Intelligent control system and control method for flue gas environment-friendly island

Abstract

The invention discloses an intelligent control system and control method for a flue gas environment-friendly island, belongs to the technical field of flue gas treatment, and solves the problem of how to improve the working condition adaptability, response speed and prediction precision of the environment-friendly island. According to the technical scheme, through intelligent integrated control over a plurality of subsystems in the denitration, dust removal and desulfurization processes, the working condition adaptability, the response speed and the prediction precision of the environment-friendly island are improved, on the premise that standard emission of pollutants is guaranteed, the comprehensive operation efficiency of all the process links is remarkably improved, energy consumption, material consumption and equipment loss are reduced, and the economic benefit is improved. According to the invention, an edge computing device can be used as an external communication module of a DCS, online robust operation is realized, the method is especially suitable for an environmental protection island SCR system in various flue gas treatment fields, safe, efficient and economical intelligent operation of the flue gas environmental protection island system can be realized on the premise of not changing original control logic, and the system has a wide application prospect. And the method has a wide popularization prospect and a remarkable industrial practical value.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of flue gas treatment, and relates to a flue gas environmental protection island intelligent control system and a control method. BACKGROUND

[0002] With the increasingly stringent environmental protection standards and the increasing requirements of industrial production for energy saving and emission reduction, the flue gas treatment system in the fields of industrial boilers, industrial furnaces, sintering machines, blast furnaces, pelletizing, coking, chemical manufacturing, ships and the like is facing unprecedented operating pressure. The denitration, dust removal and desulfurization processes of the environmental protection island are to reduce the impact of harmful substances (such as nitrogen oxides, sulfur dioxide and dust) discharged by industry on the environment, and through the collaborative work of multiple processes, to ensure that the industrial emissions meet the environmental protection standards and reduce the impact on the environment.

[0003] At present, the control strategy for flue gas treatment in industrial sites mainly relies on manual experience control, a small number of enterprises use basic PID feedback control, and a small number of researches have proposed to introduce intelligent control methods into the collaborative control operation of the environmental protection island. Among them, manual experience control relies on the experience of operating personnel. Due to the huge difference in dynamic characteristics of the control system in different load intervals, it is difficult for fixed control models and parameters to adapt to the full working condition range, and it is difficult to respond to the rapid changes of dynamic conditions under load fluctuations, which is prone to response lag. Although the traditional PID control has automatic capability, due to the frequent changes of dynamic conditions caused by boiler load fluctuations, the traditional PID control is prone to overshoot, oscillation or long regulation time. The existing technology such as the invention patent with the application publication number CN112364562A discloses a flue gas environmental protection island collaborative control method, which uses a machine learning classification algorithm to fit the optimization model and the collaborative control model of each process. However, the optimization target of each sub-model is single, and when facing complex working conditions such as unit load fluctuation, coal quality change and equipment aging, it is still difficult to ensure the response speed and prediction accuracy of the control method, and the dynamic condition adaptability is poor.

[0004] Therefore, in view of the problem that the load fluctuation condition affects the response speed and prediction accuracy of the control system in the prior art, there is an urgent need for a flue gas treatment scheme for the environmental protection island that can adapt to frequent load fluctuation conditions. SUMMARY

[0005] The technical problem to be solved by the present application is how to improve the working condition adaptability, response speed and prediction accuracy of the environmental protection island, and how to improve the comprehensive operation efficiency of each process under the premise of ensuring that the pollutants meet the emission standards, and reduce energy consumption, material consumption and equipment loss.

[0006] The present application solves the above technical problems by the following technical solutions: A smart control system for a flue gas environmental protection island, characterized in that the control system includes a denitrification control system, a dust removal control system, a desulfurization control system, and an induced draft fan control system; The denitrification control system includes a urea pyrolysis control subsystem, an ammonia injection regulation control subsystem, a denitrification reactor temperature field control subsystem, and a dilution fan regulation subsystem. The urea pyrolysis control subsystem is used to optimize the urea injection rate and burner power online, recover waste heat from flue gas, and preheat the urea solution; the ammonia injection regulation control subsystem predicts and adjusts the ammonia injection rate through time-series trend analysis; the denitrification reactor temperature field control subsystem is used to predict the temperature field distribution, outlet NOx concentration, and ammonia slip of the denitrification reactor in future periods; and the dilution fan control subsystem is used to generate the target fan speed and output the optimal dilution air volume for the current operating conditions.

[0007] Furthermore, the dust removal control system includes a bag filter cleaning control subsystem and / or an electrostatic field impedance matching subsystem; The bag filter cleaning control subsystem is used to solve the optimal pulse-jet cleaning strategy for the bag filter in future time periods; the electrostatic field impedance matching subsystem is used to dynamically correct the current and voltage values ​​of the electrostatic precipitator and adjust the high-frequency power supply output.

[0008] Furthermore, the desulfurization control system includes a gypsum dewatering control subsystem, a slurry pH control subsystem, a circulating pump optimization control subsystem, and a slurry supply control subsystem; The gypsum dewatering control subsystem is used to generate the optimal filter cake moisture content and the optimal control command for the vacuum belt filter; the slurry pH optimization control subsystem is used to dynamically track the optimal pH setpoint of the absorption tower; the circulating pump optimization control subsystem is used to generate the circulating pump start-up and shutdown and frequency conversion strategy for the current operating condition; and the slurry supply control subsystem is used to solve the optimal control increment sequence for limestone slurry.

[0009] Furthermore, the induced draft fan control system is used to generate the optimal induced draft fan speed and blade angle.

[0010] Furthermore, the urea pyrolysis control subsystem includes: The first sensing module is used to deploy a multimodal sensing network in the pyrolysis furnace to collect multi-source heterogeneous data in real time. The first fusion module is used to input the collected data into the edge computing node after synchronization by the PLC, construct the CFD-DEM digital twin synchronization data and output the prediction residual, and expand the effective working condition data based on the adversarial generative network. The first decision module is used to continuously calculate the optimal control command of the first decision module based on the online optimization of urea injection quantity and burner power, and switch the control mode of the first decision module according to abnormal conditions. The first execution module is used to receive control commands from the first decision module, adjust the spray gun and burner based on the control commands and dynamic zoning strategy, and actively intervene to remove crystals; The first energy recovery module is used to recover waste heat from flue gas and preheat urea solution through the ORC unit.

[0011] Furthermore, the ammonia injection regulation and control subsystem includes: The second data acquisition module is used to deploy a multimodal sensor network to acquire raw signals from the denitrification reactor; The second data processing module is used to preprocess the raw signal to form a standardized data stream; The second time-series trend analysis module is used to perform weighted smoothing on historical standardized data, capturing the overall trend of signal changes and short-term multi-stage predictions respectively. The second adaptive adjustment module is used to calculate the ammonia injection adjustment amount based on the deviation between the predicted value and the target set value. The second closed-loop feedback module is used to compare the predicted value with the actual measured value in real time and correct the ammonia injection amount. The second execution module is used to control the ammonia injection actuator to complete the ammonia injection based on the corrected ammonia injection quantity. The second model self-learning module is used to update and optimize model parameters online.

[0012] Furthermore, the temperature field control subsystem of the denitrification reactor includes: The third data processing module is used to collect sparse temperature measurement data from the denitrification reactor and perform data preprocessing. The third interpolation and reconstruction module is used to perform spatial interpolation and reconstruction on sparse temperature measurement data, and outputs the reconstructed temperature field, boundary gradient and near-wall mask features. The third model prediction module combines the output of spatial interpolation reconstruction with working condition data, coal type information and actuator status data as input features to construct a directional ConvLSTM spatiotemporal prediction model. The model is trained with a boundary-first masking strategy and outputs the temperature field distribution, outlet NOx concentration and ammonia escape amount for multiple future steps. The third MPC module performs multi-objective model predictive control based on the output of the predictive model, constructs a multi-objective cost function, and optimizes the control quantity of the actuator of the denitrification reactor online. The third control execution module is used to compensate for the time delay of the denitrification reactor actuator and issue the optimized control quantity, set the rollback conditions, and automatically roll back to the original control strategy when the rollback conditions are met and continue to collect data.

[0013] Furthermore, the dilution fan control subsystem includes: The fourth data sensing module is used to collect and transmit denitrification operating parameters in real time, and to preprocess the denitrification operating parameters. The fourth prediction module is used to predict the outlet NOx concentration and optimal dilution air volume in the future period based on the denitrification operation parameters using an improved two-layer LSTM neural network, and to optimize it through a process feature weighting mechanism, a cross-layer attention mechanism and a multi-objective loss function. The fourth decision module is used to output the target speed of the dilution fan based on the prediction results and the current operating conditions using a deep deterministic gradient reinforcement learning algorithm. The fourth safety control module is used to automatically take over control when the decision module outputs an abnormality, employing a fuzzy PID controller. The fourth verification module performs full-condition simulation verification of the fourth prediction module and the fourth decision module based on the digital twin platform.

[0014] Furthermore, the bag filter cleaning control subsystem includes: The fifth data acquisition module is used to collect and monitor the operating parameters of the bag filter in real time. The operating parameters include inlet and outlet pressure difference, flue gas flow rate, dust concentration, temperature and humidity. The fifth dynamic modeling module constructs a mathematical model based on the dynamic equation of filter bag pressure difference, which is used to describe the impact of dust accumulation process and pulse-jet cleaning behavior on the pressure difference change of the internal filter bags in the bag filter. The fifth prediction module, based on the mathematical model and the current operating parameters, predicts the operating status of the bag filter over a period of time in the future. The fifth optimization decision module is used to solve the optimal blowing strategy for a future period of time using a multi-objective optimization algorithm under the premise of meeting preset constraints. The blowing strategy includes blowing pressure, pulse width, blowing interval and timing arrangement. The fifth execution module, based on the rolling time-domain strategy, performs pulse-jet cleaning on the bag filter according to the instructions generated by the optimization decision module, and collects feedback information for optimization in the next cycle. The fifth adaptive update module is used to automatically correct the mathematical model parameters based on long-term running data.

[0015] Furthermore, the electrostatic impedance matching subsystem includes: The sixth data acquisition module is used to collect operating parameters and slurry quality data in real time; The sixth edge computing module is used to perform data cleaning, feature extraction, and operating condition prediction, and generate initial current and voltage setpoints. Specifically, it cleans the collected data and extracts a subset of features that are strongly correlated with current and voltage. Based on the feature subset, it predicts the initial current and voltage setpoints through a generative adversarial network model. The sixth cloud-based optimization module dynamically corrects the initial set value based on the slurry quality index model; The sixth actuator module is used to adjust the high-frequency power supply output according to the corrected current and voltage settings.

[0016] Furthermore, the gypsum dehydration control subsystem includes: The seventh data acquisition and preprocessing module is used to acquire key operating parameters in the gypsum dehydration process in real time and preprocess the key operating parameters. The seventh moisture content prediction module runs the mechanism model and the data-driven model in parallel based on the preprocessed data, and dynamically calculates the fusion weight based on the real-time prediction error of the two models, and obtains the final predicted value of filter cake moisture content by weighted fusion. The seventh multi-objective optimization and MPC control module constructs a weighted multi-objective optimization function based on constraints, uses model predictive control for real-time optimization, and employs a periodic evolutionary algorithm for global multi-objective optimization. Based on the fusion strategy, it generates the final control command and issues it to the corresponding actuator. The seventh control command execution and feedback module is used to execute the issued gypsum dehydration control commands by the actuator, collect feedback data, perform real-time deviation correction, and form a closed-loop control. The seventh adaptive learning module performs online parameter fine-tuning and periodic offline retraining of the data-driven model based on feedback data, and constructs anomaly detection and response strategies.

[0017] Furthermore, the slurry pH control subsystem includes: The eighth data acquisition and preprocessing module is used to acquire and preprocess the absorption tower's operating data; The eighth dual-objective prediction module is used to construct a hybrid prediction model using gradient boosting trees and Gaussian process regression, and output prediction results with uncertainty assessment. The eighth upper-level optimization module is used to search the Pareto optimal frontier using a multi-objective Bayesian optimization algorithm with a hybrid prediction model as the optimization objective function. The eighth lower-level MPC module is used to dynamically track the optimal pH setpoint by the Pareto optimal frontier input adaptive model predictive control execution and solve for the optimal control sequence in a rolling manner. The eighth safety constraint and rollback module is used to set safety constraints and rollback mechanisms. If the safety constraints are not met, the optimal control sequence output by the lower-level MPC module will not be executed. If the safety constraints are met, the first element of the optimal control sequence will be selected and sent to the corresponding actuator. If the rollback mechanism is not met, the control mode will be automatically switched based on a conservative control strategy to ensure the safe and stable operation of the system.

[0018] Furthermore, the circulating pump optimization control subsystem includes: The ninth data acquisition and preprocessing module is used to acquire and preprocess the operating data of the absorption tower. The ninth SO2 prediction module constructs an outlet SO2 concentration prediction model based on a long short-term memory network, and uses the system state within a sliding window as input to make rolling predictions of SO2 concentration in future periods. The ninth hard constraint module, based on the predicted SO2 concentration at the outlet and environmental emission standards, calculates the desulfurization efficiency and reverses the minimum slurry flow rate required to meet the desulfurization efficiency requirements in the future period. The ninth multi-objective optimization module is used to take the slurry flow rate demand as a hard constraint, and the power consumption of the circulating pump group, limestone consumption and the number of start-stop switching as optimization objectives, and to form a comprehensive cost function by weighting them. It uses a genetic algorithm to quickly search the frequency-start-stop combination space and generate the most economical pump start-stop and frequency conversion strategy under the current working conditions in real time. The ninth execution module is used to send the most economical pump start-up and shutdown and frequency conversion strategies to the distributed control system, and to execute the data acquisition and preprocessing module to the multi-objective optimization module in a rolling loop, and to set a rollback mechanism.

[0019] Furthermore, the slurry supply control subsystem includes: The tenth data acquisition and processing module is used to acquire multimodal raw data and preprocess it to generate process variables, derivative features and multimodal features; The tenth soft measurement estimation module is used to perform soft measurement and health estimation on multimodal features and convert them into key process parameters and equipment status indicators. The tenth rolling prediction module is used to predict the prediction sequence and sensitivity of the absorber slurry pH, outlet sulfur dioxide concentration and supersaturation in the future time domain using the absorber mechanism model. The tenth actuator model module is used to build actuator models based on the current actuator state, including slurry flow rate model, pump power model and Jacobian matrix, and to define actuator constraints; The tenth feedforward control module is used to solve the current control deviation, generate the feedforward control quantity based on the sensitivity mapping, and introduce oversaturation suppression and healthy amplitude limiting bias. The tenth multi-objective optimization module is used to construct a multi-objective optimization function with slurry acidity and alkalinity and outlet sulfur dioxide concentration as the core, while taking into account energy consumption, material consumption, supersaturation and equipment health status. Based on actuator constraints, model predictive control is solved to obtain the optimal control increment sequence. The tenth instruction mapping module is used to map the optimal control increment sequence or the feedforward reference quantity when optimization fails to the current control instruction, and to apply protective constraints based on oversaturation and equipment health score before issuing the instruction. The tenth closed-loop feedback and scene memory module performs closed-loop feedback based on the execution results, records operation events, and stores scene memory, storing the current control scene and effect in the memory bank.

[0020] Furthermore, the induced draft fan control system includes: The eleventh data acquisition module is used to collect the operating parameters of boiler and other furnace structure systems in real time. The eleventh data processing module is used to preprocess the collected operating parameters; The eleventh dual prediction module, based on the preprocessed operating parameters, uses a parallel prediction model to perform multi-step prediction of furnace negative pressure and flue gas oxygen content; The eleventh trend analysis module, based on the operation of the nitrogen oxide online meter, performs soft measurement of nitrogen oxide concentration and trend prediction of nitrogen oxide concentration, and obtains the predicted value of nitrogen oxide concentration. The eleventh air volume estimation module, based on the static characteristic model of the fan, calculates the estimated air volume, the Jacobian matrix of air volume to actuator, and the estimated fan power based on the current speed and blade angle. The eleventh feedforward calculation module is used to solve the prediction deviation between furnace negative pressure and flue gas oxygen content, to calculate the engineering feedforward component, and to calculate the engineering feedforward air volume based on the engineering feedforward component. The eleventh multi-objective optimization module is used to construct a quadratic programming problem with negative pressure deviation, oxygen content deviation, nitrogen oxide deviation and fan power as optimization objectives, and actuator limit, speed and surge risk as constraints. The eleventh programming module is used to solve quadratic programming problems and obtain the optimal control increments, including the optimal control increment sequences for the induced draft fan speed and blade angle. The eleventh control rollback module is used to take the first term of the control increment sequence as the current control command if the quadratic programming problem is solved successfully; if the solution fails or timeout occurs, the rollback control command is obtained by mapping the engineering feedforward air volume value with the Jacobian matrix. The eleventh execution module is used to perform security verification on the current control command before sending it to the induced draft fan actuator.

[0021] The present invention also provides an intelligent control method for flue gas environmental protection islands, the control method including establishing a denitrification control model, a dust removal control model, a desulfurization control model, and an induced draft fan control model; The denitrification control model includes a urea pyrolysis control sub-model, an ammonia injection regulation control sub-model, a denitrification reactor temperature field control sub-model, and a dilution fan regulation sub-model. The urea pyrolysis control sub-model is used to optimize the urea injection rate and burner power online, recover waste heat from flue gas, and preheat the urea solution. The ammonia injection regulation and control sub-model is used to predict and regulate the ammonia injection amount through time-series trend analysis. The temperature field control sub-model of the denitrification reactor is used to predict the temperature field distribution, outlet NOx concentration and ammonia slip of the denitrification reactor in future periods. The dilution fan control sub-model is used to generate the target fan speed and output the optimal dilution air volume under the current operating conditions.

[0022] Furthermore, the dust removal control model includes a bag filter cleaning control sub-model and / or an electrostatic field impedance matching sub-model; The bag cleaning control sub-model is used to solve the optimal pulse-jet cleaning strategy for the bag dust collector in future time periods through multi-objective optimization. The electrostatic field impedance matching sub-model is used to dynamically correct the current and voltage values ​​of the electrostatic dust removal equipment and adjust the high-frequency power supply output.

[0023] Furthermore, the desulfurization control model includes a gypsum dewatering control sub-model, a slurry pH control sub-model, a circulating pump optimization control sub-model, and a slurry supply control sub-model; The gypsum dehydration control sub-model is used to generate the optimal filter cake moisture content and the optimal control command for the vacuum belt filter. The pH optimization control sub-model for the slurry is used to dynamically track the optimal pH setpoint of the absorption tower. The circulating pump optimization control sub-model is used to generate the circulating pump start-up, shutdown, and frequency conversion strategies for the current operating conditions; The slurry supply control sub-model is used to solve the optimal control increment sequence of limestone slurry through multi-objective optimization; Furthermore, the induced draft fan control model is used to generate the optimal induced draft fan speed and blade angle.

[0024] Furthermore, the urea pyrolysis control sub-model includes: S201 deploys a multimodal sensor network inside the pyrolysis furnace to collect multi-source heterogeneous data in real time; S202, after the collected data is synchronized by the PLC, it is input into the edge computing node to construct the CFD-DEM digital twin synchronization data and output the prediction residual, and expand the effective working condition data based on the adversarial generative network; S203, based on the online optimization of urea injection quantity and burner power, continuously calculate the optimal control command of the first decision module, and switch the control mode of the first decision module according to abnormal conditions; S204 is used to receive control commands from the first decision module, adjust the spray gun and burner based on the control commands and dynamic zoning strategy, and actively intervene to remove crystals; S205 recovers waste heat from flue gas and preheats urea solution through the ORC unit, achieving a closed-loop energy-saving system.

[0025] Furthermore, the ammonia injection regulation control sub-model includes: S211, Deploy a multimodal sensor network to collect raw signals from the denitrification reactor; S212, preprocesses the raw signal to form a standardized data stream; S213 performs weighted smoothing on historical standardized data to capture the overall trend of signal changes and short-term multi-stage predictions. S214, Calculate the ammonia injection adjustment amount based on the deviation between the predicted value and the target set value; S215, compares the predicted value with the actual measured value in real time and corrects the ammonia injection amount; S216, control the ammonia injection actuator to complete ammonia injection based on the corrected ammonia injection quantity; S217 allows for online updates and optimization of model parameters.

[0026] Furthermore, the temperature field control sub-model of the denitrification reactor includes: S221, Collect sparse temperature measurement data of the denitrification reactor and perform data preprocessing; S222 performs spatial interpolation reconstruction on sparse temperature measurement data, and outputs the reconstructed temperature field, boundary gradient, and near-wall mask features. S223, based on the output of spatial interpolation reconstruction, combines working condition data, coal type information and actuator status data as input features to construct a spatiotemporal prediction model based on directional ConvLSTM. The model is trained with a boundary-first masking strategy and outputs the temperature field distribution, outlet NOx concentration and ammonia escape amount for multiple future steps. S224, based on the output of the prediction model, performs multi-objective model predictive control, constructs a multi-objective cost function, and optimizes the control quantity of the actuator of the denitrification reactor online; S225 performs time delay compensation on the optimized control quantity for the denitrification reactor actuator and issues it, sets back conditions, and automatically reverts to the original control strategy when the back conditions are met and continues to collect data.

[0027] Furthermore, the dilution fan control sub-model includes: S231 collects and transmits denitrification operating parameters in real time, and preprocesses the denitrification operating parameters; S232 employs an improved dual-layer LSTM neural network to predict the outlet NOx concentration and optimal dilution air volume in future periods based on the denitrification operating parameters, and optimizes them through a process feature weighting mechanism, a cross-layer attention mechanism, and a multi-objective loss function. S233 employs a deep deterministic gradient reinforcement learning algorithm to output the target speed of the dilution fan based on the prediction results and the current operating conditions. S234 employs a fuzzy PID controller, which automatically takes over control when the decision module outputs an abnormality. S235, The fourth prediction module and the fourth decision module are simulated and verified under full working conditions based on the digital twin platform.

