxygen content, and sulfur dioxide content in the furnace outlet flue. From the desulfurizing agent supply side, data includes the operating frequency and current of the limestone feeder, instantaneous and cumulative limestone feed flow rate, conveyor fan outlet air pressure, and the corresponding feed flow rate and frequency of the slaked lime feeding system. From the environmental monitoring side, data includes measured values ​​of sulfur dioxide concentration at the desulfurization tower inlet and chimney inlet. All collected analog and digital signals undergo preliminary signal conditioning, analog-to-digital conversion, and protocol encapsulation through the acquisition gateway to form a unified real-time data stream. This stream is continuously uploaded to the data processing and model layer data buffer via industrial Ethernet, providing a complete and synchronous source data foundation for various upper-layer models.

[0132] The process state prediction module adopts a prediction model based on long short-term memory networks;

[0133] The data processing and model layer also includes a data preprocessing module, which is used to clean, normalize, and perform feature selection operations on the raw data uploaded from the data acquisition layer.

[0134] It should be further explained that the long short-term memory network model used in the process state prediction module has a specific network structure, which includes an input layer, multiple long short-term memory layers, an optional discard layer, and a fully connected output layer; the sequence data received by the input layer is provided by the data preprocessing module, and its time step is preset according to the dynamic characteristics of the process.

[0135] The data preprocessing module follows this process for processing raw data: First, data cleaning is performed, identifying and removing obvious outliers by setting reasonable physical range thresholds and using statistical outlier detection methods. For missing data caused by instrument malfunctions or communication interruptions, interpolation of effective data from nearby time points or data imputation based on operating condition similarity is used to complete the missing data. Next, normalization is performed, mapping variables with different dimensions and numerical ranges to a unified numerical interval through linear or nonlinear transformations to eliminate the influence of dimensions and accelerate model training convergence. Finally, feature selection is performed, using correlation analysis, principal component analysis, or model-based feature importance assessment methods to select a subset of features that have a key impact on sulfur dioxide generation or desulfurization efficiency from all available variables, which will then serve as input variables for the long short-term memory network model.

[0136] This Long Short-Term Memory (LSTM) network model utilizes historical normal operation data, taking a preprocessed feature sequence as input and the corresponding actual sulfur dioxide concentration as the target output. Supervised training is performed using a time-series backpropagation algorithm, with mean squared error used as the loss function and an adaptive moment estimation algorithm to optimize the network weights. The trained model can continuously output predicted values ​​of sulfur dioxide concentration at the desulfurization island inlet or inside the furnace at a specified future time, based on the real-time input feature sequence. These predicted values ​​constitute one of the fundamental state information bases for the system's intelligent control and optimization decisions.

[0137] The system also includes a human-computer interaction layer;

[0138] The human-computer interaction layer is used to display the overall operating status of the system, the prediction results of the process status prediction module, the health index of the equipment health status assessment module, the control strategy generated by the global collaborative optimization unit, and the early warning information generated by the early warning and autonomous fault tolerance unit.

[0139] It should be further explained that the human-machine interaction layer is implemented through an industrial computer or server deployed in the control room. This layer runs dedicated monitoring software, which adopts a model-view-controller design pattern. It establishes real-time data subscription and communication connections with the model layer, intelligent control layer, and early warning and autonomous fault-tolerant unit through an internal data bus and data processing system.

[0140] The main interface of the monitoring software integrates multiple functional views: the system overview view dynamically renders the entire desulfurization process in the form of a process flow diagram, and simultaneously overlays key real-time data from the data acquisition layer (such as bed temperature, flue gas flow, sulfur dioxide concentration) and key predicted values ​​from the process status prediction module; the equipment health view centrally displays the real-time health index and main failure probability of each key piece of equipment in the form of a dashboard or trend curve, and uses color coding to distinguish the status level.

[0141] The Strategy and Optimization View is used to visualize the decision-making process of the global collaborative optimization unit. It can display details of currently effective collaborative control instructions, several candidate strategies evaluated in the digital twin simulation module, and their scores in the reward function (such as efficiency, cost, and health penalty).

[0142] The early warning information view presents all alarms and early warning records generated by the early warning and autonomous fault tolerance unit in the form of a list and pop-up windows. Each record contains a structured timestamp, device location, a brief description of the triggering rule, specific health indicator values ​​and suggested handling measures, and supports association retrieval with historical alarms of the same type.