[0028] Furthermore, the bag filter cleaning control sub-model includes: S301, collects and monitors the operating parameters of the bag filter in real time, including inlet and outlet pressure difference, flue gas flow rate, dust concentration, temperature and humidity; S302 is a mathematical model constructed based on the dynamic equation of filter bag pressure difference, used to describe the influence of dust accumulation process and pulse-jet cleaning behavior on the pressure difference change of internal filter bags in bag dust collectors. S303, based on a mathematical model and the current operating parameters, predicts the operating status of the bag filter over a period of time in the future; S304, Under the premise of satisfying the preset constraints, the optimal blowing strategy for a future period of time is solved by a multi-objective optimization algorithm. The blowing strategy includes blowing pressure, pulse width, blowing interval and timing arrangement. S305, based on the rolling time domain strategy, performs pulse-jet cleaning on the bag filter according to the instructions generated by the optimization decision module, and collects feedback information for optimization in the next cycle; S306 automatically corrects mathematical model parameters based on long-term operating data.

[0029] Furthermore, the electrostatic field impedance matching sub-model includes: S311, real-time acquisition of operating parameters and slurry quality data; S312, perform data cleaning, feature extraction and operating condition prediction, and generate initial current and voltage setpoints; specifically, clean the collected data and extract a subset of features strongly correlated with current and voltage; based on the feature subset, predict the initial current and voltage setpoints through a generative adversarial network model. S313, The initial set value is dynamically corrected based on the slurry quality index model; S314 adjusts the high-frequency power supply output according to the corrected current and voltage settings.

[0030] Furthermore, the gypsum dehydration control sub-model includes: S401, real-time acquisition of key operating parameters during the gypsum dehydration process, and preprocessing of the key operating parameters; S402, based on the preprocessed data, runs the mechanism model and the data-driven model in parallel, and dynamically calculates the fusion weight based on the real-time prediction error of the two models, and obtains the final predicted value of filter cake moisture content by weighted fusion. S403 constructs a weighted multi-objective optimization function based on constraints, uses model predictive control for real-time optimization, and employs a periodic evolutionary algorithm for global multi-objective optimization. Based on the fusion strategy, it generates the final control command and issues it to the corresponding actuator. S404: The actuator executes the issued gypsum dehydration control command, collects feedback data, performs real-time deviation correction, and forms a closed-loop control. S405 performs online parameter fine-tuning and periodic offline retraining of the data-driven model based on feedback data, and constructs anomaly detection and response strategies.

[0031] Furthermore, the slurry pH control sub-model includes: S411, Collect and preprocess the absorption tower operation data; S412 uses gradient boosting trees and Gaussian process regression to construct a hybrid prediction model, outputting prediction results with uncertainty assessment; S413 uses a hybrid prediction model as the optimization objective function and employs a multi-objective Bayesian optimization algorithm to search for the Pareto optimal frontier. S414, the Pareto optimal frontier input adaptive model predictive control execution dynamically tracks the optimal pH setpoint and solves the optimal control sequence in a rolling manner; S415 sets safety constraints and a backoff mechanism. If the safety constraints are not met, the optimal control sequence output by the lower-level MPC module is rejected. If the safety constraints are met, the first element of the optimal control sequence is selected and sent to the corresponding actuator. If the backoff mechanism is not met, the control mode is automatically switched based on a conservative control strategy to ensure the safe and stable operation of the system.

[0032] Furthermore, the circulating pump optimization control sub-model includes: S421, Collect absorption tower operation data and perform preprocessing; S422, based on a long short-term memory network, constructs an outlet SO2 concentration prediction model, using the system state within a sliding window as input, to make rolling predictions of SO2 concentration in future periods; S423, based on the predicted SO2 concentration at the outlet and the environmental emission standards, calculates the desulfurization efficiency and reverses it to calculate the minimum slurry flow rate required to meet the desulfurization efficiency requirements in the future period. S424 uses slurry flow rate demand as a hard constraint and power consumption of circulating pump group, limestone consumption and number of start-stop switching as optimization objectives to form a comprehensive cost function. It uses a genetic algorithm to quickly search the frequency-start-stop combination space and generate the most economical pump start-stop and frequency conversion strategy under the current working conditions in real time. S425 distributes the most economical pump start-up and shutdown and frequency conversion strategies to the distributed control system, continuously executes the data acquisition and preprocessing modules to the multi-objective optimization module, and sets a rollback mechanism.

[0033] Furthermore, the slurry supply control sub-model includes: S431: Collect multimodal raw data and preprocess it to generate process variables, derivative features and multimodal features; S432 performs soft measurement and health estimation on multimodal characteristics, and transforms them into key process parameters and equipment status indicators. S433 uses a rolling prediction sequence and sensitivity of the absorber slurry pH, outlet sulfur dioxide concentration and supersaturation in the future time domain using an absorber mechanism model. S434, establish the actuator model based on the current actuator state, including the slurry flow rate model, pump power model and Jacobian matrix, and define the actuator constraints; S435, solve for the current control deviation, generate the feedforward control quantity based on the sensitivity mapping, and introduce oversaturation suppression and healthy amplitude limiting bias; S436, construct a multi-objective optimization function with slurry acidity / alkalinity and outlet sulfur dioxide concentration as the core, while taking into account energy consumption, material consumption, supersaturation and equipment health status, and solve the model predictive control based on actuator constraints to obtain the optimal control increment sequence; S437 maps the optimal control increment sequence or the feedforward reference quantity when optimization fails to the current control command and applies protective constraints based on oversaturation and equipment health score before issuing the command. S438 performs closed-loop feedback based on the execution results, records operation events, performs scene memory storage, and stores the current control scene and effect in the memory bank.

[0034] Furthermore, the induced draft fan control model includes: S501, real-time acquisition of operating parameters of boiler and other furnace structure systems; S502, preprocesses the collected operating parameters; S503 uses a parallel prediction model to make multi-step predictions of furnace negative pressure and flue gas oxygen content based on preprocessed operating parameters. S504, based on the operation of the nitrogen oxide online meter, performs soft measurement of nitrogen oxide concentration and trend prediction of nitrogen oxide concentration to obtain the predicted value of nitrogen oxide concentration; S505, based on the static characteristic model of the fan, calculates the estimated air volume, the Jacobian matrix of air volume to actuator, and the estimated fan power according to the current speed and blade angle; S506, solves the predicted deviation between furnace negative pressure and flue gas oxygen content, is used to calculate the engineering feedforward component, and calculates the engineering feedforward air volume based on the engineering feedforward component; S507 constructs a quadratic programming problem with negative pressure deviation, oxygen content deviation, nitrogen oxide deviation and fan power as optimization objectives, and actuator limit, speed and surge risk as constraints. S508, solve the quadratic programming problem to obtain the optimal control increment, including the optimal control increment sequence for the induced draft fan speed and blade angle; S509, if the quadratic programming problem is solved successfully, the first term of the control increment sequence is taken as the current control command; if the solution fails or timeout occurs, the backoff control command is obtained by mapping the engineering feedforward air volume value with the Jacobian matrix. S510 performs a security check on the current control command and then sends it to the induced draft fan actuator.

[0035] The advantages of this invention are: (1) The technical solution of the present invention improves the adaptability, response speed and prediction accuracy of the environmental protection island by intelligently integrating and controlling multiple subsystems within the denitrification, dust removal and desulfurization processes. Under the premise of ensuring that pollutants are discharged in compliance with standards, it significantly improves the overall operating efficiency of each process link and reduces energy consumption, material consumption and equipment wear.

[0036] In the denitrification control system, the urea pyrolysis control subsystem first generates operating data and verifies its effectiveness. Based on the extended operating data, a reward function is designed to balance various indicators of the urea pyrolysis process and optimize the urea injection rate and burner power online. The ammonia injection regulation control subsystem improves the response speed and accuracy of ammonia injection control, reduces ammonia slip, and lowers operating costs by analyzing the time-series trend of ammonia injection data and making short-term predictions. The denitrification reactor temperature field control subsystem reconstructs a high-resolution temperature field online based on the collected sparse temperature measurement data, enabling the reconstructed temperature field to maintain good accuracy and robustness under wide load conditions. The dilution fan regulation subsystem, through data-driven prediction and decision-making, outputs the optimal dilution air volume for the current operating conditions in real time under conditions of load fluctuation, coal quality change, and variable ambient temperature and humidity, and stably controls the outlet NOx concentration and ammonia slip concentration within the emission standards.

[0037] In the dust removal control system, the bag filter cleaning control subsystem constructs a mathematical model to quantitatively model the dust accumulation process, filter bag resistance changes, and pulse-jet cleaning effect of the bag filter equipment. Then, through a predictive model, it comprehensively considers multiple input variables to achieve high-precision prediction of the future evolution trend of filter bag pressure difference, effectively overcoming the lag of traditional constant pressure difference or timed control. The electrostatic field impedance matching subsystem generates realistic and physically reasonable operating condition data by inputting multi-scale fused operating condition feature vectors combined with physical constraints, enabling the control system to exhibit higher accuracy and robustness in predicting the operating conditions of electrostatic precipitators.

[0038] In the desulfurization control system, the gypsum dewatering control subsystem predicts the filter cake moisture content in parallel using both predictive and mechanistic models. It employs model predictive control and periodic evolutionary algorithms for global multi-objective optimization, generating a globally optimal control sequence to achieve strong adaptability of the control system to dynamic operating conditions. The slurry pH optimization control subsystem combines gradient boosting trees and Gaussian process regression in its predictive modeling, achieving not only high-precision predictions of operating parameters but also providing uncertainty assessments. This gives pH optimization decisions risk awareness capabilities, improving adaptability to conditions such as coal quality changes and load fluctuations. The circulating pump optimization control subsystem utilizes slurry demand inversion and full... The lifecycle operating cost function integrates pump group power consumption, limestone material consumption, and start-stop switching penalties into a multi-objective optimization framework. A genetic algorithm is used to quickly solve for the optimal start-stop and frequency conversion combination in the rolling time domain, generating the most economical pump start-stop and frequency conversion adjustment commands under the current operating conditions, which are then sent to the actuators. The slurry supply control subsystem uses an absorption tower mechanism model to provide a predictive sequence and sensitivity of the system's future state, representing the coupling relationships between key variables. Under the constraints and guidance of the mechanism model, subsequent slurry supply optimization control is executed, ensuring that the optimization strategy always conforms to physical reality and the desulfurization reaction mechanism, effectively improving the operating condition adaptability and robustness of the desulfurization slurry supply control system.

[0039] In the induced draft fan control system, the furnace negative pressure and flue gas oxygen content are predicted in multiple steps through a parallel predictive model. The sensitivity of the future air volume to the predicted target is calculated in advance. The two are treated as a whole for coordinated prediction and compensation. The required compensation air volume is calculated in advance and the optimizer sends instructions to the induced draft fan actuator to realize the pre-intervention of the induced draft fan outlet air volume. This significantly reduces control lag, shortens the recovery time under varying operating conditions, and improves the operating condition adaptability of the control system.

[0040] (2) The present invention can use edge computing devices as external communication modules of DCS to achieve robust online operation. It is particularly suitable for environmental protection island SCR systems in various flue gas treatment fields. It can achieve safe, efficient and economical intelligent operation of flue gas environmental protection island systems without changing the original process control logic. It has broad promotion prospects and significant industrial practical value.

[0041] This invention is applicable to flue gas treatment scenarios in environmental islands involving industrial boilers (industries such as thermal power generation, gas power generation, biomass power generation, and waste incineration power generation that convert heat energy through fuel combustion), industrial kilns (including but not limited to glass, cement, ceramics, and building materials industries that achieve high-temperature processes such as calcination, smelting, and sintering through fuel combustion or electric heating), sintering machines, blast furnaces, pelletizing plants, coking plants, chemical manufacturing plants, marine power systems, metallurgy, and VOC treatment industries, where one or more of the following processes—desulfurization, denitrification, and dust removal—are involved. The flue gas treatment processes include, but are not limited to: desulfurization processes (such as dry desulfurization, wet desulfurization, and semi-dry desulfurization, with desulfurizing agents including but not limited to calcium-based, sodium-based, and magnesium-based desulfurizing agents); denitrification processes (such as selective non-catalytic reduction (SNCR), selective catalytic reduction (SCR), organic polymer denitrification (PNCR), ammonia denitrification, and ammonia-free denitrification); and dust removal processes, such as electrostatic precipitators, bag filters, and electrostatic-bag composite dust collectors. Attached Figure Description

[0042] Figure 1 This is a schematic diagram of the structure of an intelligent control system for a flue gas environmental protection island according to Embodiment 1 of the present invention; Figure 2 This is a schematic diagram of the execution logic of the urea pyrolysis control subsystem according to Embodiment 1 of the present invention; Figure 3 This is a schematic diagram of the execution logic of the ammonia injection regulation and control subsystem according to Embodiment 1 of the present invention; Figure 4 This is a schematic diagram of the execution logic of the temperature field control subsystem of the denitrification reactor in Embodiment 1 of the present invention; Figure 5 This is a schematic diagram of the execution logic of the dilution fan control subsystem according to Embodiment 1 of the present invention; Figure 6 This is a schematic diagram of the execution logic of the bag cleaning control subsystem according to Embodiment 1 of the present invention; Figure 7 This is a schematic diagram of the execution logic of the electrostatic field impedance matching subsystem in Embodiment 1 of the present invention; Figure 8 This is a schematic diagram of the execution logic of the gypsum dehydration control subsystem according to Embodiment 1 of the present invention; Figure 9 This is a schematic diagram of the execution logic of the slurry pH control subsystem according to Embodiment 1 of the present invention; Figure 10 This is a schematic diagram of the execution logic of the circulating pump optimization control subsystem according to Embodiment 1 of the present invention; Figure 11 This is a schematic diagram of the execution logic of the slurry supply control subsystem according to Embodiment 1 of the present invention; Figure 12This is a schematic diagram of the execution logic of the induced draft fan control system according to Embodiment 1 of the present invention. Detailed Implementation

[0043] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0044] This invention relates to an intelligent control system for a flue gas environmental protection island, comprising multiple control subsystems, each with a complete and independent technical solution. The symbols used in this invention follow these conventions: for clarity, the same symbols may be used to represent different parameters in the descriptions of embodiments of different control subsystems; the meaning of these parameters is limited to their respective embodiments, and the parameters and symbols involved in different control subsystems or different embodiments are unrelated.

[0045] The technical solution of the present invention will be further described below with reference to the accompanying drawings and specific embodiments: Example 1 like Figure 1 As shown, specifically, an intelligent control system for a flue gas environmental protection island is disclosed, which includes a denitrification control system, a dust removal control system, a desulfurization control system, and an induced draft fan control system.

[0046] The denitrification control system includes a urea pyrolysis control subsystem, an ammonia injection regulation control subsystem, a denitrification reactor temperature field control subsystem, and a dilution fan regulation subsystem. The denitrification control system is used to intelligently control various treatment devices and components involved in the flue gas denitrification process to remove nitrogen oxide concentrations from flue gas emissions.

[0047] The urea pyrolysis control subsystem is used to optimize the urea injection rate and burner power online, recover waste heat from flue gas, and preheat the urea solution; the ammonia injection regulation control subsystem predicts and adjusts the ammonia injection rate through time-series trend analysis; the denitrification reactor temperature field control subsystem is used to predict the temperature field distribution, outlet NOx concentration, and ammonia slip of the denitrification reactor in future periods; and the dilution fan control subsystem is used to generate the target fan speed and output the optimal dilution air volume for the current operating conditions.

[0048] The dust removal control system includes a bag filter cleaning control subsystem and / or an electrostatic field impedance matching subsystem; the dust removal control system is used to intelligently control various processing devices and components involved in the flue gas dust removal process to remove solid particulate matter from the flue gas.

[0049] The bag filter cleaning control subsystem is used to solve the optimal pulse-jet cleaning strategy for the bag filter in future time periods; the electrostatic field impedance matching subsystem is used to dynamically correct the current and voltage values ​​of the electrostatic precipitator and adjust the high-frequency power supply output.

[0050] The desulfurization control system includes a gypsum dewatering control subsystem, a slurry pH control subsystem, a circulating pump optimization control subsystem, and a slurry supply control subsystem. The desulfurization control system is used to intelligently control various treatment devices and components involved in the flue gas desulfurization process to remove sulfur dioxide from the flue gas.

[0051] The gypsum dewatering control subsystem is used to generate the optimal filter cake moisture content and the optimal control command for the vacuum belt filter; the slurry pH optimization control subsystem is used to dynamically track the optimal pH setpoint of the absorption tower; the circulating pump optimization control subsystem is used to generate the circulating pump start-up and shutdown and frequency conversion strategy for the current operating condition; and the slurry supply control subsystem is used to solve the optimal control increment sequence for limestone slurry.

[0052] The induced draft fan control system is used to generate the optimal induced draft fan speed and blade angle. The induced draft fan control system is used to maintain negative pressure in the furnace, control flue gas flow, and ensure normal flue gas emission.

[0053] In a preferred embodiment, such as Figure 2 As shown, the urea pyrolysis control subsystem specifically includes: The first sensing module is used to deploy a multimodal sensor network in the pyrolysis furnace to collect multi-source heterogeneous data in real time, including temperature, flue gas velocity, NH3 concentration, HNCO concentration, NOx concentration, and droplet diameter data.

[0054] In this embodiment, the first sensing module is located in the pyrolysis furnace body and collects multi-source heterogeneous data such as temperature field, flow field, chemical field, and atomization quality of the pyrolysis furnace by deploying a multimodal sensor network. Specifically, the temperature field data is obtained by deploying a thermocouple array to collect temperature data of the pyrolysis furnace; the flow field data is obtained by deploying a laser particle velocimeter (PIV) to monitor the flue gas velocity distribution in real time and thus obtain flue gas velocity data; the chemical field data is obtained by deploying a TDLAS laser spectrometer to measure the concentration data of NH3, HNCO, and NOx online; and the atomization quality data is obtained by acquiring droplet images with a high-speed camera and analyzing the Sauter Mean Diameter (SMD) of the droplets according to existing image recognition algorithms to obtain droplet diameter data; wherein, the Sauter Mean Diameter (SMD) is a core parameter characterizing the average particle size of the spray droplets.

[0055] This embodiment provides a hardware deployment method for the first sensing module, specifically: a 32-channel thermocouple array is arranged at equal intervals along the height of the pyrolysis furnace from 0 to 6 meters; a TDLAS laser spectrometer is deployed across the furnace axis for 1.2 meters, with an accuracy of ±1 ppm; an acoustic sensor array is evenly arranged at 8 points around the furnace wall to collect the high-frequency acoustic emission characteristics of crystalline particles impacting the wall surface. All sensors can be fixed to PTFE gaskets via M5 to M12 threads, and the sensor data signals are connected to the PLC via shielded cables. The above hardware deployment method is only a preferred embodiment and can be adapted to meet actual needs.

[0056] Furthermore, the temperature field data can also be obtained by deploying any one of infrared thermal imagers, fiber optic gratings (FBG), wireless SAW sensors, or RTDs (platinum resistance thermometers), requiring only the use of any sensor model that meets the requirements of spatial resolution ≤5cm and response time ≤50ms.

[0057] The first fusion module is used to input the collected data into the edge computing node after synchronization by the PLC, construct synchronized CFD-DEM digital twin data and output prediction residuals, and expand the effective operating condition data based on the generative adversarial network; specifically, the first fusion module includes the following: The first data synchronization unit is used to synchronize multi-source heterogeneous data through the field PLC at a first preset cycle. It uses the OPC UA (OPC Unified Architecture) communication protocol to upload a 35-dimensional vector representing the running data, including 32 channels of temperature, NH3 concentration, HNCO concentration, and NOx inlet concentration, and inputs it to the edge computing node. The first data fusion unit is used to add 8-dimensional acoustic FFT energy spectrum and 2-dimensional PIV velocity field components to a 35-dimensional vector at a second preset period via edge computing nodes. The added vector is packaged and stored in a MySQL time-series database using the MQTT (Message Queuing Telemetry Transport) communication protocol. The MySQL time-series database retains 2 years of data for historical backtracking, data mining, or model training. The acoustic features are acquired by acoustic sensors deployed in S1.

[0058] This embodiment provides a hardware deployment method for the fusion layer, specifically: The Siemens S7-1516-3 PLC is responsible for high-speed I / O synchronization, updating data with a first preset cycle of 4ms to ensure the time consistency of sensor data, and is used for real-time synchronization of temperature field, flow field, chemical field and atomization quality data. The NVIDIA Jetson AGX Orin boasts 275 TOPS of AI computing power and appends multi-dimensional vectors in a second preset cycle of 50ms for running CFD-DEM models and lightweight AI inference. The Jetson AGX Orin is an AI supercomputer released by NVIDIA that can be used for various forms of edge embedded computing with local AI inference latency of less than 10ms.

[0059] In this embodiment, the above hardware deployment method is only a preferred implementation method and can be adapted according to actual needs.

[0060] The first data cleaning unit is used to clean the data by imputing missing values, EMA denoising, and Z-score normalization, and to construct derived features, including the NOx concentration change rate (denoted as ΔNOx / Δt), unit urea consumption, and temperature non-uniformity index. .

[0061] In this embodiment, the missing value imputation can employ linear interpolation or nearest neighbor imputation; the EMA denoising specifically uses the exponential moving average method, which reduces the impact of noise by assigning higher weights to closer data points. In this embodiment, the smoothing factor α is set to 0.2 to make the EMA response slower, resulting in a stronger smoothing effect, effectively denoising and avoiding short-term fluctuations from interfering with long-term trends; the Z-score normalization is used to eliminate the dimensional differences between different features in the imputed and denoised data, improving the stability of the model. Z-score normalization is updated based on a rolling window, and the standard deviation is... and mean Whenever a new set of data arrives, the mean and standard deviation of that set of data are calculated within the rolling window.

[0062] In this embodiment, the constructed derived features include the NOx concentration change rate (denoted as ΔNOx / Δt), unit urea consumption, and temperature non-uniformity index. Derived features are used to transform static data collected by raw sensors into dynamic indicators that can reflect process dynamics, early faults, or optimization potential in real time. This allows the model to capture higher-order nonlinear relationships at a lower dimension, improving prediction accuracy, reducing training sample requirements, and enhancing generalization ability under operating conditions. In this embodiment, calculations are performed in real time at Jetson edge nodes, directly outputting the unit urea consumption and temperature non-uniformity index every 50ms frame.

[0063] Furthermore, this embodiment also defines a temperature uniformity index. The following logic is used to evaluate the temperature uniformity of the pyrolysis furnace cross-section:

[0064] in, Indicates the temperature uniformity index. A value closer to 1 indicates a more uniform temperature across the furnace cross-section. The design goal of this embodiment is to meet this requirement. ; This represents the temperature non-uniformity index, which in this embodiment is used to represent the standard deviation of the furnace cross-sectional temperature. This indicates the average temperature of the furnace cross-section.

[0065] In this embodiment, the standard deviation of the 32 thermocouples deployed on the same cross-section of the furnace should not exceed 5% of the average temperature of the same cross-section, which is equivalent to a relative uniformity of ≥95% when the temperature difference is ≤15K, denoted as .when When the temperature is below 0.95, the system is considered to have uneven temperature. In this case, you can choose to reduce urea injection or increase burner power to promote temperature uniformity. Recovery.

[0066] The first digital twin construction unit is used to construct a digital twin based on the CFD-DEM coupling model as a virtual reactor to simulate the reaction process in the pyrolysis furnace, and output the predicted residual by synchronizing physical field data in real time.

[0067] This embodiment uses computational fluid dynamics (CFD) to simulate the flow, heat transfer, and mass transfer processes of fluids (such as flue gas and urea droplets), and combines it with the discrete element method (DEM) to simulate the motion, collision, and deposition behavior of particles (such as crystalline particles). The CFD provides flow field information to the DEM, and the DEM then feeds back the influence of particle motion on the flow field to achieve coupled influence. This simulates the entire process of urea atomization, evaporation, and pyrolysis in a pyrolysis furnace. Then, by synchronizing physical field data in real time, the predicted residual σ is output. Specifically, the physical field data is the data output by the first data cleaning unit.

[0068] Specifically, the output prediction residual σ is as follows: First, a CFD-DEM coupled model is constructed. Based on the input physical field data, the CFD-DEM coupled model solves the Navier-Stokes equation and the particle momentum equation, and outputs the predicted physical field data with a period of 50ms, including NH3 concentration, NOx outlet concentration, cross-sectional average temperature and crystallization risk index CRI.

[0069] Secondly, the gradient boosting tree (GBDT) is used to fit the nonlinear main trend of the predicted physical field data, and the physical residual Δ is output as the training label, where the physical residual Δ is the difference between the predicted physical field data and the true value.

[0070] Then, Gaussian process regression (GPR) is used to train the model with training labels, outputting the predicted residual σ and the 95% confidence interval. The model bias is corrected based on the predicted residual σ and the uncertainty of the prediction is quantified. When the predicted residual σ > 0.05, the CFD-DEM fine mesh recalculation is triggered, and the updated model version is stored in the model library to ensure the continuous evolution of the digital twin.

[0071] In this embodiment, the physical residual is calculated based on the prediction results output by the CFD-DEM coupled model and used as a training label for machine learning, replacing the training label of traditional experimental data. This enables high-fidelity model training without experiments. At the same time, the data generated by GBDT fitting the CFD-DEM output results can feed back into the CFD-DEM, and the physical residual between the CFD-DEM output results and the true values ​​can train GBDT, enabling the digital twin to have the ability to self-perceive, self-train, self-evaluate, and self-evolve.

[0072] The first load case generation unit is used to generate virtual load case data using WGAN-GP and to verify the load case data using a CFD-DEM coupled model and Wasserstein-1 distance; specifically, it includes the following: The generator subunit is used to build the GAN generator, which fakes working condition data to increase the amount of training data.

[0073] The verification subunit is used to perform preliminary verification of the generated operating condition data. Specifically, the generated operating condition data is input into the CFD-DEM coupled model. If the model output error is ≤5%, the generated operating condition data is considered to have passed the preliminary verification. If the model output error is >5%, the generated operating condition data is considered to have failed the preliminary verification, and the process returns to the generator subunit.

[0074] The discrimination sub-unit is used to verify the consistency between the generated working condition data and the actual data distribution based on the Wasserstein-1 distance. If the verification condition of Wasserstein-1 distance is met, the data is considered consistent. If so, the generated operating condition data is considered valid data.

[0075] In this embodiment, WGAN-GP (Wasserstein GAN with Gradient Penalty) is a form of Generative Adversarial Network (GAN). It introduces the Wasserstein-1 distribution distance metric to measure the shortest path between the real data distribution and the synthetic data distribution. This embodiment verifies the consistency between the generated working condition data and the real data distribution based on the Wasserstein-1 distance, using the following logical representation:

[0076] in, express and Wasserstein-1 distance, Represents the true data distribution. Indicates the distribution of synthetic data. Indicates all and For the joint distribution set of the edges, Indicates the joint distribution. Represents random data pairs. This represents the sum of all data pairs sampled from the joint distribution γ. ,calculate The average value of (x,y) γ represents random data pairs Follows a joint distribution γ, express The L1 vector distance between them.

[0077] The verification condition designed in this embodiment is when... At that time, the operating condition data generated by the GAN generator was considered to be valid data.

[0078] Furthermore, to prevent the generator from "lazily" copying the original data, this embodiment introduces a fingerprint regularization term based on WGAN-GP, and uses the following logic to represent the total loss function of GAN:

[0079] in, This represents the total loss function of GAN. Indicating resistance to loss, Indicates the fingerprint regularization weight. Represents the fingerprint feature vector. This represents the output of the GAN generator. This represents a random noise vector.

[0080] This embodiment introduces the WGAN-GP generative small-sample reinforcement learning framework to address the scarcity of operating condition data in actual industrial production. First, virtual operating condition data is generated by the GAN generator, expanding the original less than 1,000 valid operating condition data to more than 10 times the amount of training data. Then, the validity of the generated operating condition data is initially verified based on the CFD-DEM coupling model. Next, the consistency between the distribution of the generated operating condition data and the real data is verified based on the Wasserstein-1 distance. The quality of the data distribution output by the generator is improved through multiple game iterations, and the verified valid operating condition data is retained in the data pool. This allows for continuous iteration and rapid adaptation in coal-fired units, gas-fired units, and cement kiln tail units.

[0081] The first decision module is used to continuously calculate the optimal control command based on online optimization of urea injection quantity and burner power, and switch the control mode of the first decision module according to abnormal conditions; specifically, the first decision module includes the following: The first optimization unit aims to achieve the highest denitrification efficiency, lowest ammonia slip, lowest energy consumption, and lowest crystallization risk. Based on a multi-objective Bayesian optimizer and using EHVI as the acquisition function, it searches for the Pareto front within the equivalent urea injection rate range of 3.0-8.0 pH. It then optimizes the urea injection rate and burner power online using the PPO reinforcement learning algorithm.

[0082] In this embodiment, multi-objective Bayesian optimization is used to construct the initial Pareto front offline, providing an initial reference point for PPO with a minute-level response; the PPO reinforcement learning algorithm is used for online policy optimization, adjusting the urea injection quantity and burner power according to the real-time status with a second-level response to adapt to transient conditions such as load fluctuations and NOx mutations.

[0083] In this embodiment, the multi-objective Bayesian optimization process is represented by the following logic: I. Generate initial design points using Latin hypercube sampling; II. Construct Gaussian process surrogate models for each objective; Ⅲ. Calculate the EHVI value for each candidate point, and select the point with the largest EHVI for simulation; IV. Update the Pareto frontier and surrogate model, iterating until convergence; V. Output the Pareto optimal solution set, which is used to initialize the PPO policy network and reward function weights.

[0084] In this embodiment, the PPO reinforcement learning algorithm optimizes the urea injection rate and burner power online by designing a reward function, using the following logical representation:

[0085] in, This represents the instantaneous reward value at time t. This represents the weighting coefficient for denitrification efficiency. Indicates denitrification efficiency. This represents the ammonia slip weighting coefficient. The ammonia escape concentration at the outlet is measured in real time by the TDLAS laser ammonia meter at the outlet, read into the control cycle by the PLC, and directly used as one of the input items of the PPO reward function, in ppm. This represents the energy consumption weighting coefficient. This indicates the total power of the system, in kW. This represents the crystallization risk weighting coefficient. Indicates the crystallization risk index. The value ranges from 0 to 1, and to quantify the crystallization tendency 30 seconds in advance, the crystallization risk index is represented by the following logic. :

[0086] in, This indicates the concentration of isocyanate (HNCO concentration) at the furnace outlet at time t, in ppm. This indicates the upper limit of the permissible concentration of isocyanate; in this example, it can be taken as 50 ppm. This represents the lowest furnace wall temperature at time t, in K. This represents the isocyanate dew point temperature, taken as 588K; This represents the Heaviside step function, which takes the value 1 if the condition is met, and 0 otherwise.

[0087] The first control unit performs rolling calculations of the optimal control command based on Model Predictive Control (MPC) constraints. It updates the A, B, and C matrices of the internal prediction model of the MPC using the recursive least squares (RLS) method and outputs a 40-dimensional control vector, including 32 spray gun flow rates and 8 burner power.

[0088] In this embodiment, to enable the model to track changes in the pyrolysis furnace operating conditions as quickly as possible, RLS employs an exponential forgetting weighted objective function. This can be represented using the following logic:

[0089] in, Represents the forgetting factor, taking , express Time-regression vector, This represents the vector of parameters to be estimated, formed by stacking the elements of matrices A, B, and C, which are updated by the recursive least squares method.

[0090] Furthermore, when the prediction residual σ > 0.05 or the system load undergoes a step change, the covariance matrix of the RLS algorithm will be forcibly reset, so that it can quickly discard the influence of old data, prevent old data from continuing to dominate, quickly track new operating conditions, ensure rapid convergence after sudden changes in operating conditions, and improve the adaptability of operating conditions.

[0091] In this embodiment, MPC determines the optimal control sequence by minimizing future errors and control increments. The rolling calculation of the optimal control command utilizes the following logical representation:

[0092] in, Indicates the future starting from time k. Step-optimal control sequence This indicates the prediction time-domain step size; in this embodiment, the time-domain step size is 1 second. Take 30; This represents the model's predicted output vector for the i-th step. This represents the reference trajectory vector for the i-th future step, which is a pre-set value; This represents the output error weight matrix. This represents the control increment vector. This represents the control increment weight matrix. This represents the squared error between the predicted output and the target value, weighted by matrix Q. This represents the squared cost of the control increment vector weighted by matrix E.

[0093] To prevent control actions from exceeding the allowable range of the process, the rolling optimization of MPC must simultaneously satisfy the MPC constraints, as represented by the following logic:

[0094] in, This indicates the predicted ammonia concentration at the outlet in step k, in ppm. This represents the standard deviation of temperature at step k. This represents the temperature at the i-th measurement point in the k-th step. This represents the average temperature at step k. Indicates the number of measurement points. This represents the HNCO concentration at step k, in ppm.

[0095] Furthermore, this embodiment sets the following constraints on the 40-dimensional control vector: The flow rates of all 32 spray guns are between 0.1 and 5 L / min; The power of each of the eight burners is between 10 and 50 kW; ammonia escape from export .

[0096] The first safety unit is when the predicted residual σ > 0.05 or the crystallization risk index. When the pH value is greater than 0.8, an anomaly is considered to exist. The system immediately switches to a conservative PID control mode with a fixed pH of 7.0, and the spray gun flow rate is set to 80% of the flow rate corresponding to the previous response cycle. Once the anomaly is resolved, the first optimization unit control is automatically restored.

[0097] Furthermore, if the predicted residual σ < 0.03 and no anomaly is observed after two response cycles, the anomaly is considered to have been contacted.

[0098] In this embodiment, the first optimization unit aims to achieve the highest denitrification efficiency, lowest ammonia slip, lowest energy consumption, and lowest crystallization risk. Based on the PPO reinforcement learning algorithm, it optimizes the urea injection rate and burner power online by designing a reward function to balance the above indicators, enabling the system to continuously approach the optimal operating point. In the first control unit, the optimal control command is calculated based on MPC constraints, predicting the trend within the next 30 seconds. When model mismatch, crystallization risk, or severe blockage occurs, the first safety unit switches the control mode of the first control unit to the traditional PID control mode. After the anomaly is resolved, it switches back to the first optimization unit control, greatly reducing the risk of abnormal situations, reducing the degree of manual intervention, ensuring the availability of the equipment and long-term safe and stable operation, and reducing economic losses caused by unplanned downtime.

[0099] The first execution module is used to receive control commands, adjust the spray gun and burner based on the control commands and dynamic zoning strategy, and actively intervene to remove crystals.

[0100] In this embodiment, the first execution module first receives a control vector from the first decision module, which includes the spray gun flow rate and the burner power. Based on the control command, the spray gun flow rate and burner power are adjusted. Specifically, the dynamic zoning strategy involves dividing the pyrolysis furnace along its height into multiple independent control zones, each associated with a set of spray guns. When the temperature sensor in a certain independent control zone detects a temperature below 320°C, the burner power of that independent control zone is increased and the spray gun flow rate is reduced to prevent unreacted urea crystallization. When the decision layer predicts that the NOx concentration will increase by 20 mg / Nm³ within the next 10 seconds, the urea flow rate in the corresponding zone is increased in advance. Simultaneously, crystallization is removed in real time using an acoustic soot remover.

[0101] This embodiment provides a hardware deployment method for the first execution module, specifically: 32 independent servo spray guns and 8 proportional burners are configured, wherein the spray gun flow rate range is 0.1~5 L / min and the burner power is 10~50kW; the spray guns are inserted downwards at 15° and the gun head is 450mm from the center of the furnace; at the same time, the sonic soot remover has a power of 500W and a frequency of 20kHz pulsed sonic wave to remove the initial crystallization.

[0102] In this embodiment, the above hardware deployment method is only a preferred implementation method and can be adapted according to actual needs.

[0103] The first energy recovery module is used to recover waste heat from flue gas and preheat urea solution through the ORC unit.

[0104] In this embodiment, an Organic Rankine Cycle (ORC) waste heat power generation system is used to recover the high-temperature flue gas from the pyrolysis furnace outlet. The ORC unit uses R245fa as the working fluid, with an evaporation pressure of 2.8 MPa, a condensation pressure of 0.45 MPa, and a power generation capacity of 260 kW. It absorbs heat and evaporates it into steam to drive a generator to generate electricity. The recovered electrical energy is used to preheat the urea solution to 80°C. In this embodiment, the above process parameters are set for the specific operation and can be selected according to actual needs, not limited to the scope of this embodiment. The flue gas after heat transfer flows into the subsequent dust removal process. The following logic represents the thermal efficiency of the ORC unit in converting flue gas waste heat into electrical energy:

[0105] in, Indicates ORC thermal efficiency. This represents the net power generation of ORC, which is taken as 260kW in this embodiment. This represents the ORC input thermal power.

[0106] This embodiment collects multi-source heterogeneous data in real time based on the deployed sensor network, and inputs it into the edge computing node after synchronization with the PLC. By introducing the WGAN-GP small sample reinforcement learning framework, it addresses the problem of scarce operating condition data in actual industrial production. First, virtual operating condition data is generated by the GAN generator, expanding the original small amount of effective operating condition data to more than 10 times the amount of training data. Then, the effectiveness of the generated operating condition data is initially verified based on the CFD-DEM coupling model. Next, the consistency between the distribution of the generated operating condition data and the real data is verified by the Wasserstein-1 distance. The quality of the data distribution output by the generator is improved through multiple game iterations, and the verified effective operating condition data is retained in the data pool. It can continuously iterate and quickly adapt to coal-fired units, gas-fired units, and cement kiln tail units, effectively improving the operating condition adaptability of the control system.

[0107] Based on the extended effective operating data, this embodiment takes the highest denitrification efficiency, lowest ammonia slip, lowest energy consumption, and lowest crystallization risk as the optimization objectives in the first optimization module. Based on the PPO reinforcement learning algorithm, the urea injection quantity and burner power are optimized online by designing a reward function to balance the above indicators, so that the system can continuously approach the optimal operating point. At the control layer, the optimal control command is calculated based on MPC constraints and the trend in the next 30 seconds is predicted, which effectively improves the control accuracy of the system.

[0108] Example 2 The difference between this embodiment and Embodiment 1 is that, as Figure 3 As shown, the ammonia injection regulation and control subsystem includes: The second data acquisition module is used to deploy a multimodal sensor network to acquire raw signals from the denitrification reactor; In this embodiment, the multimodal sensor network in the second data acquisition module includes NO sensors respectively arranged at the inlet, inside, and outlet of the denitrification reactor. x Concentration sensors, ammonia concentration sensors, flue gas temperature sensors, flue gas velocity sensors, load signal sensors, dust concentration sensors, and various combustion status sensors arranged in the boiler combustion system.

[0109] The second data processing module is used to preprocess the raw signal to form a standardized data stream, including: The second data filtering unit uses exponential smoothing filtering to remove noise and high-frequency interference from the acquired raw signal, resulting in a stationary signal sequence. This can be represented using the following logic:

[0110] in, The original signal, The signal is a stable signal after filtering. This represents the filtered signal value from the previous moment. This is the first smoothing coefficient, and its value ranges from 0 to... <1; The second data denoising unit performs denoising processing, specifically detecting and correcting outliers or transient interference signals in the signal, using the following logic:

[0111] in, The signal value after denoising. This is the length of the median filter window. To replace outliers with the median of the signal within the window, For the noise reduction threshold, when Time and The difference between the filtered signal values ​​at each time step is greater than the noise reduction threshold. When the data is considered abnormal, then... If the signal value after the time-based filtering is processed, then median filtering is performed; otherwise, median filtering is not performed and the original data is retained. The second data normalization unit converts all types of signals to a standard dimension range and normalizes them. Intervals are used to form standardized data sequences. This can be represented using the following logic:

[0112] in, The value is the normalized value. This refers to the minimum value of the corresponding signal type under the sampling history or set operating conditions. This represents the maximum value of the corresponding signal type under the sampling history or set operating conditions.

[0113] The second time-series trend analysis module is used to perform weighted smoothing on historical standardized data, capturing both the overall trend of signal changes and short-term multi-stage predictions, including: The second smoothing unit processes the historical standardized data sequence using a weighted smoothing method. To capture the overall trend of signal changes, the following logical representation is used:

[0114]

[0115] in, The signal value after smoothing. For the first The historical normalized signal value at time L, where L is the historical window length. For the first Dynamic weights at time points, This indicates exponentiation. This is the weight decay coefficient; The second prediction unit performs short-term multi-stage prediction using the smoothed signal, combining the current signal value and the rate of change of the signal value to generate a prediction sequence for the next H time points, represented by the following logic:

[0116]

[0117] in, For the predicted sequence at the next H time points, The smoothed signal value at the current moment is used as... Trend signals at any given moment; The smoothed signal value from the previous time step is used as... Trend signal at time -1; The increment is the trend signal, and K is the number of prediction stages. , The first The first and second coefficients of the stage For the first Dynamic weights for each stage.

[0118] In this embodiment, the overall trend of signal change is first captured by weighted smoothing based on historical time series data. Then, short-term multi-stage prediction is performed by combining the current signal with the rate of change of signals at different times. This enables the system to accurately capture short-term trends, suppress transient fluctuations, and adapt to load fluctuations and flue gas parameter changes in denitrification conditions, providing a stable and reliable prediction basis for adaptive adjustment of ammonia injection.

[0119] The second adaptive adjustment module is used to calculate the ammonia injection adjustment amount based on the deviation between the predicted value and the target set value. Specifically: The target setpoint and prediction sequence within the system The NOx concentration in the sample is compared to obtain the deviation value. The ammonia injection adjustment amount is calculated based on the deviation, using the following logical expression:

[0120] in, This is the adjustment value for ammonia injection rate. The predicted NOx concentration for the next time step. Set a value for the target NOx concentration. This is the proportionality coefficient. These are the differential coefficients. The time derivative represents the rate of change of the deviation.

[0121] The second closed-loop feedback module is used to compare the predicted value with the actual measured value in real time and correct the ammonia injection amount. In this embodiment, the corrected ammonia injection rate is represented using the following logic:

[0122] in, This represents the ammonia injection rate after closed-loop correction. The ammonia injection rate calculated by the adaptive adjustment module. This represents the actual measured NOx concentration. This is the closed-loop correction coefficient.

[0123] The second execution module is used to control the actuator to complete the ammonia injection based on the corrected ammonia injection quantity. The actuator includes an ammonia injection pump and a valve. By receiving the ammonia injection quantity command from the adaptive adjustment module and the closed-loop feedback module, the module converts the control command into a physical operation, adjusting the ammonia injection pump flow rate or the valve opening to ensure that the ammonia injection quantity is accurate to the set value. The ammonia injection quantity output by the execution module is represented by the following logic. :

[0124] in, The final output of the execution module is the amount of ammonia injected. This is the response function for the actuator.