[0143] Furthermore, this interaction layer provides an access-managed interface, allowing operators to query data and decision logs for any historical time period, confirm and archive non-critical alarms, and, in specific modes, manually review and confirm automatically generated control strategies before execution. Such manual interventions are also recorded by the system and can be used as reference data for subsequent model optimization. All view displays are refreshed at set intervals, ensuring operators can comprehensively, intuitively, and promptly grasp the overall system operating status, intelligent decision-making basis, and equipment maintenance needs.

[0144] This system effectively integrates process status prediction, equipment health assessment, and multi-objective optimization decision-making by constructing an integrated architecture of "data acquisition - multi-model collaborative analysis - global intelligent optimization - closed-loop execution feedback," achieving multi-dimensional collaborative control of the flue gas desulfurization process. It can dynamically coordinate the desulfurizing agent addition strategy inside and outside the furnace based on real-time operating conditions, optimizing the balance between desulfurizing agent consumption and system energy consumption while ensuring that emission indicators meet requirements. Furthermore, through quantitative assessment and integrated decision-making of the operating status of key equipment, it enhances the system's adaptability to load and coal quality fluctuations and improves overall operational stability.

[0145] Furthermore, the system's embedded cross-validation and self-learning mechanisms enhance the reliability of core state perception, while the early warning and autonomous fault-tolerance functions based on equipment health status enable a shift from post-fault handling to pre-fault early warning and strategy adjustment. This reduces the risk of unplanned downtime and performance degradation caused by potential equipment failures, while providing transparent decision-making basis and status visualization, alleviating the burden of manual judgment on operators, and providing technical support for the long-term economic, stable, and reliable operation of the desulfurization system.

[0146] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0147] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A flue gas desulfurization system based on intelligent control, characterized in that, include: The data acquisition layer is used to collect multi-source data related to the desulfurization process in real time. The data processing and modeling layer is used to process, analyze, and model the collected data. And an intelligent control layer, used to generate and execute control strategies based on model output; The intelligent control layer includes a global collaborative optimization unit and at least one local execution unit; The global collaborative optimization unit is used to simulate and evaluate multiple control strategy combinations based on the state information output by the data processing and model layer, under a preset optimization objective, and generate global collaborative control instructions. The local execution unit is used to receive and execute the global collaborative control command to adjust the designated desulfurization process equipment; The data processing and modeling layer includes: The process status prediction module is used to predict the concentration or amount of substances generated at key desulfurization nodes. The equipment health status assessment module is used to assess and output quantitative indicators of the operating status of key equipment based on non-process parameter data of equipment operation; And a strategy verification and correction module, used to cross-verify and correct the key prediction results output by the process state prediction module; When generating the global collaborative control command, the global collaborative optimization unit simultaneously considers the prediction results of the process state prediction module and the quantitative indicators of the operating state output by the equipment health status assessment module.

2. The flue gas desulfurization system based on intelligent control according to claim 1, characterized in that: The global collaborative optimization unit includes a digital twin simulation module and a strategy optimization module; The digital twin simulation module has a built-in dynamic simulation model of the desulfurization system. The strategy optimization module is used in the digital twin simulation module to simulate in parallel multiple future control time domains various control strategies formed by different combinations of control parameters of the local execution unit, and to evaluate the comprehensive benefits of each strategy according to a preset reward function. The reward function integrates at least the desulfurization efficiency index, the material consumption index, and the equipment status penalty item related to the output of the equipment health status assessment module. The strategy optimization module selects a control strategy that meets the overall benefit requirements based on the evaluation results and generates the global collaborative control command.

3. The flue gas desulfurization system based on intelligent control according to claim 2, characterized in that: The equipment health status assessment module includes an equipment feature extraction network and a health quantification model; The device feature extraction network is used to extract features from the historical and real-time runtime sequence data of key equipment, wherein the time sequence data includes at least one of vibration, current, and pressure data. The health measurement model is used to output a health index representing the current health level of the device, as well as the probability of a specific type of failure occurring within a set future time period, based on the extracted features. The global collaborative optimization unit adjusts the control parameter constraint range of the corresponding device or its weight in the reward function based on the health index or the failure probability.