[0125] The second model self-learning module is used to update and optimize model parameters online, including: The second parameterization unit is used to parameterize the prediction model of the time series trend analysis module and the control model of the adaptive adjustment module, and uniformly represent them in the following linear parameterization form:

[0126] in, For model output, Let be the parameter vector to be optimized. , This represents the weighting coefficients at the system input. Indicates the input lag order. This represents the lag coefficient at the system output. Indicates the output lag order. This represents the transpose of the parameter vector to be optimized. for The historical data matrix at any given time consists of historical input signals and their statistical characteristics; In this embodiment, the fixed lag orders n and m in the second parameterization unit cannot accurately describe the dynamic process of the aforementioned changes throughout the entire lifecycle, resulting in insufficient long-term adaptability of the model. Therefore, this embodiment introduces a second lag order adaptive adjustment unit, enabling the system to adjust the lag orders n and m online or periodically based on the current actual dynamic characteristics.

[0127] The second lag order adaptive adjustment unit is used to switch the current model and adjust the model lag order based on the model's predictive performance. Specifically, it continuously monitors the short-term predictive performance of the current model, runs candidate models in the candidate model set in parallel and evaluates them comprehensively, determines the optimal lag order, and smoothly switches models, including: I. Continuously monitor the short-term predictive performance of the current model and generate a candidate model set; wherein, the current model is denoted as... The orders of the model input and output are respectively The short-term prediction performance is obtained by calculating the root mean square error (RMSE) or the mean absolute percentage error (MAPE); the candidate model set is based on the current order. Generate based on, denoted as This indicates that the input and output terminals will be increased or decreased by 0 or 1 respectively based on the current order.

[0128] In this embodiment, when the model prediction performance is continuously lower than a preset threshold, or when the system detects a significant change in operating conditions, the subsequent lag order adjustment process is triggered.

[0129] II. Run candidate models in the candidate model set in parallel and evaluate them comprehensively, specifically as follows: Within a preset time window, each candidate model in the candidate model set runs in parallel, performing an independent RLS algorithm to identify its corresponding optimal parameter vector. .

[0130] After the window period ends, each candidate model is comprehensively evaluated using information criteria. These criteria can be either the Akaike Information Criterion (AIC) or the Bayesian Information Criterion (BIC). An evaluation function is used to balance the goodness of fit and complexity of the model, as expressed in the following logical expression:

[0131] in, This represents the evaluation criterion value of the candidate model. To evaluate the amount of data within the window, P is a penalty factor. In this embodiment, the information criterion tends to select a model that can accurately fit the data without being overly complex.

[0132] III. Decision-making on the optimal order and smooth model switching, specifically: Compare the evaluation criterion values ​​of all candidate models and select the lag order corresponding to the model with the smallest criterion value. This is the optimal order for the new generation.

[0133] The newly determined optimal model and its parameter vector Set as the primary model, the original model The RLS algorithm will switch to standby mode and its main update thread will be based on the new regressor vector. For the optimal model The corresponding optimal Continuous online updates are provided.

[0134] In this embodiment, to achieve a disturbance-free switching model, parameter initialization or data buffering techniques can be used at the switching point to avoid abrupt changes in the model output.

[0135] By introducing the aforementioned adaptive adjustment mechanism for lag order, this invention enables the model to not only update parameters via RLS but also adjust its core structure (lag order), thereby better adapting to the slow or rapid changes in the dynamic characteristics of the system. When facing complex scenarios such as wide-range operating condition fluctuations and equipment aging, the system can automatically maintain high-precision modeling and control effects. This reduces the reliance on engineers manually tuning the model structure and improves the system's intelligence level and engineering application value.

[0136] The second parameter update unit is used to update the gain matrix and covariance matrix, using newly acquired data samples. The parameter estimates are updated using the recursive least squares method. Specifically: First, calculate the RLS gain matrix. :

[0137] Secondly, update the parameter estimates. :

[0138] Then, update the covariance matrix. :

[0139] in, for Gain matrix at time step for The covariance matrix at time t, for The covariance matrix at time -1 for The parameter estimates at time -1 Forgetting factor, take ; The second parameter optimization unit is used to set parameter optimization trigger conditions, including periodic triggering, performance triggering, and operating condition change triggering. The periodic triggering specifically refers to updating parameters once every fixed time step. In this embodiment, each fixed time step can be used. The parameters are updated once every minute or each sampling period as a fixed time step; The performance trigger specifically involves calculating the model prediction error. When the prediction error exceeds a preset threshold, parameter updates are triggered. In this embodiment, the model prediction error can be the root mean square error (RMSE) or the mean absolute error (MAE), and the preset threshold can be the historical average value. The aforementioned triggering of operating condition changes specifically involves monitoring the rate of change or deviation of key process variables. Parameter updates are automatically triggered when any of the following conditions are met: (1) Triggering based on rate of change: The instantaneous rate of change of unit load, flue gas temperature or ammonia concentration exceeds their respective set first threshold. For example, the change in unit load per unit time exceeds ±2% of its rated load, or the change in flue gas temperature per unit time exceeds ±5°C.

[0140] (2) Triggering based on deviation: The deviation between the instantaneous value of the unit load, flue gas temperature or ammonia concentration and its set value or moving average value under the current operating conditions exceeds the second threshold set by each. For example, the absolute value of the deviation between the flue gas temperature and the set temperature exceeds ±10°C.

[0141] (3) Trend-based triggering: When the time series data of key process variables are linearly fitted, if the absolute value of the slope exceeds the set third threshold, it indicates that the variable shows a significant monotonically increasing or decreasing trend.

[0142] The second optimization processing unit optimizes the model based on optimization strategies, including phased parameter optimization, introduction of physical constraints, and parameter update results through exponential smoothing. The parameter optimization is carried out in stages, specifically by using independent RLS to optimize the time series trend analysis module and the adaptive adjustment module respectively. To ensure the physical validity of the optimization results and the safety of the system, multi-level constraints are introduced to comprehensively limit the model's output and internal parameters: (1) Bottom-level static physical constraints: These are used to ensure the executability of the actuator and directly affect the final output of the adaptive adjustment module. The constraints include, but are not limited to: the ammonia injection rate setting value is a non-negative number, the valve opening range is 0% to 100%, and other actuator operating range limitations.

[0143] (2) High-level dynamic coordination constraints: Utilizing the output results of other related modules in the system, dynamic and intelligent constraint boundaries are formed to guide and limit the optimization process, specifically including: Parameter constraints based on the model self-learning module: The parameter confidence intervals (e.g., the main diagonal elements of the parameter covariance matrix) identified online by the model self-learning module are used as constraints. When the parameter values ​​updated by the RLS algorithm exceed the reasonable range in its historical statistical significance, they are limited or penalized to prevent the model from diverging due to short-term data interference.

[0144] Output constraints based on the closed-loop feedback module: The control deviation calculated by the closed-loop feedback module (such as the difference between the actual NOx concentration and the set value) and its changing trend are used as constraints. For example, when the feedback module detects that the NOx concentration is continuously higher than the set value and the deviation is widening, the optimization processing unit can temporarily tighten the upper limit of the output value of the adaptive adjustment module, or force it to output a more aggressive adjustment amount to ensure that the system quickly returns to a stable state.

[0145] Exponential smoothing is used to process parameter update results to avoid drastic control fluctuations caused by parameter jumps. This is represented by the following logic:

[0146] in, For the present The parameter estimate at time t is also the latest parameter value obtained through the RLS algorithm; In the current The smoothing parameter value at time step, For the previous moment ( The smoothing parameter value at time -1. The second smoothing coefficient, 0 < <1.

[0147] The second performance evaluation unit is used to evaluate model performance and adjust it online. Specifically: First, the model performance indicators are calculated, including model prediction error and control effect indicators. The model prediction error includes root mean square error (RMSE) and mean absolute error (MAE). The control effect indicators include denitrification efficiency and ammonia slip rate, which are used to quantify model performance. The ammonia slip rate can be directly obtained by reading the measured value of the ammonia concentration instrument at the outlet of the denitrification reactor. The denitrification efficiency is calculated by the percentage change in NOx concentration at the outlet and inlet of the denitrification reactor relative to the inlet NOx concentration, which is used to measure the purification capacity of the denitrification system.

[0148] Secondly, cross-validation is used to divide historical data into training and validation sets to evaluate the generalization ability of the optimized model on the validation set and avoid overfitting.

[0149] Then, based on the model performance metrics, adjust the parameters online to optimize the frequency or forgetting factor. To optimize the speed and stability of parameter updates; compare the updated model predictions with the actual measurements in real time; if the model performance deteriorates after the parameter update, roll back to the previous set of optimal parameters to ensure stable system operation.

[0150] This embodiment significantly improves the response speed and accuracy of ammonia injection control by introducing time-series trend analysis and short-term prediction mechanisms, reducing ammonia escape and lowering operating costs. Simultaneously, the control system possesses adaptive adjustment capabilities and a closed-loop feedback mechanism, enabling stable operation under load fluctuations or changes in operating conditions, exhibiting excellent intelligence, robustness, and scalability. The introduction of a model self-learning module further enhances the system's adaptive capabilities and long-term operational stability. Through online parameter optimization and model performance evaluation, the system ensures efficient and accurate ammonia injection control even under complex operating conditions such as unit load fluctuations, coal quality changes, and equipment aging.

[0151] Example 3 The difference between this embodiment and Embodiments 1 and 2 is that, as Figure 4 As shown, the temperature field control subsystem of the denitrification reactor includes: The third data processing module is used to collect sparse temperature measurement data from the denitrification reactor and perform data preprocessing, including: Collect sparse temperature data of the denitrification reactor, including temperature, valve position, load, total flue gas volume, outlet NOx concentration, outlet NH3 concentration, and outlet SO2 concentration; perform data preprocessing on the collected sparse temperature data, including data time alignment, data cleaning, and data normalization.

[0152] The third interpolation and reconstruction module is used to perform spatial interpolation and reconstruction on sparse temperature measurement data, and outputs the reconstructed temperature field, boundary gradient, and near-wall mask features, specifically including the following: First, define the wall grid and calibrate the coordinates of the measuring points; Secondly, spatial interpolation reconstruction of sparse temperature measurement data is performed based on the anisotropic Kriging interpolation algorithm, specifically as follows: Constructing anisotropic covariance functions As a kernel function, it is represented using the following logic:

[0153] In the formula, Let be the spatial displacement vector between two points. Indicates to Transpose operation; The signal variance characterizes the overall fluctuation amplitude of the temperature field; To measure the noise variance, This indicates exponentiation. For the Dirac function, An anisotropic metric tensor, used to define the shape and orientation of spatial correlation, is represented by the following logic. :

[0154] In the formula, The relevant lengths along the principal axis, It is a rotation matrix, dominated by the flue gas direction angle. This is determined to align the principal axis of the covariance ellipsoid with the direction of flue gas flow. Kriging estimation can be expressed using the following logic:

[0155] In the formula, Let x be the temperature value to be predicted at location point x. The column vector consisting of temperature observations from all N actual temperature measurement points. , Given the covariance matrix between the measurement points, Points to be interpolated The covariance vector between the known measurement points and the total covariance vector. Next, virtual wall points are set and boundary gradient constraints are applied; The specific steps of setting virtual wall points are as follows: placing virtual wall points at preset intervals along the wall surface, and setting the virtual wall surface temperature. And the uncertainty of the temperature at that virtual point. ;in, This represents the variance of the actual measured temperature values ​​in the near-wall region. for The corresponding weighting coefficients; The boundary gradient constraint specifically refers to: reconstructing the gradient of the temperature field along the wall normal direction. No more than one maximum allowed value This can be represented using the following logic:

[0156] In the formula, Represents the temperature field. Indicates the direction of the wall normal. Indicates the boundary gradient constraint value. This indicates the preset heat transfer limit value; Then, the hyperparameters of the Kriging interpolation algorithm are updated online using maximum likelihood estimation. Record hyperparameters The log-likelihood function is obtained by taking the logarithm of the maximum likelihood estimation function. The log-likelihood function is then expressed using the following logic. :

[0157] In the formula, Representing the covariance matrix, exponential smoothing filtering is used to smooth the hyperparameters. Periodic online updates are performed using the following logical representation:

[0158] In the formula, These are the hyperparameters updated at time t. This is the maximum likelihood estimate at the current moment. This is the smoothing coefficient.

[0159] The third model prediction module, based on the output of spatial interpolation reconstruction combined with operating condition data, coal type information, and actuator status data as input features, constructs a directional ConvLSTM-based spatiotemporal prediction model. The model is trained using a boundary-first masking strategy and outputs the temperature field distribution, outlet NOx concentration, and ammonia slip for future multiple steps. Specifically, the third model prediction module includes the following: First, define the control vector. Coal type indicator vector and working condition vector The input tensor of the prediction model is constructed by combining the output results. ,remember ;in, , , The output consists of a three-channel image sequence comprising the temperature field, boundary gradient field, and near-wall mask of the past K frames; the input tensor is... Total number of input channels Recorded as Where K represents the number of historical frames, Indicates the vector dimension; Secondly, a spatiotemporal prediction model based on directional ConvLSTM is constructed; The prediction model includes a shared encoder and a multi-task decoder. The shared encoder is composed of stacked L layers of directional ConvLSTM units. The multi-task decoder includes a temperature prediction head and an emission prediction head. The temperature prediction head includes several deconvolutional layers or upsampling layers to decode the encoded spatiotemporal features into a temperature field sequence for the next few steps. The emission prediction head includes a global pooling layer and a fully connected layer to regress and predict future NOx concentrations and NH3 escape amounts. Furthermore, directional weights are introduced during the update process of each of the directional ConvLSTM units. , This can be represented using the following logic:

[0160] In the formula, This represents the hidden state at time t. and These represent one-dimensional convolution operations along the x and y directions, respectively, where b is the bias term and the directional weights. , By setting up a small gating network Dynamically generated based on the current operating conditions, using the following logical representation:

[0161] In the formula, For the Sigmoid function, The dominant direction of flue gas.

[0162] Furthermore, the prediction model employs a multi-task loss function based on uncertainty, which is represented by the following logic. :

[0163] In the formula, The predicted losses are for temperature, NOx, and NH3, respectively. These are the uncertainty terms for temperature, NOx, and NH3, respectively.

[0164] Next, during the model training phase, a boundary-first masking strategy is used to handle missing tests and fault points, specifically by using a masking method with a higher probability than the core region. Higher probability Randomly select the near-wall region of the input tensor. The block is masked by setting all feature values ​​within the block to 0, forcing the prediction model to rely on contextual information and boundary features to infer the value of the missing point; Then, a sub-model library is constructed based on working condition clustering, and the optimal sub-model is switched or fine-tuned online according to the current operating status and prediction error. Specifically, the sub-model library is constructed by using a clustering algorithm to cluster the working condition feature vectors in historical operating data, grouping similar working conditions into the same cluster; a dedicated directional ConvLSTM sub-model is trained for each working condition cluster, forming the sub-model library. During online operation, the maximum posterior probability is calculated. , choose to The sub-model with the highest probability To make a prediction, the following logical representation can be used:

[0165] In the formula, Indicates proportional to, Indicates the current input Next choice The posterior probability of each sub-model This indicates the current operating condition characteristics and the j-th cluster center. The distance between them This represents the center vector of the j-th working condition cluster. Adjustable weights are used to balance the similarity of operating conditions and the accuracy of recent predictions; Let be the root mean square error of the j-th sub-model's predictions on recent historical data.

[0166] The third MPC module performs multi-objective model predictive control based on the output of the predictive model, constructs a multi-objective cost function, and optimizes the control quantity of the actuator online. In this embodiment, the multi-objective cost function is represented using the following logic. :

[0167] In the formula, Indicates the temperature field tracking term. , , , , These are the weight coefficients for each optimization objective. Let represent the predicted temperature field at time τ. For reference or target temperature field; This represents the boundary gradient penalty term; Indicates performance loss. Represents the performance loss function; This indicates the number of prediction steps for the temperature field. Indicates the control step size. This indicates the control of incremental penalty terms. Indicates control increment; This indicates the penalty term for the control amplitude. The control amplitude is represented; the temperature field efficiency loss function is represented using the following logic. :

[0168] In the formula, This means taking the average over all i grid points in the space. For the first Regional weights of spatial grid points This represents the temperature gradient at that point. The optimal reaction temperature for SCR is... is the regularization coefficient.

[0169] Furthermore, the aforementioned Adjustments are made by introducing an online adaptive update law for the weights, using the following logical representation:

[0170] In the formula, for Spatial region weights at time points, For the clipping function, through The function limits the updated weights to a preset safety range. Inside, These represent the minimum and maximum values ​​of the spatial region weights, respectively. Forgetting factor, For learning rate, It is the first Several performance metrics.

[0171] Furthermore, the multi-objective cost function The optimization solution applies constraints, including hard constraints and soft constraints; the hard constraints specifically mean that the magnitude and increment of the control quantity are strictly limited to the physically permissible range, which is represented by the following logic:

[0172]

[0173] In the formula, , These represent the maximum and minimum permissible values ​​of the control amplitude, respectively. , These represent the maximum and minimum allowable values ​​for the control quantity increment, respectively; The soft constraints specifically involve constraining key output variables such as predicted ammonia slip and NOx emissions, and are represented by the following logic:

[0174]

[0175] In the formula, This represents the predicted ammonia release at time t+τ. This indicates the maximum permissible amount of ammonia escaping. This represents the predicted NOx concentration at the export point at time t+τ. The emission limit for NOx concentration at the export site.

[0176] The third control execution module is used to compensate for the delay of the optimized control quantity and issue it, set the rollback conditions, and automatically roll back to the original control strategy when the rollback conditions are met and continue to collect data.

[0177] This embodiment considers the actual temperature field under wide load conditions. By fusing virtual wall points and boundary gradient constraints, an anisotropic Kriging interpolation algorithm based on wall constraints is constructed. The temperature at unknown locations is predicted based on the correlation between the collected sparse temperature measurement data. The high-resolution temperature field is reconstructed online, and the boundary gradient field and near-wall region mask features are generated simultaneously. Furthermore, the hyperparameters of the interpolation algorithm are updated online adaptively through maximum likelihood estimation to ensure that the reconstructed temperature field maintains good accuracy and robustness under wide load conditions.

[0178] Example 4 The difference between this embodiment and embodiments one through three is that, as Figure 5 As shown, the dilution fan control subsystem includes: The fourth data sensing module is used to collect and transmit denitrification operating parameters in real time, and to preprocess the denitrification operating parameters. The denitrification operating parameters include flue gas flow rate, SCR inlet NOx concentration, SCR outlet NOx concentration, dilution air temperature, dilution air pressure, ammonia injection flow rate, catalyst layer pressure difference, unit load, and coal quality characteristics.

[0179] The fourth prediction module is used to predict the outlet NOx concentration and optimal dilution air volume in the future period based on the denitrification operation parameters using an improved two-layer LSTM neural network, and to optimize it through a process feature weighting mechanism, a cross-layer attention mechanism and a multi-objective loss function. In this embodiment, the two-layer LSTM neural network includes: The first layer of the LSTM network is used to predict the SCR outlet NOx concentration for future periods based on historical multivariate time series data. The historical multivariate time series data is obtained by collecting historical flue gas flow rate, unit load, coal quality characteristics, dilution air temperature, dilution air pressure, and catalyst layer pressure difference as multivariate time series data. In this embodiment, the time period from the next 30 seconds to 120 seconds is used for prediction to avoid passive response of the denitrification system.

[0180] The second-layer LSTM network is used to calculate the predicted NOx concentration at the SCR outlet. Using current operating data as input, the system outputs the optimal dilution airflow for future time periods. The current operating data includes ammonia injection flow rate, dilution air temperature, dilution air pressure, and catalyst layer pressure difference. In this embodiment, the optimal dilution airflow is output through the second-layer LSTM network. Used to provide reference target values ​​for reinforcement learning decisions in the decision-making module.

[0181] Furthermore, the two-layer LSTM neural network in this embodiment uses the Adam optimizer and sets the initial learning rate to 0.001, the batch size to 256, uses a sliding window as the sample production method, sets the Dropout probability to 0.2, and uses an early stopping strategy to prevent the model from overfitting during training.

[0182] In this embodiment, the process feature weighting mechanism dynamically adjusts the feature weights input to the first-layer LSTM network based on three key process parameters: ammonia-nitrogen molar ratio, catalyst activity coefficient, and flue gas residence time; wherein, the ammonia-nitrogen molar ratio is expressed as NH3 / NO. x The catalyst activity coefficient is calculated based on the operating temperature and usage time, and the flue gas residence time is derived from the flue gas flow rate and flue volume.

[0183] Furthermore, the cross-layer attention mechanism is embedded between the two-layer LSTM networks, adjusting the feature weights of the second-layer LSTM network according to process constraints. The process constraints include the optimal reaction range for the ammonia-nitrogen ratio and the reaction temperature window. When the predicted value deviates from the process feasible solution, the weights of key features are automatically increased to force the model to converge.

[0184] In this embodiment, the optimal reaction range for the ammonia-to-nitrogen ratio can be 1.0~1.2, and the reaction temperature window can be when the reaction temperature is <300℃, NO x The removal rate drops sharply. By transforming the above process mechanism into an attention weight constraint, when the ammonia-nitrogen ratio corresponding to the predicted value of the SCR outlet NOx concentration in the first-layer LSTM network deviates from the optimal range, the attention layer automatically enhances the ammonia injection flow rate, dilution air volume, etc., as input feature weights in the second LSTM network, forcing the model to converge toward the process feasible solution. This can effectively solve the defect in general LSTM models where the predicted value conforms to the data trend but violates physical laws.

[0185] Furthermore, the multi-objective loss function The timing prediction error and process constraint error are used to train the two-layer LSTM network, and their expression is:

[0186] in, The root mean square error (RMSE) is used to measure the accuracy of time series forecasting. This represents process constraint error, used to calculate the deviation between predicted values ​​and theoretical values ​​derived from reaction kinetics. For process constraint weights.

[0187] In this embodiment, a dual-layer Long Short-Term Memory (LSTM) network structure is employed to fully exploit the temporal correlations in the dilution fan control. Since general LSTMs only rely on data temporal correlation modeling, this embodiment overcomes the shortcomings of traditional LSTMs in industrial scenarios—being used only for data-driven applications and lacking practical physical meaning—by basing them on process mechanism constraints and fusing temporal features.

[0188] The fourth decision module is used to output the target speed of the dilution fan based on the prediction results and the current operating conditions using a deep deterministic gradient reinforcement learning algorithm. In this embodiment, the deep deterministic policy gradient reinforcement learning algorithm includes: The state space is expanded by adding specific state variables for the denitrification process to enhance the industrial adaptability of the state description. These specific state variables include ammonia-nitrogen ratio redundancy, catalyst temperature margin, and dilution air mixing uniformity; wherein, ammonia-nitrogen ratio redundancy... This is the ratio of the difference between the actual ammonia-nitrogen ratio and the optimal ammonia-nitrogen ratio to the optimal ammonia-nitrogen ratio, used to reflect the redundancy of the current ammonia injection rate; catalyst temperature margin. The difference between the current reaction temperature and the minimum effective temperature is used to measure the safety margin of the reaction temperature; the uniformity of dilution air mixing is calculated by collecting data from distributed wind speed sensors in the flue to determine the wind speed difference at each point of the ammonia injection grid, which is used to quantify the rationality of the air volume distribution.

[0189] The reward function is designed with a dynamic weight adjustment mechanism to adaptively adjust the NOx exceedance penalty weight and energy consumption penalty weight according to the operational risk level.

[0190] Furthermore, the dynamic weights of the reward function are used to classify operational risk levels based on the differences between the export NOx concentration and the ammonia slip concentration and the limits, including low-risk, medium-risk, and high-risk zones, wherein: High-risk areas will be given priority in ensuring emissions meet standards, and the penalty weight for NOx exceeding standards will be increased; for example, when NO... x >45mg / Nm 3 At that time, the penalty weight for exceeding the NOx limit will be adjusted. The energy consumption penalty weight is increased from the base value of 0.4 to 0.6. The emission standard was reduced from 0.2 to 0.1, prioritizing compliance with emission standards. In low-risk areas, the focus is on energy-saving optimization, with weight values ​​dynamically adjusted using linear interpolation to increase the weight of energy consumption penalties; The weights of medium-risk areas are dynamically adjusted through linear interpolation; for example, when NO x <30mg / Nm 3 And when ammonia slip is <2 ppm, Increased to 0.3, The efficiency was reduced to 0.2, with a focus on energy-saving optimization.

[0191] In this embodiment, the reward function is represented using the following logic. :

[0192] in, , The dynamic weights are assigned to nitrogen oxides and dilution fan power. , The target limits are for nitrogen oxide concentration and ammonia concentration. To dilute the power of the fan, , To dilute the fan speed at time t and time t-1, , The penalty weights are the ammonia concentration and rotational speed of the fan.

[0193] Safety constraints are set for the operating space, including speed change rate constraints based on fan bearing temperature and minimum air volume constraints based on ammonia injection flow rate.

[0194] The speed change rate constraint is specifically: the maximum speed change rate is dynamically adjusted according to the fan bearing temperature; in this embodiment, when the bearing temperature is >65℃, the speed change rate is reduced from 50rpm / s to 30rpm / s to reduce mechanical shock. The minimum air volume constraint is specifically calculated as follows: the minimum dilution air volume is calculated based on the product of the ammonia injection flow rate and the lower limit of the dilution ratio, and then converted into the lower limit of the rotation speed to avoid uneven ammonia mixing due to insufficient air volume; in this embodiment, when the ammonia injection flow rate is 100 kg / h, the minimum rotation speed is not less than 900 rpm.

[0195] The fourth safety control module employs a fuzzy PID controller to automatically take over control when the decision module outputs an abnormality; the fuzzy PID controller includes: A customized fuzzy rule library is generated based on the characteristics of the denitrification process. Specifically, ammonia slip deviation and deviation change rate are used as input variables, and PID incremental parameters are used as output variables. 49 control rules are generated based on membership functions (NB, NM, NS, Z, PS, PM, PB), and 3 new fuzzy rules specific to the denitrification process are added. These specific fuzzy rules are as follows: (1) If the ammonia slip deviation is positive (PB, indicating that the ammonia slip is much higher than the target value) and the deviation change rate is negative (NS, indicating that the deviation is decreasing but at a slow speed), then increase the proportional gain ΔKp (+0.2), increase the integral gain ΔKi (+0.1), and decrease the derivative gain ΔKd (-0.1). The process basis is that although the ammonia slip is high, the trend is slowing down, so it is necessary to strengthen the proportional adjustment to respond quickly, but suppress the derivative to avoid oscillation.

[0196] (2) If the ammonia slip deviation is negative (NB, indicating that the ammonia slip is much lower than the target value) and the deviation change rate is positive (PB, indicating that the deviation is increasing rapidly), then reduce the proportional gain ΔKp (-0.1), reduce the integral gain ΔKi (-0.05), and increase the derivative gain ΔKd (+0.1). The process basis is that the ammonia slip is too low and is still decreasing, so it is necessary to weaken the integral effect to prevent the air volume from being reduced too much and causing NOx to exceed the standard, while increasing the derivative to predict the change.

[0197] (3) If the ammonia slip deviation is zero (Z, indicating that the ammonia slip is near the target value) and the deviation change rate is positive (PB, indicating that the deviation is increasing rapidly), then the proportional gain ΔKp (+0.05), integral gain ΔKi (+0.03), and derivative gain ΔKd (+0.02) are increased slightly. The process basis is that the ammonia slip is currently normal but the upward trend is obvious, and the PID parameters need to be fine-tuned in advance to prevent future exceedances.

[0198] The parameter self-tuning mechanism divides various operating conditions according to unit load, coal sulfur content and ambient temperature, and presets the optimal membership function parameters for each operating condition. When the control error exceeds the threshold, the parameters are automatically switched.

[0199] The fourth verification module performs full-condition simulation verification of the prediction module and the decision module based on the digital twin platform; the digital twin platform adopts a multi-scale coupled simulation architecture, including: Microscopic layer: Based on reaction kinetic equations, the reaction process at the catalyst particle scale is simulated to calculate the catalyst activity distribution; In the meso-level, the flow and mixing processes at the scale of flue gas and ammonia injection grid are simulated using CFD to calculate the velocity field and concentration field distribution. At the macro level, the prediction and decision modules are embedded in MATLAB / Simulink and interact with the CFD model in real time through the OPC interface to simulate the closed-loop process of control strategy and equipment response.

[0200] This embodiment uses data-driven prediction and decision-making to output the optimal dilution air volume in real time under operating conditions such as load fluctuations, coal quality changes, and varying ambient temperature and humidity. It stably controls the outlet NOx concentration and ammonia slip concentration within emission standards. Compared with the traditional PID control mode, it reduces the average power consumption of the dilution fan, while maintaining the smoothness of speed adjustment and reducing mechanical shock. Compared with traditional control, it reduces frequent large-scale speed adjustments and damper opening and closing, reducing mechanical damage. At the same time, it reduces operation and maintenance costs, and reduces the frequency of manual intervention and failure rate.

[0201] Example 5 The difference between this embodiment and embodiments one through four is that, as Figure 6 As shown, the bag filter cleaning control subsystem includes: The fifth data acquisition module is used to collect and monitor the operating parameters of the bag filter in real time. The operating parameters include inlet and outlet pressure difference, flue gas flow rate, dust concentration, temperature and humidity. The fifth dynamic modeling module constructs a mathematical model based on the dynamic equation of filter bag pressure difference, which is used to describe the impact of dust accumulation process and pulse-jet cleaning behavior on the pressure difference change of the internal filter bags in the bag filter. In this embodiment, the mathematical model is represented using the following logic:

[0202] In the formula, For a moment The pressure difference of the filter bag Indicates time The rate of change of pressure difference in the filter bag, For flue gas flow rate, The concentration of dust at the inlet. The dust deposition coefficient is... For jet control input, , Take 0 or 1, This represents the dust removal efficiency coefficient.

[0203] The fifth prediction module, based on the mathematical model and the current operating parameters, predicts the operating status of the bag filter over a future period. Specifically, this module discretizes the time-domain mathematical model constructed in the dynamic modeling module, performs multi-step predictions using the discretized state-space model, and records the discrete time step for each interval. For a given moment, the discretized state-space model can be represented using the following logic:

[0204] in, for The pressure difference of the filter bag at any given time for The pressure difference of the filter bag at any given time Represents the discrete first... The sampling point, the first A moment represents time. ; for The amount of smoke at any given time for The concentration of inlet dust at any given time. for The timing of the blow control input, , This is the natural decay coefficient, which can be ignored or combined. , a) is the natural attenuation factor of the filter bag pressure difference, used to reflect the self-maintaining characteristic of the pressure difference when there is no pulse-jet cleaning; b) is the gain coefficient of dust accumulation, representing the driving force of flue gas flow and dust concentration on the increase of pressure difference; c) is the efficiency factor of pulse-jet cleaning, reflecting the multiplicative attenuation capability of the pulse-jet cleaning action on the current pressure difference. Together, the physical processes of dust accumulation and cleaning are quantified into key parameters in the discrete state equation.

[0205] Furthermore, the discrete state-space model is rewritten in state-space form:

[0206] in, for The state variable at time t, for The state variable at time t, This is used to reflect the clogging status of the filter bag; Let the disturbance input be denoted as . , This indicates the transpose operation. This is the state-space equation.

[0207] The fifth optimization decision module is used to solve the optimal blowing strategy for a future period of time using a multi-objective optimization algorithm under the premise of meeting preset constraints. The blowing strategy includes blowing pressure, pulse width, blowing interval and timing arrangement. In this embodiment, the specific steps of using a multi-objective optimization algorithm to solve for the optimal jetting strategy over a future period are as follows: By solving a finite-time multi-objective optimization problem in each control cycle to predict the system behavior in the next N steps and optimize the control sequence, the multi-objective optimization problem is represented by a model-predicted control objective function:

[0208] in, For differential pressure tracking error, For the target pressure difference, 800 Pa can be used in this embodiment. These are error weight, energy consumption weight, and lifetime weight, respectively. For the future The pressure difference of the filter bag in the step; As an energy consumption item, in this embodiment, energy consumption can be compressed air consumption. For filter bag wear model, For the future The wear life of the step is used to reflect the impact of the jet cleaning on the filter bag under high pressure differential. , For the future Pulse width (blow duration) of the step. For the future The deviation between the step injection pressure and the target value, For the blowing pressure, For pulse width, For the future Step-by-step jet control input.

[0209] Furthermore, the preset constraints include minimizing energy consumption, maximizing dust removal efficiency, extending filter bag lifespan, and ensuring emission compliance. These preset constraints are represented using the following logic:

[0210] in, This indicates the maximum number of pulses per unit time. This indicates the maximum differential pressure limit. These represent the minimum and maximum spray intervals, respectively. This indicates the spray interval within the control cycle.

[0211] The fifth execution module, based on a rolling time-domain strategy, performs pulse-jet cleaning operations on the bag filter according to the instructions generated by the optimization decision module, and collects feedback information for optimization in the next cycle; the rolling time-domain strategy specifically means: at time... Based on the current state And predictive models, to solve for the future Optimal control sequence of steps , obtain new The multi-objective optimization problem needs to be solved again.

[0212] The fifth adaptive update module is used to automatically correct the mathematical model parameters based on long-term running data.

[0213] The fifth adaptive update module updates the dust deposition coefficient using a recursive least squares method. This can be represented using the following logic:

[0214] in, for The latest estimate of the dust deposition coefficient at any given time. for The estimated value of the dust deposition coefficient at time (previous time). Here is the gain matrix. Let be the regression vector, denoted as .

[0215] This embodiment describes the dynamic process of the pressure difference in the internal filter bags of a baghouse dust collector as it changes with dust accumulation and pulse-jet cleaning using differential equations. For the first time, it quantitatively models key physical processes in a baghouse dust collection system, such as dust accumulation, filter bag resistance changes, and pulse-jet cleaning effectiveness, constructing a dynamic prediction model based on a fusion of mechanism and data-driven approaches. This model comprehensively considers multi-dimensional input variables such as flue gas flow rate, dust concentration, temperature, and pressure difference trends, achieving high-precision prediction of the future evolution trend of the filter bag pressure difference. It overcomes the lag inherent in traditional constant pressure difference or time-based control, significantly improving the foresight and adaptability of the control.

[0216] Example 6 The difference between this embodiment and embodiments one through five is that, as Figure 7 As shown, the electrostatic impedance matching subsystem includes: The sixth data acquisition module is used to collect operating parameters and slurry quality data in real time. The operating parameters include sintering machine feed rate, coal feed rate, flue gas flow rate, and inlet dust concentration data. The slurry quality data includes circulating pump power consumption, slurry supply flow rate, and outlet dust concentration data.

[0217] The sixth edge computing module is used to perform data cleaning, feature extraction, and operating condition prediction, and generate initial current and voltage setpoints. Specifically, it cleans the collected data and extracts a subset of features that are strongly correlated with current and voltage. Based on the feature subset, it predicts the initial current and voltage setpoints through a generative adversarial network model. In this embodiment, the sixth edge computing module includes: The sixth data cleaning unit is used to detect anomalies and fill in missing values ​​for the collected operating parameters. Specifically, in this embodiment, the Z-score algorithm is used to identify abnormal data points, the isolated forest algorithm is combined to detect systematic deviations, and the K-nearest neighbor (KNN) interpolation method is used to fill in missing data.

[0218] The sixth feature extraction unit is used to perform dimensionality reduction processing on the cleaned operating parameters using principal component analysis, extract principal component features, and use mutual information method to screen and construct a feature subset strongly correlated with current and voltage; the sixth feature extraction unit includes: The sixth dimensionality reduction subunit is used for dimensionality reduction processing of operating parameters, specifically: First, data standardization is performed on all operating parameters. Then, the covariance matrix is ​​calculated based on the standardized data. Next, eigenvalues ​​and corresponding eigenvectors are obtained from the covariance matrix through eigenvalue decomposition. Finally, the principal components corresponding to the top N largest eigenvalues ​​are selected to form the dimensionality-reduced eigenvectors. The sixth principal component feature extraction subunit, based on the dimensionality-reduced feature vector, obtains a linear combination of principal component features and original operating condition parameters. This linear combination feature reflects the inherent structure of the data, specifically including the following: First, the principal component scores are calculated. Specifically, the standardized data is projected onto the corresponding principal component feature vectors to obtain the coordinates of each sample in the principal component space, which are then used as the principal component scores. Next, the principal component scores of each sample are combined into a feature vector, which serves as the input for subsequent model training. In this embodiment, for flue gas flow rate and inlet dust concentration, a four-dimensional feature vector can be formed by extracting the first two principal component scores.

[0219] The sixth strongly correlated feature screening subunit uses the mutual information method to screen and construct a subset of features strongly correlated with current and voltage. Specifically, it calculates the mutual information value between each principal component feature and the target variable. When the mutual information value is greater than the mutual information threshold, the corresponding principal component feature is considered to be strongly correlated with the target variable. It traverses all principal component features and selects a subset of features strongly correlated with the target variable.

[0220] The sixth operating condition prediction unit is used to run a generative adversarial network (GAN) model, using the feature subset as input, to predict the initial current and voltage setpoints; the GAN model includes: The generator includes an input layer, an encoder module, and a decoder module. It receives the operating condition feature vector through the input layer, fuses the time-series features and static features through the encoder module, and then outputs the predicted current and voltage values ​​through the decoder module. The discriminator consists of an input layer, a feature extraction layer, a physical constraint module, and an output layer. It receives real data through the input layer. With generated data The feature extraction layer improves training stability through activation functions and spectral normalization, and embeds a physical constraint module to determine whether the predicted current and voltage values ​​satisfy the dynamic equivalent resistance relationship. Finally, the discrimination result is output through the output layer. The generative adversarial network model is trained using a hybrid loss function, including adversarial loss, L1 regression loss, and gradient penalty term. The hybrid loss function is represented by the following logic. :

[0221] in, Indicating resistance to loss, Indicates L1 regression loss, Represents the gradient penalty term. , These are the weight coefficients corresponding to the L1 regression loss and the gradient penalty term, respectively. In this embodiment, we can take... , The following logic represents how to counteract losses. :

[0222] in, The expectation operator represents the expectation of the true data distribution. The statistical average; This is a sample of real-world working conditions. This is the discriminator function, used to output the probability that the input data is the true data. This is a noise vector with a dimension of 1×16, used to enhance the generalization ability of the generator. Specifically, it is generated from a standard normal distribution. Random sampling in the middle, The generator function takes a noise vector as input. To output the predicted current I and voltage V.

[0223] In this embodiment, the L1 regression loss is represented using the following logic. :

[0224] in, , The voltage and current values ​​predicted by the generator. , These are voltage and current values ​​from actual operating conditions. It is an L1 norm.

[0225] In this embodiment, the gradient penalty term is represented using the following logic. :

[0226] in, Indicates valid data. , These are random interpolation coefficients.

[0227] The sixth cloud-based optimization module dynamically corrects the initial set value based on the slurry quality index model; the sixth cloud-based optimization module includes: The sixth slurry quality assessment unit is used to construct a slurry quality index model. The slurry quality index is a weighted sum of slurry flow rate, outlet dust concentration, and circulating pump power consumption, and is represented by the following logic:

[0228] in, For slurry flow rate, The dust concentration at the outlet. For the power consumption of the circulating pump, , , They are respectively , , The corresponding weighting coefficients are used in this embodiment to construct an SQI model to balance dust removal efficiency and slurry treatment costs, thereby avoiding energy waste caused by excessive electrostatic precipitators.

[0229] The sixth dynamic correction unit, based on a fuzzy logic controller, calculates the correction amounts for current and voltage according to the deviation between the target value and the measured value of the slurry quality index, using the following logic representation:

[0230]

[0231] in, , These are the proportional and derivative coefficients of the fuzzy logic controller, respectively. , These are the voltage and current correction values, respectively. Indicates the slurry quality index as a function of time. The rate of change.

[0232] The sixth feedforward control unit predicts the flue gas flow rate, dust concentration, and slurry pH value at future times based on the operating condition prediction model. When the predicted flue gas flow rate, dust concentration, and slurry pH value deviate from their respective reference values ​​by more than a preset feedforward threshold, the unit adjusts the current setpoint and voltage setpoint in advance based on feedforward control rules. The adjustment of the current setpoint in advance based on feedforward control rules... Voltage setting value Specifically:

[0233]

[0234] in, , These are the reference current and voltage values, which can be set according to the equipment's rated power and withstand voltage rating. , , These are the baseline dust concentration, slurry pH value, and flue gas flow rate, respectively, taken as stable values ​​under typical operating conditions. , , These are adjustment coefficients corresponding to dust concentration, slurry pH value, and flue gas flow rate, respectively. These can be calibrated experimentally. In this embodiment, we take... , , .

[0235] The sixth feedback control unit is used to monitor slurry density, redox potential, outlet dust concentration and sulfur dioxide concentration in real time; based on the deviation between the monitoring data and the corresponding target values, it calculates the closed-loop correction of current and voltage, and corrects the set value of the feedforward control. In this embodiment, the closed-loop correction amount of current and voltage is calculated based on the deviation between the monitoring data and the corresponding target value. , Specifically:

[0236]

[0237] in, , , , These are the monitored outlet dust concentration, slurry density, sulfur dioxide concentration, and redox potential, respectively. , , , These are the target values ​​for outlet dust concentration, slurry density, sulfur dioxide concentration, and redox potential, respectively. , , , These are correction factors for the corresponding outlet dust concentration, slurry density, sulfur dioxide concentration, and redox potential, respectively. These can be calibrated experimentally; in this embodiment, they are taken as... , , , .

[0238] In this embodiment, the sampling frequency of the feedback control module is set to collect data once every 1 minute, and the correction cycle is to perform closed-loop correction once every 3 minutes.

[0239] The sixth actuator module is used to adjust the high-frequency power supply output according to the corrected current and voltage settings; the high-frequency power supply control uses pulse width modulation (PWM) technology to adjust the power supply output power and directly act on the electrostatic field to realize dynamic adjustment of current and voltage.

[0240] Furthermore, it also includes a sixth equipment health management module, used to allocate pressure coefficients for each electric field, construct a health status index model to quantify the current operating state of the electric field, and design objective functions to optimize the current and voltage setpoints of each electric field, including: First, define the electric field pressure coefficient. Quantify the priority of each electric field in dynamic resource allocation; Secondly, a health status index model is constructed to quantify the current operating state of the electric field, represented by the following logic:

[0241] in, , These are the actual voltage and the rated voltage of the electric field, respectively. , These represent the number of historical failures and the total number of operations, respectively. , These refer to the continuous operating time of the electric field and its design life, respectively. , , These are the weighting coefficients corresponding to electric field voltage, number of runs, and running time, respectively. Then, the weighted least squares method is used to allocate parameters to design the objective function. By minimizing the weighted sum of squared errors, the current and voltage setpoints are optimized, using the following logical expression:

[0242] in, The regularization coefficient is . , The first The actual voltage of the electric field, voltage, , The first The set voltage and current of the electric field.

[0243] This embodiment inputs multi-scale fused operating condition feature vectors into the GAN model and introduces physical constraints into the model structure. By using a hybrid loss function, it balances the realism and physical rationality of the generated data, avoiding the model from getting trapped in local optima and generating unreasonable predictions. In addition, to avoid the generator gradient vanishing due to an overly strong discriminator during training, this invention introduces a gradient penalty term Lgp in the loss function to replace the JS divergence of the traditional GAN ​​model to improve training stability. This results in the control system exhibiting higher accuracy and robustness in predicting the operating conditions of electrostatic precipitators, avoiding lag or overfitting in the prediction results.