4. The flue gas desulfurization system based on intelligent control according to claim 3, characterized in that: The strategy verification and correction module includes a first prediction unit, a second prediction unit, and a result calibration unit; The first prediction unit uses a first type of algorithm model to predict the sulfur dioxide concentration at the inlet of the desulfurization island and outputs a first predicted value. The second prediction unit uses a second type of algorithm model, different from the first type of algorithm model, to predict the sulfur dioxide concentration of the same object and outputs a second prediction value. The result calibration unit is used to introduce a third type of state information when the deviation between the first predicted value and the second predicted value exceeds a preset threshold, and to perform fusion calibration on the first predicted value and the second predicted value based on the third type of state information, and output a calibrated predicted value with a confidence interval. The third type of state information is a qualitative pattern label obtained by evaluating the combustion state through image recognition technology.

5. The flue gas desulfurization system based on intelligent control according to claim 4, characterized in that: The system also includes a self-learning feedback loop; The self-learning feedback loop is used to continuously compare the actual measured value with the calibration prediction value after the control strategy is executed based on the calibration prediction value; When the comparison error persists and the confidence interval is lower than the set standard, the offline analysis mechanism is triggered. The offline analysis mechanism packages the raw data, intermediate predicted values, calibration process data and final actual values ​​within the time period into special working condition samples, and extracts potential unmodeled features or new rules through anomaly detection and causal analysis algorithms. The analysis results output by the offline analysis mechanism are fed back to the process state prediction module or the strategy verification and correction module for parameter updates or structural fine-tuning of the online model.

6. The flue gas desulfurization system based on intelligent control according to claim 5, characterized in that: The local execution unit includes an in-furnace desulfurizing agent control module and an external desulfurizing agent control module; The in-furnace desulfurizing agent control module is used to adjust the operating parameters of the equipment related to adding limestone into the furnace according to the global collaborative control command; The external desulfurizing agent control module is used to adjust the operating parameters of the equipment related to adding quicklime to the desulfurization tower according to the global collaborative control command; The global collaborative optimization unit, through the digital twin simulation module, collaboratively optimizes the control parameters of the in-furnace desulfurizer control module and the out-of-furnace desulfurizer control module to achieve dynamic load distribution and cost balance between the in-furnace and out-of-furnace desulfurization processes.

7. A flue gas desulfurization system based on intelligent control according to claim 6, characterized in that: The system also includes an early warning and autonomous fault-tolerant unit; The early warning and autonomous fault-tolerant unit is connected to the equipment health status assessment module and the global collaborative optimization unit; The early warning and autonomous fault-tolerant unit is pre-set with a rule library, which defines control constraint adjustment strategies and early warning information corresponding to different quantitative indicators of equipment health status and failure probabilities. When the indicators output by the device health status assessment module trigger the rule base, the early warning and autonomous fault tolerance unit sends a constraint adjustment signal to the global collaborative optimization unit and generates early warning information containing maintenance suggestions.

8. The flue gas desulfurization system based on intelligent control according to claim 7, characterized in that: The data acquisition layer is connected to the boiler control system, desulfurizer feeding control system, and flue gas monitoring system through various industrial communication protocols. The collected data points include boiler load, total air volume, primary air volume and temperature, secondary air volume and temperature, bed temperature, bed pressure, coal feed rate, flue gas parameters at furnace outlet, desulfurizing agent feed flow rate and frequency, conveying air pressure, and sulfur dioxide concentration at the inlet and outlet of the desulfurization tower.

9. A flue gas desulfurization system based on intelligent control according to claim 8, characterized in that: The process state prediction module adopts a prediction model based on long short-term memory networks; The data processing and model layer also includes a data preprocessing module, which is used to clean, normalize, and perform feature selection operations on the raw data uploaded by the data acquisition layer.

10. A flue gas desulfurization system based on intelligent control according to claim 9, characterized in that: The system also includes a human-computer interaction layer; The human-computer interaction layer is used to display the overall operating status of the system, the prediction results of the process status prediction module, the health index of the equipment health status assessment module, the control strategy generated by the global collaborative optimization unit, and the early warning information generated by the early warning and autonomous fault tolerance unit.

Images