[0244] This embodiment provides a feedforward-feedback composite control architecture, which combines operating condition prediction results with real-time slurry quality data to form a closed-loop control link. The feedforward control adjusts the current and voltage setpoints in advance based on the operating condition prediction results, and the feedback control corrects the current and voltage setpoints in a closed loop based on real-time slurry quality and flue gas parameters, thereby correcting the setpoints of the feedforward control in a closed loop. This forms a composite control link that combines feedforward and feedback to ensure stable dust removal efficiency.

[0245] Example 7 The difference between this embodiment and embodiments one through six is ​​that, as Figure 8 As shown, the gypsum dehydration control subsystem includes: The seventh data acquisition and preprocessing module is used to acquire key operating parameters in the gypsum dehydration process in real time and preprocess the key operating parameters, specifically as follows: First, key operating parameters during the gypsum dewatering process are collected in real time, including feed characteristics, process parameters, environmental parameters, and auxiliary parameters. Among them, the feed characteristics include solid content, particle size distribution, and feed temperature; the process parameters include filter cake thickness, vacuum degree, zoned washing water volume, and belt speed; and the environmental parameters include inlet water temperature and washing water pressure.

[0246] Secondly, the key operating parameters collected were normalized. Next, outlier detection and removal are performed on the normalized parameters; Then, the data is smoothed using the sliding window averaging method; Finally, the processed data is stored and synchronized.

[0247] The seventh moisture content prediction module, based on preprocessed data, runs the mechanism model and the data-driven model in parallel, and dynamically calculates the fusion weight based on the real-time prediction errors of the two models. The weighted fusion yields the final predicted value of the filter cake moisture content, including: The seventh mechanistic model unit is used to construct a mechanistic model based on the reaction mechanism of the gypsum dehydration process and output the first predicted value of the filter cake moisture content; the mechanistic model includes the following: First, Darcy's law of percolation was used to determine the filtrate flow rate. Model the filtrate flow rate using the following logic. :

[0248] in, The filter cake permeability coefficient, For filter area, For vacuum degree, The viscosity of the filtrate, The thickness of the filter cake; Secondly, the Kozeny–Carman relation is used to solve for the filter cake permeability coefficient. This can be represented using the following logic:

[0249] in, Let be the particle shape constant. Porosity It is the ratio of the total surface area of ​​the solid particles to the total volume of the solid particles; Next, the filtrate flow rate was measured. The volume of filtrate is obtained by integration. This can be represented using the following logic:

[0250] in, t represents the current time point when the filter cake begins to form; Finally, the moisture content of the filter cake was calculated. Expressed in the form of mass fraction :

[0251] in, , These represent the total volume of water contained in the filter cake and the total volume of dry solid particles in the filter cake, respectively. , The density is that of water and the solid phase.

[0252] The seventh data-driven model unit constructs a data-driven model based on a long short-term memory network and outputs a second predicted value of the filter cake moisture content; the data-driven model includes an input layer, a hidden layer, and an output layer; The input layer is used to receive input feature vectors, including but not limited to multi-dimensional input feature vectors such as temperature, vacuum degree, belt speed, and washing water volume, and to pass the input feature vectors to the hidden layer; The hidden layer includes several LSTM layers, and each LSTM layer contains multiple LSTM units. Each LSTM unit calculates its internal forget gate after receiving input. Input gate Output gate and update its own cell state. and hidden state Each LSTM unit follows the following gating logic:

[0253]

[0254]

[0255]

[0256]

[0257] in, For activation function, For the previous time step t The hidden state of 1 Let be the input vector at time step t. For the previous time step t Cell state 1, It is the hyperbolic tangent function. , , , These are the weight matrices for the corresponding gates. , , , These are all bias terms, which are learnable parameters; The output layer is used to receive the hidden state of the final time step in the last LSTM layer. , merge and map to the second predicted value .

[0258] This embodiment provides a configuration method for a data-driven model, specifically, the hidden layer includes 3 LSTM layers, each LSTM layer contains 100 LSTM units, the data-driven model uses the Adam optimizer, the training data comes from historical running logs, the batch size is 32, and the epoch is 100, which means that the model will adjust its parameters once every 32 samples to the average prediction error generated by the samples, and the model will learn from the entire training dataset for 100 rounds, while supporting an early stopping mechanism to prevent the model from overfitting during training.

[0259] The seventh dynamic fusion unit dynamically calculates the fusion weights based on the real-time prediction errors of the two models, and then performs weighted fusion to obtain the final predicted value of the filter cake moisture content, specifically as follows: First, the real-time prediction errors of the computational model and the data-driven model are calculated separately, and represented using the following logic:

[0260] in, The prediction error of the mechanistic model, For the prediction error of the data-driven model, This is the first predicted value output by the mechanistic model. This is the second predicted value output by the data-driven model. To obtain the true moisture content measured online, The number of samples; Secondly, the fusion weights are dynamically calculated using the following logical representation:

[0261] Finally, the weighted fusion yields the final predicted value of the filter cake moisture content, expressed using the following logic:

[0262] in, This is the final predicted value for the moisture content of the filter cake.

[0263] The seventh multi-objective optimization and MPC control module constructs a weighted multi-objective optimization function based on constraints, uses model predictive control for real-time optimization, and employs a periodic evolutionary algorithm for global multi-objective optimization. Based on a fusion strategy, it generates final control commands and issues them to the actuators, including: The seventh constraint unit is used to define constraints, with washing water consumption, energy consumption, moisture content deviation, and flux as optimization objectives. After normalizing the above optimization objectives, a weighted multi-objective optimization function is constructed. This can be represented using the following logic:

[0264] in, This represents the actual amount of washing water consumed. This is a normalized reference value for washing water consumption. This represents the total energy consumption of the actual system. This serves as a normalized reference value for energy consumption. The target value for moisture content, This represents the actual gypsum treatment throughput. This is the upper limit of flux. , , and The weights for washing water consumption, energy consumption, moisture content deviation, and flux are respectively used; the constraints are represented using the following logic:

[0265] in, For filter cake thickness, , These represent the maximum and minimum values ​​of the filter cake thickness, respectively. For the first Water volume for zoned washing , These represent the maximum and minimum washing water volume, respectively. For vacuum degree, , These represent the maximum and minimum vacuum levels, respectively. For belt speed, , These represent the maximum and minimum belt speeds, respectively. The seventh rolling optimization unit is used to solve the MPC optimization problem using a nonlinear fusion prediction model and perform rolling optimization. Specifically, it establishes a nonlinear fusion prediction model based on the mechanism model and the data-driven model to predict the filter cake moisture content in a rolling manner, and outputs the predicted filter cake moisture content curve, using the following logical representation:

[0266] in, This is the predicted value of filter cake moisture content to be output at the nth time in the future. Indicates the current moment. As a mechanistic model, For the mechanism model function, For data-driven model functions, For data-driven models, This is the control sequence corresponding to the interval from the current time to the nth time in the future. , These are the internal state vectors of the mechanistic model and the deep learning model, respectively. For dynamic fusion weights; The MPC optimization problem Using the following logical representation:

[0267] in, This represents the prediction output of the nonlinear fusion prediction model at the nth time in the future. for Reference value, This indicates the transpose operation. , Both are weight matrices. This represents the control increment at the nth time in the future; The seventh multi-objective optimization unit is used to perform global multi-objective optimization using a periodic evolutionary algorithm to solve for the Pareto optimal solution set of the control variables. Specifically: Population initialization involves randomly generating a set of control variables to initialize the parent population. A multi-objective function is defined to evaluate washing water consumption, energy consumption, moisture content deviation, and flux. Non-dominated ordination is performed by stratifying the population according to Pareto dominance relationships; if individual A is superior to individual B on all objectives, then A dominates B. All non-dominated individuals are identified, forming the first non-dominated stratum. These individuals are then removed, and non-dominated individuals are identified from the remaining individuals to form the second non-dominated stratum, and so on. Crowding calculation involves calculating the crowding distance for individuals within the same non-dominated stratum. To measure its distribution density in the target space; for a single individual, its crowding distance is the normalized value of the sum of the distances between two adjacent individuals on each objective function; genetic operations, specifically generating offspring through simulated binary crossover and polynomial mutation; based on the elite retention strategy, merging the parent and offspring populations, and performing non-dominated sorting and crowding calculation on the merged population again; according to the non-dominated level and crowding degree, selecting individuals with the same number as the initial parent population from high to low to form a new parent population; iterative selection, specifically repeating S333-S336 until the iteration converges.

[0268] The seventh instruction issuing unit generates the final control instruction based on the fusion strategy and issues it to the actuator; the final control instruction includes a control sequence consisting of the target filter cake thickness, the washing water volume of each zone, and the belt speed. The method of generating the final control command based on the fusion strategy is as follows: Model predictive control is used to generate the initial control quantity in real time according to the prediction results, and a periodic evolutionary algorithm is used to provide the Pareto optimal solution set as the initial solution of the MPC optimization problem.

[0269] The seventh control command execution and feedback module is used to execute the issued control commands by the actuator, collect feedback data, perform real-time deviation correction, and form closed-loop control; in this embodiment, the real-time deviation correction specifically involves: First, calculate the execution deviation. :

[0270] in, For the actual output of the implementing agency, Output the target value. If... Then, a proportional correction is applied, expressed using the following logic:

[0271] in, The revised control commands, The original control command, The proportionality coefficient is used in this embodiment, and can be selected based on the system response tuning. It is 0.5. To preset the correction threshold, this embodiment can take... The real-time deviation correction is 2%, which serves as a rapid feedback correction for model predictive control and can effectively prevent the accumulation of small disturbances during the model predictive control process.

[0272] The seventh adaptive learning module performs online parameter fine-tuning and periodic offline retraining of the data-driven model based on feedback data, and constructs anomaly detection and response strategies; specifically, the online parameter fine-tuning and periodic offline retraining of the data-driven model based on feedback data involves: First, define the triggering conditions for the adaptive learning mechanism; Secondly, the parameters of the data-driven model are updated online using an online stochastic gradient descent algorithm, represented by the following logic:

[0273] in, These are the model parameters before the update. For the updated model parameters, This represents the loss function under the current model parameters. This is the most recent batch of data; Finally, the data-driven model is periodically retrained offline. Specifically, based on the full dataset, the data-driven model is retrained offline weekly using the Adam optimizer, and the online model is replaced after passing the validation set test.

[0274] This embodiment aims to improve dehydration efficiency, reduce water and energy consumption, and enhance operational adaptability. It employs nonlinear MPC rolling optimization to solve the MPC optimization problem. Simultaneously, the NSGA-II algorithm is used to solve for the Pareto optimal solution set of the control variables, periodically optimizing the control variables directly to generate a globally optimal control sequence, supplementing the local optimization capabilities of MPC. Based on the optimization results, control commands are generated, executed by various devices, and feedback values ​​from these devices are collected and initially corrected, forming a complete closed-loop control chain.

[0275] Example 8 The difference between this embodiment and embodiments one through seven is that, as Figure 9 As shown, the slurry pH optimization control subsystem includes: The eighth data acquisition and preprocessing module is used to acquire and preprocess the absorption tower operation data. The absorption tower operation data includes inlet SO2 concentration, outlet SO2 concentration, slurry pH value, circulating slurry flow rate, flue gas flow rate, limestone slurry concentration, absorption tower liquid level, and slurry temperature. The preprocessing includes missing value processing, noise reduction, and normalization.

[0276] The eighth dual-objective prediction module is used to construct a hybrid prediction model using gradient boosting trees and Gaussian process regression, and outputs prediction results with uncertainty assessment; the eighth dual-objective prediction module includes: The eighth basic model unit is based on a gradient boosting tree to build a basic model, including a desulfurization efficiency prediction model and a limestone consumption prediction model, and outputs the predicted values ​​of desulfurization efficiency and limestone consumption. In this embodiment, two independent GBDT models are first constructed using historical operating data to predict desulfurization efficiency α and limestone unit consumption β, respectively. Specifically, the input feature vector X is defined to include pre-processed operating parameters such as inlet SO2 concentration, flue gas flow rate, and slurry pH value. Based on GBDT, a desulfurization efficiency prediction model and a limestone consumption prediction model are constructed to fit the main trend of the dual objective function, using the following logical representation:

[0277]

[0278] Where X represents the input feature vector, which is a set of values ​​describing the current operating conditions. Each component within it has an independent physical unit, such as: slurry pH value (dimensionless), inlet SO2 concentration (unit: mg / m³). 3 ), flue gas flow rate (unit: m) 3 / h) etc.; and These represent the trained GBDT models used to predict desulfurization efficiency and limestone consumption, respectively. The predicted desulfurization efficiency value is expressed in units of %. The predicted unit consumption value of limestone is expressed in kg / h.

[0279] The eighth residual unit is used to calculate the prediction residuals of the base model and construct the residual dataset; First, using the trained GBDT model and The historical dataset used at that time Calculate the prediction residual for each historical sample point. The prediction residual is defined as the difference between the actual value and the predicted value, expressed using the following logic:

[0280]

[0281] in, and These represent the actual desulfurization efficiency (in %) and limestone consumption (in kg / h) for the i-th historical sample, respectively. and These are the predicted values ​​of desulfurization efficiency and limestone unit consumption for the i-th historical sample, respectively. and These are the desulfurization efficiency residual (in %) and limestone consumption residual (in kg / h) of the i-th historical sample, respectively.

[0282] Furthermore, the above formula was used to generate two residual datasets for training the GPR residual model. and These two datasets reflect systematic biases and random noise related to operating conditions (input feature vector X) that the GBDT-based model failed to capture.

[0283] The eighth prediction distribution unit constructs a residual model based on Gaussian process regression and outputs a prediction distribution. The residual model includes a desulfurization efficiency residual model and a limestone consumption residual model. The prediction distribution includes the residual mean and residual variance predicted by the residual model. In this embodiment, two independent GPR models are trained using the two residual datasets constructed in S22, respectively. and , represent the GPR models for desulfurization efficiency and limestone consumption, respectively. The core advantage of the GPR model is that it can not only predict the expected value of the residuals, but also represent the confidence level of the prediction results through variance, for any new input condition. The GPR model outputs a value related to the residuals. The posterior prediction distribution is given by a Gaussian distribution:

[0284]

[0285] in, and These respectively represent the new operating conditions Probabilistic prediction of desulfurization efficiency residuals and limestone consumption residuals, where N(...) represents a Gaussian distribution. and They are respectively and The mean of the predicted residuals is used to correct the prediction bias of GBDT, and the units are % and kg / h, respectively; and They are respectively and The predicted residual variance is used to quantify the uncertainty of the prediction, and its unit is % (percentage). 2 and (kg / h) 2 .

[0286] The eighth hybrid prediction unit, by superimposing the basic model and the residual model, constructs a hybrid prediction model. Given an input feature vector, it outputs predictions for desulfurization efficiency and limestone consumption. The hybrid prediction model includes a hybrid prediction model for desulfurization efficiency. Hybrid prediction model for limestone consumption The prediction results include the prediction mean and the prediction variance; In this embodiment, the predicted values ​​of the GBDT basic model and the predicted distribution of the GPR residual model are linearly superimposed to form the final hybrid prediction model. For a given complete input feature vector X containing a specific slurry pH value, the predicted mean and predicted variance of the desulfurization efficiency hybrid prediction model fa(X) are represented by the following logic:

[0287]

[0288] in, and Let represent the mean and variance of the desulfurization efficiency prediction model for input X, respectively. This represents the predicted value of the basic GBDT model for input X. This represents the residual mean prediction of the input X by the residual GPR model. This represents the residual variance prediction of the residual GPR model for the input X.

[0289] The following logic represents the mixed prediction model for limestone consumption. Output of predicted mean and predicted variance:

[0290]

[0291] in, and Limestone consumption mixed prediction model Given an input feature vector X, the output is the predicted mean and predicted variance. This represents the predicted limestone consumption trend of the basic GBDT model for input X. This represents the mean residual of limestone consumption predicted by the residual GPR model for input X. This represents the variance prediction of the residual GPR model for limestone consumption of input X, i.e., the uncertainty of the final prediction.

[0292] The eighth upper-level optimization module is used to search the Pareto optimal frontier using a hybrid prediction model as the optimization objective function and a multi-objective Bayesian optimization algorithm, including: The eighth dual-objective function unit is used to combine the desulfurization efficiency prediction model. Hybrid prediction model for limestone consumption As a biobjective function, it can be represented logically as follows: ,

[0293] in, This can be expressed as the objective function for desulfurization efficiency, with the optimization goal being to maximize the desulfurization efficiency value. This is expressed as the objective function for limestone consumption, and its optimization objective is to maximize the unit consumption of limestone. The prediction results of the biobjective function follow a Gaussian distribution, which can be represented by the following logic:

[0294] in, Indicates the first One objective function, ; Indicates a Gaussian distribution. This represents the predicted mean. Indicates the prediction variance; The eighth optimization unit, based on the desired hypervolume improvement to obtain the function, is within the pH constraint range. The inner iterative search is performed on the Pareto optimal solution set; where... , These are the upper and lower limits of the preset pH range, respectively.

[0295] Furthermore, the search for the optimal pH setting specifically involves: Define the currently collected running samples as the input data for EHVI, denoted as . { } Among them, index Indicates the first One historical data sampling point, This represents the pH setpoint corresponding to that sampling point. and These are respectively represented at the sampling points The actual measured desulfurization efficiency and negative limestone consumption; Based on all currently observed sample points, the current Pareto optimal front is extracted using a non-dominated sorting algorithm. ; Preset pH range Internally sample candidate pH values ​​to generate a candidate point set; Calculate the predicted distribution for each candidate point, specifically by calculating the objective function corresponding to each pH candidate value. The predicted mean and predicted variance; For each candidate point, calculate its relationship with the current Pareto front. Potential increase in supervolume The super-volume increment Defined as the hypervolume dominated by the new Pareto front after adding the new points predicted by pH to the current observation set, minus The supervolume it governs; Integrating the predicted distribution yields the EHVI value, which can be represented logically as follows:

[0296] in, This represents the hypervolume increment of the candidate point relative to the current Pareto front. Represents the mathematical expectation; The pH value corresponding to the maximum EHVI value is selected as the optimal pH setting, expressed using the following logic:

[0297] in, This indicates the optimal pH setting.

[0298] The eighth lower-level MPC module is used to dynamically track the optimal pH setpoint by executing the Pareto optimal frontier input adaptive model predictive control, and to solve for the optimal control sequence in a rolling manner, including: The eighth optimal setting unit selects the current optimal setpoint pH from the Pareto optimal frontier based on the plant's current operating strategy. * Specifically, a weighted comprehensive utility function is introduced to... The optimal set point pH was obtained by evaluating and screening the alternative solutions. * This can be represented using the following logic:

[0299] in, This represents the set of Pareto optimal solutions. This represents the unique optimal pH setting selected through decision-making. and These represent the normalized desulfurization efficiency and normalized limestone consumption corresponding to a specific pH value in the Pareto solution set, respectively. and These are the weighting coefficients for environmental benefits and economic costs, respectively. ; The eighth local dynamic prediction unit establishes and identifies a local dynamic prediction model around the set point pH*; specifically, an independent linear time-varying state-space model is used to describe the local dynamic prediction model based on the limestone slurry dosage. As a control input and slurry pH value The dynamic relationship between the controlled outputs is represented by the following logic:

[0300]

[0301] in, , Let these represent the system state vectors at times k and k+1, respectively. This represents the control input at time k. This represents the system measurement output at time k. , , Let these represent the system's state transition matrix, input matrix, and output matrix, respectively. Indicates process noise. Indicates measurement noise; The eighth online update unit uses online system identification to perform online adaptive updates of the local dynamic prediction model; specifically, the recursive least squares method is used for online adaptive updates of the local dynamic prediction model, using the following logical representation:

[0302]

[0303]

[0304] in, for Gain matrix at time step for The covariance matrix at time t, for The covariance matrix at time -1 for The regression vector at time step, Forgetting factor, take ; This is the model parameter vector, which includes the state transition matrix. Control input matrix Element; The eighth rolling optimization unit solves for the optimal control sequence based on rolling time-domain optimization. Specifically, it finds the optimal future control action sequence by solving a constrained optimization problem, which is defined as minimizing the cost function. J This can be represented using the following logic:

[0305] in, It is the cost function to be minimized. This represents a prediction of the output slurry pH value at time k+i based on information from time k. Indicates a time index. The weight matrix represents the output bias. The weight matrix represents the control increment and reflects the requirements for the smoothness of the control action; Indicates the future Control increment at any time, , 、 They represent , Time-based control input, Indicates the prediction step size. This indicates the control step size.

[0306] The eighth safety constraint and rollback module is used to set safety constraints and rollback mechanisms. If the safety constraints are not met, the optimal control sequence output by the lower-level MPC module is rejected. If the safety constraints are met, the first element of the optimal control sequence is selected and sent to the actuator. If the rollback mechanism is not met, the control mode is automatically switched based on a conservative control strategy to ensure the safe and stable operation of the system. Specifically, the failure to meet the safety constraints refers to: when the optimal pH setpoint exceeds the constraint range. Or the predicted variance exceeds the variance threshold σ max If the safety constraints are not met, the optimal control sequence output by S4 will not be executed.

[0307] Furthermore, the failure to meet the rollback mechanism specifically refers to: the system pH value deviation exceeding the deviation threshold ΔpH within two consecutive optimization cycles.max When necessary, the system automatically switches control modes based on a conservative control strategy to ensure safe and stable operation.

[0308] This invention introduces a multi-objective Bayesian optimization algorithm, taking desulfurization efficiency and limestone consumption as dual optimization objectives. Under the premise of ensuring SO2 emissions meet standards, it dynamically optimizes the pH setpoint to achieve a balance between desulfurization efficiency and operational economy, avoiding resource waste or environmental risks caused by traditional single-objective optimization strategies. In predictive modeling, this invention combines gradient boosting trees and Gaussian process regression, which not only obtains high-precision predictions of operating parameters but also provides prediction uncertainty assessment, enabling pH optimization decisions to have risk perception capabilities. This improves the robustness and stability under operating conditions such as coal quality changes and load fluctuations, effectively enhancing the adaptability of the control system to different operating conditions.

[0309] Example 9 The difference between this embodiment and embodiments one through eight is that, as Figure 10 As shown, the circulating pump optimization control subsystem includes: The ninth data acquisition and preprocessing module is used to acquire and preprocess the absorption tower's operating data; specifically, the acquired absorption tower operating data is the raw data stream of all operating parameters, including the unit load P. unit Flue gas flow rate G, inlet SO2 concentration C in SO2 concentration at the outlet C out slurry pH, density ρ sl CaCO3 content fraction w CaCO3 and the start / stop status of each circulating pump. i Frequency f i Current I i and power P i .

[0310] Furthermore, the preprocessing includes using a sliding window quartile anomaly detection method for anomaly removal; using cubic spline interpolation to repair missing values ​​to fill in missing values ​​caused by anomaly removal or sensor failure; and performing Z-score standardization on each data point to eliminate dimensional differences between parameters.

[0311] The ninth SO2 prediction module constructs an outlet SO2 concentration prediction model based on a long short-term memory network. It uses the system state within a sliding window as input to make rolling predictions of SO2 concentrations for future periods.

[0312] In this embodiment, the outlet SO2 concentration prediction model includes an input layer, a dual-layer LSTM encoder, an attention layer, a single-layer LSTM decoder, and an output layer. The input layer is used to receive the input sequence. The length of the input sequence corresponds to each time step of the sliding window, and the input dimension of each time step is d; Each layer of the dual-layer LSTM encoder contains multiple LSTM units; wherein the first layer LSTM encoder outputs the first layer hidden state. The second-layer LSTM encoder uses Given the input, output the second hidden state. and cell state Each LSTM unit calculates its own forget gate after receiving input. Input gate Output gate and update its own cell state. and hidden state ; The attention layer is used to calculate the hidden state sequence output by the dynamically weighted two-layer LSTM encoder. At each prediction time step t of the decoder, the attention weights of the hidden state sequence are calculated. ; The initial state of the single-layer LSTM decoder is set to the final state of the encoder, and the input is the previous predicted value and the context vector. splicing; The output layer is specifically a fully connected layer that uses a linear activation function; the output of the single-layer LSTM decoder is used to generate multi-step prediction values ​​through the fully connected layer. .

[0313] The ninth hard constraint module, based on the predicted SO2 concentration at the outlet and environmental emission standards, calculates the desulfurization efficiency and, inversely, calculates the minimum slurry flow rate required to meet the desulfurization efficiency requirements in the future, including: The ninth desulfurization efficiency unit is based on the predicted SO2 concentration at the outlet for future periods. Combined with inlet SO2 concentration and emission limits A slurry flow rate demand calculation model was constructed to calculate the required desulfurization efficiency. :

[0314] The ninth slurry flow rate demand unit, based on the mass transfer-reaction model of the slurry absorption tower, converts the desulfurization efficiency into slurry flow rate demand. After fitting using an empirical formula, it can be represented logically as follows:

[0315] Where G is the flue gas flow rate, C out For the SO2 concentration at the outlet, take K is the overall mass transfer coefficient, and A is the effective mass transfer area. The logarithmic mean concentration difference is calculated using the following formula. :

[0316] in, The SO2 saturation concentration at the slurry interface is determined by the slurry pH value, limestone activity, and temperature parameters. The simplified slurry flow rate requirement is obtained by regressing historical data from the above formula. :

[0317] in, All are regression coefficients, which can be obtained by fitting offline data using the least squares method; T is the slurry temperature.

[0318] The ninth multi-objective optimization module is used to take slurry flow rate demand as a hard constraint and the power consumption of the circulating pump group, limestone consumption, and the number of start-stop switching as optimization objectives. A weighted comprehensive cost function is constructed, and a genetic algorithm is used to quickly perform a global search within the frequency-start-stop combination space to generate the most economical pump start-stop and frequency conversion strategy under the current operating conditions in real time. The ninth multi-objective optimization module includes: The ninth constraint unit is used to control the slurry flow rate. To solve the hard constraints of the optimization problem, a comprehensive cost function J for the entire life cycle operation is constructed. This comprehensive cost function J is determined by the pump group power consumption E. pump Limestone consumption E material and start / stop switching penalty N switch Weighted composition; in this embodiment, the comprehensive cost function J is represented using the following logic:

[0319] in, E respectively pump E material and N switch The corresponding weighting coefficients satisfy The weight values ​​were determined using the analytic hierarchy process (AHP).

[0320] The ninth global search unit is used to perform a fast global search within the frequency-start-stop combination solution space using a genetic algorithm to generate the optimal pump start-stop state and operating frequency, which serves as the most economical pump start-stop and frequency conversion strategy under the current operating conditions. Specifically: First, define the solution space encoding vector. ,in Indicates the start / stop status of the circulating pump. Indicates the operating frequency of the circulating pump; Secondly, a genetic algorithm is used to perform a global search within the frequency-start / stop combination solution space, specifically: Initialize the population and generate randomly. Individual; Tournament selection is employed, and the individual selection probability is calculated based on fitness values. As shown in the following formula:

[0321] Among them, J i Indicates the first The fitness function value of each individual; An adaptive single-point crossover operation is employed, adjusting the crossover probability based on population diversity. As shown in the following formula:

[0322] in, The initial crossover probability, Here, is the decay coefficient, and diversity is the average Hamming distance of the population; An adaptive uniform mutation operation is employed to adjust the mutation probability based on population diversity. As shown in the following formula:

[0323] in, The initial mutation probability; Based on the elite retention operation, the best is retained in each generation. Individual; The convergence criterion is defined as follows: convergence occurs when the maximum number of iterations is reached or the improvement of the optimal solution over multiple consecutive generations is less than a preset threshold.

[0324] The ninth execution module is used to send the most economical pump start-up and shutdown and frequency conversion strategies to the distributed control system, and to execute the data acquisition and preprocessing module to the multi-objective optimization module in a rolling loop, and to set a rollback mechanism.

[0325] This embodiment constructs a multi-step prediction model for outlet SO2 concentration based on LSTM to accurately identify and pre-quantify the coupled effects of multiple factors such as unit load, flue gas flow rate, inlet sulfur content, and slurry activity on slurry demand. Then, by utilizing slurry demand inversion and the full life cycle operating cost function, the pump group power consumption, limestone material consumption, and start-up / shutdown switching penalties are uniformly incorporated into a multi-objective optimization framework. The optimal start-up / shutdown and frequency conversion combination is quickly solved in the rolling time domain using a genetic algorithm. The most economical pump start-up / shutdown and frequency conversion adjustment commands under the current operating conditions are generated and sent to the actuators, and the minimum slurry flow rate required for the circulating pump to meet the desulfurization efficiency requirements is obtained.

[0326] Example 10 The difference between this embodiment and embodiments one through nine is that, as Figure 11 As shown, the slurry supply control subsystem includes: The tenth data acquisition and processing module is used to acquire multimodal raw data and preprocess it to generate process variables, derivative features and multimodal features; the multimodal raw data includes, but is not limited to, the pH of the absorber slurry, the sulfur dioxide concentration SO2_in at the absorber inlet, the sulfur dioxide concentration SO2_out at the absorber outlet, the slurry density ρ, the total slurry supply flow rate Q_feed, the internal circulation volume Q_circ, and the slurry mass flow rate L; The process variables include, but are not limited to, the variance of pH of the absorber slurry, pH_var, the variance of inlet sulfur dioxide concentration, SO2_in_var, the variance of outlet sulfur dioxide concentration, SO2_out_var, the variance of slurry density, ρ_var, the variance of total slurry flow rate, Q_feed_var, the variance of internal circulation volume, Q_circ_var, and the variance of slurry mass flow rate, L_var. The derivative characteristics include, but are not limited to, the rate of change of slurry pH (dpH_dt), the rate of change of sulfur dioxide concentration (dSO2_dt), the rate of change of slurry density (dρ_dt), and the rate of change of total slurry flow rate (dQ_dt). The multimodal features include, but are not limited to, the visual feature vision_feat, the acoustic feature acoustic_feat, the vibration feature vib_feat, and the spectral feature spectrum_feat.

[0327] The tenth soft measurement estimation module is used to perform soft measurement and health estimation on multimodal features, and convert them into key process parameters and equipment status indicators, specifically: Based on multimodal characteristics, online estimation of slurry quality parameters, including limestone particle size distributions D50 and D90, effective alkalinity alk_eff, and Ca / S demand CaS_need; Estimate the supersaturation degree SI and its uncertainty, wherein the supersaturation degree SI is represented using the following logic:

[0328] Where IAP is the ion activity product and K_sp is the solubility product constant of the corresponding compound; The device health score H_score and health status tags are evaluated by combining acoustic characteristics, vibration characteristics, current signals, and power spectrum. The device health score H_score is represented using the following logic:

[0329] in, For the first A fusion feature, For the first The weights corresponding to each fusion feature, σ is the activation function, and the output is normalized to the range of 0-1, where 1 represents a healthy state and 0 represents a serious fault; based on the device health score H_score, a threshold is set to generate health_tags; Adaptive operating conditions are achieved using a hybrid expert and attention mechanism.

[0330] The tenth rolling prediction module is used to predict the prediction sequence and sensitivity of the absorber slurry pH, outlet sulfur dioxide concentration and supersaturation in the future time domain using the absorber mechanism model. In this embodiment, the following logic represents the prediction of the current sulfur dioxide concentration at the absorber outlet using an absorber mechanism model:

[0331] Where K represents the absorption coefficient, L represents the mass flow rate of the circulating slurry in the absorption tower, and G represents the flue gas volume. ρ represents the density of the slurry. Indicates the effective alkalinity factor. , and then predict the sulfur dioxide concentration sequence SO2_out_pred[ ] at the outlet of the absorption tower in the future time domain by rolling prediction; The following logic can be used to predict the current pH of the absorber slurry using an absorber mechanism model:

[0332] Among them, among them, Indicates the rate of change of pH value of the slurry over time. This represents the gain coefficient for fresh slurry replenishment. This represents the attenuation coefficient of sulfur dioxide absorption. This indicates the pH balance value of the slurry. The pH_pred[ ] represents the pH time constant, and is used to predict the pH sequence of the absorber slurry in the future time domain by rolling. The following logic can be used to predict the current supersaturation of the absorber tower using a mechanistic model:

[0333] in, This represents the temperature-dependent solubility product, where T represents the slurry temperature. This indicates the effective concentration of calcium ions in the slurry. The effective concentration of sulfate ions in the slurry is represented by SI_pred[ ], which is then used to predict the supersaturation sequence of the absorber in the future time domain by rolling.

[0334] The tenth actuator model module is used to build actuator models based on the current actuator state, including slurry flow rate model, pump power model and Jacobian matrix, and to define actuator constraints; In this embodiment, the slurry flow rate model is represented using the following logic:

[0335] in, Q This represents the volumetric flow rate of fresh slurry delivered by the slurry pump to the absorption tower, and represents the core control input. This represents the empirical coefficient for slurry flow rate. This indicates the operating speed of the slurry pump, which is the main variable for flow rate regulation. Indicates the density of the slurry; The pump power model can be represented using the following logic:

[0336] in, Indicates the pump's shaft power. This represents the empirical coefficient for pump power. The Jacobian matrix can be represented using the following logic: ,

[0337] in, This represents the partial derivative of the slurry flow rate with respect to the pump speed. This represents the partial derivative of the slurry density with respect to the valve opening. This indicates the degree of valve opening.

[0338] Furthermore, the actuator constraints specifically include the following: Upper and lower limit constraints, specifically, require that the actuator control quantity be within the physical limit range, such as valve opening degree 0-100% and speed 0-1500 r / min; Rate constraint, specifically, limits the rate of change of the control variable in a single step, such as a change in rotational speed ≤ 50 r / min / s; The allocation coefficient constraint specifically ensures that the sum of the flow allocation coefficients for multiple actuators is always 1. Density range constraints, specifically, the slurry density must be within a safe range, such as 1050-1250 kg / m³ to prevent sedimentation or overload; The dwell time constraint specifically requires that the actuator remain in place for ≥60 seconds after the action to avoid frequent actions. The flushing sequence constraint is specifically the sequential logic of the actuator, such as executing the flushing valve first, then the slurry valve, and finally the flushing valve closed in sequence.

[0339] The tenth feedforward control module is used to solve the current control deviation, generate the feedforward control quantity based on the sensitivity mapping, and introduce oversaturation suppression and healthy amplitude limiting bias to form dist_ff[ ] and switch_hint, including: To solve the current control deviation, we first calculate the current control deviation based on the predicted sequence of the absorber tower for the next multiple steps output by S3. Specifically, we calculate the current control deviation by taking the first predicted value of the outlet sulfur dioxide concentration sequence SO2_out_pred[] and the target value of the outlet sulfur dioxide, using the following logic:

[0340] Wherein, ΔSO2 represents the current control deviation of sulfur dioxide at the outlet. This represents the first predicted value of SO2_out_pred[]. This indicates the export target value for sulfur dioxide; The control deviation of the slurry's pH can be determined using the following logical expression:

[0341] Wherein, ΔpH represents the current control deviation of the slurry's acidity / alkalinity. This represents the first-step prediction value of pH_pred[ ]. Indicates the target value for the pH of the slurry; Feedforward control quantities are generated based on sensitivity mapping, including feedforward slurry flow rate Q_feed_ff and feedforward density rho_ff.

[0342] Oversaturation suppression and health limiting bias are introduced to form the multi-actuator allocation ratio dist_ff[ ] and the switching hint switch_hint.

[0343] Furthermore, the multi-objective optimization function is represented in S6 using the following logic:

[0344] in, These are the cost function weights corresponding to slurry pH, outlet sulfur dioxide, energy consumption, material consumption, supersaturation, and health limit, respectively. express Target value, , All are nonlinear power coefficients of the deviation term. This represents the control increment, specifically the difference between the current control value and the value at the previous moment. This represents the control increment penalty matrix.

[0345] The tenth multi-objective optimization module is used to construct a multi-objective optimization function with slurry acidity and alkalinity and outlet sulfur dioxide concentration as the core, while taking into account energy consumption, material consumption, supersaturation and equipment health status. Based on actuator constraints, model predictive control is solved to obtain the optimal control increment sequence. The tenth instruction mapping module is used to map the optimal control increment sequence or the feedforward reference quantity when optimization fails to the current control instruction, and to apply protective constraints based on oversaturation and equipment health score before issuing the instruction. The tenth closed-loop feedback and scene memory module provides closed-loop feedback based on the execution results, records operation events, and stores scene memories, storing the current control scene and effects in the memory bank, including: Closed-loop feedback is performed based on the execution results to optimize subsequent control cycles; the errors in slurry pH and outlet sulfur dioxide concentration, e_pH and e_SO2, are calculated, specifically... , ;in, This indicates the actual value of the slurry's pH level. Indicates the target value for the pH of the slurry. This indicates the actual value of sulfur dioxide concentration at the export site. This indicates the target value for sulfur dioxide concentration at the export level; Record operation events, store scene memories, and store the current controlled scene and effects in the memory bank; In this embodiment, the operation event is a snapshot of the control action, including time, scene, action, source, effect, and constraint. By converting the operation event into a structured "scene-action-effect" combination and storing it in the memory bank, the current control is optimized using historical experience when a similar scene is encountered.

[0346] Specifically, the time refers to the action time, the scene refers to the working conditions at the time of the action, the action refers to the specific control action executed, the source refers to the specific source of the control action (e.g., from the feedforward module or the optimizer), the effect refers to the change after the action is executed, and the constraint refers to the constraint conditions of the control action. The structured "scene-action-effect" combination specifically uses a set of standardized values ​​to represent the working conditions at the time (scene F), the control action (action S), the change after the action is executed (effect E), and the fingerprint ID. The fingerprint is specifically a representation of the scene value. In this embodiment, a hash string can be used as the fingerprint for different scenes to quickly find different scenes.

[0347] When encountering similar scenarios, optimizing the current control using historical experience specifically involves comparing the current operating conditions with scenarios in the memory bank. If the similarity between the current operating conditions and scenarios in the memory bank exceeds 85%, a similar scenario is considered found, and this memory is reset for a preset cooling time to avoid repeatedly using the same strategy in a short period and prevent control oscillations. Then, the optimal control action from the historical scenario is used as the starting point for the current optimizer, allowing it to calculate results faster instead of starting from scratch. Furthermore, if the implementation of the historical scenario was ineffective (e.g., oversaturation increased previously), more emphasis is placed on suppressing oversaturation (increasing the weight of oversaturation); if energy consumption was reduced too much previously, the energy consumption limit is slightly relaxed (decreasing the weight of energy consumption).

[0348] In this embodiment, the absorption tower mechanism model is used not only to provide a predicted sequence and sensitivity of the future state of the system and to represent the coupling relationship between key variables, but also to use the absorption tower mechanism model as a constraint condition to execute subsequent optimization control under the constraint and guidance of the mechanism model, so that the optimization strategy always conforms to the physical reality and the desulfurization reaction mechanism, effectively improving the working condition adaptability and robustness of the desulfurization slurry supply control system.

[0349] The absorption tower mechanism model predicts the future time-domain sequence in a rolling manner, providing a feasible initial solution for model predictive control. When the system experiences strong disturbances, it can act in advance within a short control cycle, calculating the slurry pH and outlet sulfur dioxide concentration required to maintain stability, thus shortening the solution time and greatly improving the number of iterations and dynamic response speed. In addition, when abnormal triggering of execution steps occurs, such as degradation or rollback, the generated feedforward control quantity can be directly used as the control command, ensuring that the system continues to operate under acceptable performance and guaranteeing the robustness of the control system.

[0350] Example 11 The difference between this embodiment and embodiments one through ten is that, as Figure 12 As shown, the induced draft fan control system includes: The eleventh data acquisition module is used to acquire the operating parameters of the boiler and other furnace structural systems in real time, specifically the raw physical quantities of the instruments acquired on-site, including: furnace negative pressure. Oxygen content in flue gas Nitrogen oxide concentration Pressure difference before and after the air preheater induced draft fan speed Leaf angle Fan current 1. Exhaust fan outlet air volume , control sequence of dampers or valve positions at all levels Unit load and furnace temperature .

[0351] The eleventh data processing module is used to preprocess the collected operating parameters, specifically: First, median filtering and low-pass filtering are applied to the key signal to obtain the estimated value and variance of the key signal. The key signals include furnace negative pressure. Oxygen content in flue gas Nitrogen oxide concentration induced draft fan speed and leaf angle The furnace negative pressure estimate is obtained by filtering. Estimated oxygen content in flue gas Estimated nitrogen oxide concentration Estimated speed of induced draft fan Estimated value of moving leaf angle And the estimated variances corresponding to each estimated value; Secondly, calculate the first-order difference characteristics, including the rate of change of furnace negative pressure. Rate of change in oxygen content in flue gas and unit load change rate ; Next, obtain the calibration air volume. ; Then, calculate the estimated value of flue resistance. ;in, This represents a linear or empirical formula.

[0352] The eleventh dual-prediction module, based on preprocessed operating parameters, employs a parallel prediction model to perform multi-step predictions of furnace negative pressure and flue gas oxygen content, specifically: First, a linear state-space model or Kalman filtering is used for multi-step prediction; then, through multi-step recursion, the following is obtained: , , where, This represents the estimated negative pressure in the furnace during the future time period T. This represents the estimated oxygen content of the flue gas during the future time period T. Secondly, the prediction results of multiple models are weighted and merged, specifically by weighting and combining them according to the inverse of the estimated variance, as shown in the following formula: ,

[0353] in, This represents the predicted value of the furnace negative pressure after weighted fusion. This represents the predicted value of oxygen content in the flue gas after weighted fusion. Indicates the estimated value of furnace negative pressure The corresponding estimated variance, This represents the estimated oxygen content in the flue gas. The corresponding estimated variance; Then, simultaneously, based on the prediction results, an empirical sensitivity to future wind volume changes is output. and .

[0354] The eleventh trend analysis module, based on the operation of the nitrogen oxide online meter, performs soft measurement of nitrogen oxide concentration and trend prediction of nitrogen oxide concentration, and obtains the predicted value of nitrogen oxide concentration. The eleventh air volume estimation module, based on the fan's static characteristic model, calculates the estimated air volume, the Jacobian matrix of air volume to actuator, and the estimated fan power based on the current rotational speed and blade angle. Specifically: First, based on the static characteristics of the wind turbine, offline fitting was performed to obtain... and fit the power ; Obtain characteristic curves provided by the wind turbine manufacturer offline or obtain wind turbine operating data through on-site testing, and establish the relationship between wind turbine flow rate and power and induced draft fan speed. With leaf angle The functional relationship, obtained through table lookup or bivariate polynomial fitting, is represented by the following logic:

[0355]

[0356] in, For fan flow rate, For the power of the induced draft fan, , , respectively, are the fitting coefficients for fan flow rate and induced draft fan power, obtained through offline fitting, and i and j are the influencing factors of induced draft fan speed and blade angle, respectively; Then, airflow calibration and Jacobian matrix estimation are performed, specifically as follows: When the control system is equipped with an air volume meter, the fan is calibrated using the air volume meter, and the result is obtained by referring to a table. The obtained functional relationship is calibrated using the following formula: ;in, These are calibration coefficients; Define the Jacobian matrix as ; When the control system is not equipped with an air volume meter, the flow rate prediction and Jacobian matrix estimation are directly performed using the fitting function or table lookup data from step S51.

[0357] The eleventh feedforward calculation module is used to solve the predicted deviation between furnace negative pressure and flue gas oxygen content, to calculate the engineering feedforward component, and to calculate the engineering feedforward air volume based on the engineering feedforward component. Specifically: First, take and The prediction bias is calculated from the initial predicted value using the following logical expression: ,

[0358] in, express The deviation between the initial predicted value and the reference value. express The first predicted value in the middle, This indicates the target value of the furnace negative pressure. express The deviation between the initial predicted value and the reference value. express The first predicted value in the middle, This indicates the target value for the oxygen content in the flue gas; Secondly, the engineering feedforward components, including the furnace negative pressure deviation feedforward component, are calculated based on the predicted deviation. Feedforward component of flue gas oxygen content deviation and penalties for exceeding nitrogen oxide limits This can be represented using the following logic:

[0359]

[0360]

[0361] in, This represents the feedforward component of the furnace negative pressure deviation. This represents the feedforward component of the deviation in oxygen content of flue gas. This indicates the penalty amount for exceeding the limit for nitrogen oxides. Represents the saturation function. This represents the proportionality coefficient of the furnace negative pressure deviation. This represents the differential coefficient of the furnace negative pressure deviation. This indicates the rate of change of the negative pressure deviation in the furnace. This represents the limit value of the differential term of the furnace negative pressure deviation. This indicates the safety limit for the negative pressure deviation in the furnace. This represents the proportionality coefficient for the deviation in oxygen content of flue gas. This represents the differential coefficient of the deviation in oxygen content of flue gas. This indicates the rate of change in the oxygen content of the flue gas. This represents the limit value of the differential term of the deviation in flue gas oxygen content. This indicates the safety limit for deviations in the oxygen content of flue gas. Indicates the nitrogen oxide penalty coefficient. This indicates the upper limit for nitrogen oxide emissions; Next, synthesize the reference feedforward air volume. This can be represented using the following logic:

[0362] in, They represent The weighting coefficients range from 0 to 1. Then, calculate the energy consumption cost of the wind turbine. , used to Provided to the optimizer.

[0363] The eleventh multi-objective optimization module is used to construct a quadratic programming problem with negative pressure deviation, oxygen content deviation, nitrogen oxide deviation, and fan power as optimization objectives, and actuator limiting, speed, and surge risk as constraints. Specifically: First, based on the linearized system dynamics, a sensitivity matrix is ​​constructed at the current operating point. ; Secondly, the negative pressure in the furnace Oxygen content in flue gas Nitrogen oxide concentration Exhaust fan power To optimize the objective, a quadratic programming problem QP is constructed, and the coefficient matrix of the quadratic terms in QP is obtained. With the linear term coefficient vector ; ,

[0364] in, This represents the control increment penalty matrix. This represents the vector of predicted values ​​for the output variable. A vector representing the reference values ​​of the output variables; Then, linear inequality constraints are added, including rate constraints, upper and lower bound constraints, and surge line approximation constraints. Soft constraints.

[0365] The eleventh programming module is used to solve quadratic programming problems and obtain the optimal control increments, including the optimal control increment sequences for the induced draft fan speed and blade angle. The eleventh control backoff module is used to take the first term of the control increment sequence as the current control command if the quadratic programming problem is solved successfully; if the solution fails or a timeout occurs, it combines the engineering feedforward air volume value with the Jacobian matrix to obtain the backoff control command, specifically: Approximated by the Jacobian matrix Solve for the rollback control instruction, specifically: , ,

[0366] in, This indicates a deviation from the airflow target. This indicates the target air volume of the induced draft fan. This indicates the current air volume of the induced draft fan. This indicates the adjustment amount of the induced draft fan speed. This indicates the amount of adjustment for the blade angle. Indicates the allocation coefficient; Furthermore, it also includes application dead-time limiting, rate cutoff, and actuator reallocation operations; wherein, the application dead-time limiting specifically refers to... but Returning to 0, the rate cutoff specifically means , The upper limit of the rate of change of rotation speed is defined as the inter-actuator redistribution operation, which specifically involves transferring the remaining control quantity to another actuator when any actuator reaches its limit. Furthermore, before calculation, it is checked whether the running point will be pushed into the surge region. If the distance from the running point to the surge line is less than the preset threshold surge_threshold, the distance is reduced. Alternatively, you can reject the command and log it.

[0367] The eleventh execution module is used to perform security verification on the control commands before sending them to the induced draft fan actuator.

[0368] This embodiment is based on the strong coupling relationship between furnace negative pressure and flue gas oxygen content. It uses a parallel prediction model to predict furnace negative pressure and flue gas oxygen content in multiple steps, and calculates the sensitivity of future air volume to the predicted target in advance. The two are treated as a whole for coordinated prediction and compensation. This allows the system to pre-calculate the required compensation air volume before disturbances (such as load changes and coal quality fluctuations) actually affect the controlled parameters. The system then sends instructions to the induced draft fan actuator through the optimizer, realizing pre-intervention of the induced draft fan outlet air volume. This significantly reduces control lag, shortens the recovery time under varying operating conditions, and improves the operating condition adaptability of the control system. Furthermore, due to the pre-action, the system does not need to compensate for disturbances through large-scale correction, significantly reducing overshoot. This makes the furnace negative pressure and oxygen content curves smoother and more stable, greatly improving the stability of the combustion process.

[0369] Example 12 The present invention also provides a control method for the intelligent control system of the flue gas environmental protection island described in Embodiments 1 to 11 above, comprising the following steps: S100, establish denitrification control model, dust removal control model, desulfurization control model and induced draft fan control model.

[0370] S200, the denitrification control model includes a urea pyrolysis control sub-model, an ammonia injection regulation control sub-model, a denitrification reactor temperature field control sub-model, and a dilution fan regulation sub-model; The urea pyrolysis control sub-model is used to optimize the urea injection rate and burner power online, recover waste heat from flue gas, and preheat the urea solution. The ammonia injection regulation and control sub-model is used to predict and regulate the ammonia injection amount through time-series trend analysis. The temperature field control sub-model of the denitrification reactor is used to predict the temperature field distribution, outlet NOx concentration and ammonia slip of the denitrification reactor in future periods. The dilution fan control sub-model is used to generate the target fan speed and output the optimal dilution air volume under the current operating conditions; S300, the dust removal control model includes a bag cleaning control sub-model and / or an electrostatic field impedance matching sub-model; The bag cleaning control sub-model is used to solve the optimal pulse-jet cleaning strategy for the bag dust collector in future time periods through multi-objective optimization. The electrostatic field impedance matching sub-model is used to dynamically correct the current and voltage values ​​of the electrostatic dust removal equipment and adjust the high-frequency power supply output. S400, the desulfurization control model includes a gypsum dewatering control sub-model, a slurry pH control sub-model, a circulating pump optimization control sub-model, and a slurry supply control model; The gypsum dehydration control sub-model is used to generate the optimal filter cake moisture content and the optimal control command for the vacuum belt filter. The pH optimization control sub-model for the slurry is used to dynamically track the optimal pH setpoint of the absorption tower. The circulating pump optimization control sub-model is used to generate the circulating pump start-up, shutdown, and frequency conversion strategies for the current operating conditions; The slurry supply control sub-model is used to solve for the optimal control increment sequence of limestone slurry through multi-objective optimization.

[0371] S500, the induced draft fan control model is used to generate the optimal induced draft fan speed and blade angle.

[0372] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A smart control system for a flue gas environmental protection island, characterized in that, The control system includes a denitrification control system, a dust removal control system, a desulfurization control system, and an induced draft fan control system; The denitrification control system includes a urea pyrolysis control subsystem, an ammonia injection regulation control subsystem, a denitrification reactor temperature field control subsystem, and a dilution fan regulation subsystem. The urea pyrolysis control subsystem is used to optimize the urea injection rate and burner power online, recover waste heat from flue gas, and preheat the urea solution; the ammonia injection regulation control subsystem predicts and adjusts the ammonia injection rate through time-series trend analysis; the denitrification reactor temperature field control subsystem is used to predict the temperature field distribution, outlet NOx concentration, and ammonia slip of the denitrification reactor in future periods; and the dilution fan control subsystem is used to generate the target fan speed and output the optimal dilution air volume for the current operating conditions.

2. The intelligent control system for the flue gas environmental protection island according to claim 1, characterized in that, The dust removal control system includes a bag cleaning control subsystem and / or an electrostatic field impedance matching subsystem; The bag filter cleaning control subsystem is used to solve the optimal pulse-jet cleaning strategy for the bag filter in future time periods; the electrostatic field impedance matching subsystem is used to dynamically correct the current and voltage values ​​of the electrostatic precipitator and adjust the high-frequency power supply output.

3. The intelligent control system for the flue gas environmental protection island according to claim 1, characterized in that, The desulfurization control system includes a gypsum dewatering control subsystem, a slurry pH control subsystem, a circulating pump optimization control subsystem, and a slurry supply control subsystem. The gypsum dewatering control subsystem is used to generate the optimal filter cake moisture content and the optimal control command for the vacuum belt filter; the slurry pH optimization control subsystem is used to dynamically track the optimal pH setpoint of the absorption tower; the circulating pump optimization control subsystem is used to generate the circulating pump start-up and shutdown and frequency conversion strategy for the current operating condition; and the slurry supply control subsystem is used to solve the optimal control increment sequence for limestone slurry.

4. The intelligent control system for the flue gas environmental protection island according to claim 1, characterized in that, The induced draft fan control system is used to generate the optimal induced draft fan speed and blade angle.

5. The intelligent control system for the flue gas environmental protection island according to claim 1, characterized in that, The urea pyrolysis control subsystem includes: The first sensing module is used to deploy a multimodal sensing network in the pyrolysis furnace to collect multi-source heterogeneous data in real time. The first fusion module is used to input the collected data into the edge computing node after synchronization by the PLC, construct the CFD-DEM digital twin synchronization data and output the prediction residual, and expand the effective working condition data based on the adversarial generative network. The first decision module is used to continuously calculate the optimal control command of the first decision module based on the online optimization of urea injection quantity and burner power, and switch the control mode of the first decision module according to abnormal conditions. The first execution module is used to receive control commands from the first decision module, adjust the spray gun and burner based on the control commands and dynamic zoning strategy, and actively intervene to remove crystals; The first energy recovery module is used to recover waste heat from flue gas and preheat urea solution through the ORC unit.

6. The intelligent control system for the flue gas environmental protection island according to claim 1, characterized in that, The ammonia injection regulation and control subsystem includes: The second data acquisition module is used to deploy a multimodal sensor network to acquire raw signals from the denitrification reactor; The second data processing module is used to preprocess the raw signal to form a standardized data stream; The second time-series trend analysis module is used to perform weighted smoothing on historical standardized data, capturing the overall trend of signal changes and short-term multi-stage predictions respectively. The second adaptive adjustment module is used to calculate the ammonia injection adjustment amount based on the deviation between the predicted value and the target set value. The second closed-loop feedback module is used to compare the predicted value with the actual measured value in real time and correct the ammonia injection amount. The second execution module is used to control the ammonia injection actuator to complete the ammonia injection based on the corrected ammonia injection quantity. The second model self-learning module is used to update and optimize model parameters online.

7. The intelligent control system for the flue gas environmental protection island according to claim 1, characterized in that, The temperature field control subsystem of the denitrification reactor includes: The third data processing module is used to collect sparse temperature measurement data from the denitrification reactor and perform data preprocessing. The third interpolation and reconstruction module is used to perform spatial interpolation and reconstruction on sparse temperature measurement data, and outputs the reconstructed temperature field, boundary gradient and near-wall mask features. The third model prediction module combines the output of spatial interpolation reconstruction with working condition data, coal type information and actuator status data as input features to construct a directional ConvLSTM spatiotemporal prediction model. The model is trained with a boundary-first masking strategy and outputs the temperature field distribution, outlet NOx concentration and ammonia escape amount for multiple future steps. The third MPC module performs multi-objective model predictive control based on the output of the predictive model, constructs a multi-objective cost function, and optimizes the control quantity of the actuator of the denitrification reactor online. The third control execution module is used to compensate for the time delay of the denitrification reactor actuator and issue the optimized control quantity, set the rollback conditions, and automatically roll back to the original control strategy when the rollback conditions are met and continue to collect data.

8. The intelligent control system for the flue gas environmental protection island according to claim 1, characterized in that, The dilution fan control subsystem includes: The fourth data sensing module is used to collect and transmit denitrification operating parameters in real time, and to preprocess the denitrification operating parameters. The fourth prediction module is used to predict the outlet NOx concentration and optimal dilution air volume in the future period based on the denitrification operation parameters using an improved two-layer LSTM neural network, and to optimize it through a process feature weighting mechanism, a cross-layer attention mechanism and a multi-objective loss function. The fourth decision module is used to output the target speed of the dilution fan based on the prediction results and the current operating conditions using a deep deterministic gradient reinforcement learning algorithm. The fourth safety control module is used to automatically take over control when the decision module outputs an abnormality, employing a fuzzy PID controller. The fourth verification module performs full-condition simulation verification of the fourth prediction module and the fourth decision module based on the digital twin platform.

9. The intelligent control system for the flue gas environmental protection island according to claim 2, characterized in that, The bag filter cleaning control subsystem includes: The fifth data acquisition module is used to collect and monitor the operating parameters of the bag filter in real time. The operating parameters include inlet and outlet pressure difference, flue gas flow rate, dust concentration, temperature and humidity. The fifth dynamic modeling module constructs a mathematical model based on the dynamic equation of filter bag pressure difference, which is used to describe the impact of dust accumulation process and pulse-jet cleaning behavior on the pressure difference change of the internal filter bags in the bag filter. The fifth prediction module, based on the mathematical model and the current operating parameters, predicts the operating status of the bag filter over a period of time in the future. The fifth optimization decision module is used to solve the optimal blowing strategy for a future period of time using a multi-objective optimization algorithm under the premise of meeting preset constraints. The blowing strategy includes blowing pressure, pulse width, blowing interval and timing arrangement. The fifth execution module, based on the rolling time-domain strategy, performs pulse-jet cleaning on the bag filter according to the instructions generated by the optimization decision module, and collects feedback information for optimization in the next cycle. The fifth adaptive update module is used to automatically correct the mathematical model parameters based on long-term running data.

10. The intelligent control system for the flue gas environmental protection island according to claim 2, characterized in that, The electrostatic impedance matching subsystem includes: The sixth data acquisition module is used to collect operating parameters and slurry quality data in real time; The sixth edge computing module is used to perform data cleaning, feature extraction, and operating condition prediction, and generate initial current and voltage setpoints. Specifically, it cleans the collected data and extracts a subset of features that are strongly correlated with current and voltage. Based on the feature subset, it predicts the initial current and voltage setpoints through a generative adversarial network model. The sixth cloud-based optimization module dynamically corrects the initial set value based on the slurry quality index model; The sixth actuator module is used to adjust the high-frequency power supply output according to the corrected current and voltage settings.

11. The intelligent control system for the flue gas environmental protection island according to claim 3, characterized in that, The gypsum dehydration control subsystem includes: The seventh data acquisition and preprocessing module is used to acquire key operating parameters in the gypsum dehydration process in real time and preprocess the key operating parameters. The seventh moisture content prediction module runs the mechanism model and the data-driven model in parallel based on the preprocessed data, and dynamically calculates the fusion weight based on the real-time prediction error of the two models, and obtains the final predicted value of filter cake moisture content by weighted fusion. The seventh multi-objective optimization and MPC control module constructs a weighted multi-objective optimization function based on constraints, uses model predictive control for real-time optimization, and employs a periodic evolutionary algorithm for global multi-objective optimization. Based on the fusion strategy, it generates the final control command and issues it to the corresponding actuator. The seventh control command execution and feedback module is used to execute the issued gypsum dehydration control commands, collect feedback data, perform real-time deviation correction, and form a closed-loop control. The seventh adaptive learning module performs online parameter fine-tuning and periodic offline retraining of the data-driven model based on feedback data, and constructs anomaly detection and response strategies.

12. The intelligent control system for the flue gas environmental protection island according to claim 3, characterized in that, The slurry pH control subsystem includes: The eighth data acquisition and preprocessing module is used to acquire and preprocess the absorption tower's operating data; The eighth dual-objective prediction module is used to construct a hybrid prediction model using gradient boosting trees and Gaussian process regression, and output prediction results with uncertainty assessment. The eighth upper-level optimization module is used to search the Pareto optimal frontier using a multi-objective Bayesian optimization algorithm with a hybrid prediction model as the optimization objective function. The eighth lower-level MPC module is used to dynamically track the optimal pH setpoint by the Pareto optimal frontier input adaptive model predictive control execution and solve for the optimal control sequence in a rolling manner. The eighth safety constraint and rollback module is used to set safety constraints and rollback mechanisms. If the safety constraints are not met, the optimal control sequence output by the lower-level MPC module will not be executed. If the safety constraints are met, the first element of the optimal control sequence will be selected and sent to the corresponding actuator. If the rollback mechanism is not met, the control mode will be automatically switched based on a conservative control strategy to ensure the safe and stable operation of the system.

13. The intelligent control system for the flue gas environmental protection island according to claim 3, characterized in that, The circulating pump optimization control subsystem includes: The ninth data acquisition and preprocessing module is used to acquire and preprocess the operating data of the absorption tower. The ninth SO2 prediction module constructs an outlet SO2 concentration prediction model based on a long short-term memory network, and uses the system state within a sliding window as input to make rolling predictions of SO2 concentration in future periods. The ninth hard constraint module, based on the predicted SO2 concentration at the outlet and environmental emission standards, calculates the desulfurization efficiency and reverses the minimum slurry flow rate required to meet the desulfurization efficiency requirements in the future period. The ninth multi-objective optimization module is used to take the slurry flow rate demand as a hard constraint, and the power consumption of the circulating pump group, limestone consumption and the number of start-stop switching as optimization objectives, and to form a comprehensive cost function by weighting them. It uses a genetic algorithm to quickly search the frequency-start-stop combination space and generate the most economical pump start-stop and frequency conversion strategy under the current working conditions in real time. The ninth execution module is used to send the most economical pump start-up and shutdown and frequency conversion strategies to the distributed control system, and to execute the data acquisition and preprocessing module to the multi-objective optimization module in a rolling loop, and to set a rollback mechanism.

14. The intelligent control system for the flue gas environmental protection island according to claim 3, characterized in that, The slurry supply control subsystem includes: The tenth data acquisition and processing module is used to acquire multimodal raw data and preprocess it to generate process variables, derivative features and multimodal features; The tenth soft measurement estimation module is used to perform soft measurement and health estimation on multimodal features and convert them into key process parameters and equipment status indicators. The tenth rolling prediction module is used to predict the prediction sequence and sensitivity of the absorber slurry pH, outlet sulfur dioxide concentration and supersaturation in the future time domain using the absorber mechanism model. The tenth actuator model module is used to build actuator models based on the current actuator state, including slurry flow rate model, pump power model and Jacobian matrix, and to define actuator constraints; The tenth feedforward control module is used to solve the current control deviation, generate the feedforward control quantity based on the sensitivity mapping, and introduce oversaturation suppression and healthy amplitude limiting bias. The tenth multi-objective optimization module is used to construct a multi-objective optimization function with slurry acidity and alkalinity and outlet sulfur dioxide concentration as the core, while taking into account energy consumption, material consumption, supersaturation and equipment health status. Based on actuator constraints, model predictive control is solved to obtain the optimal control increment sequence. The tenth instruction mapping module is used to map the optimal control increment sequence or the feedforward reference quantity when optimization fails to the current control instruction, and to apply protective constraints based on oversaturation and equipment health score before issuing the instruction. The tenth closed-loop feedback and scene memory module performs closed-loop feedback based on the execution results, records operation events, and stores scene memory, storing the current control scene and effect in the memory bank.

15. The intelligent control system for the flue gas environmental protection island according to claim 4, characterized in that, The induced draft fan control system includes: The eleventh data acquisition module is used to collect the operating parameters of boiler and other furnace structure systems in real time. The eleventh data processing module is used to preprocess the collected operating parameters; The eleventh dual prediction module, based on the preprocessed operating parameters, uses a parallel prediction model to perform multi-step prediction of furnace negative pressure and flue gas oxygen content; The eleventh trend analysis module, based on the operation of the nitrogen oxide online meter, performs soft measurement of nitrogen oxide concentration and trend prediction of nitrogen oxide concentration, and obtains the predicted value of nitrogen oxide concentration. The eleventh air volume estimation module, based on the static characteristic model of the fan, calculates the estimated air volume, the Jacobian matrix of air volume to actuator, and the estimated fan power based on the current speed and blade angle. The eleventh feedforward calculation module is used to solve the prediction deviation between furnace negative pressure and flue gas oxygen content, to calculate the engineering feedforward component, and to calculate the engineering feedforward air volume based on the engineering feedforward component. The eleventh multi-objective optimization module is used to construct a quadratic programming problem with negative pressure deviation, oxygen content deviation, nitrogen oxide deviation and fan power as optimization objectives, and actuator limit, speed and surge risk as constraints. The eleventh programming module is used to solve quadratic programming problems and obtain the optimal control increments, including the optimal control increment sequences for the induced draft fan speed and blade angle. The eleventh control rollback module is used to take the first term of the control increment sequence as the current control command if the quadratic programming problem is solved successfully; if the solution fails or timeout occurs, the rollback control command is obtained by mapping the engineering feedforward air volume value with the Jacobian matrix. The eleventh execution module is used to perform security verification on the current control command before sending it to the induced draft fan actuator.

16. A method for intelligent control of a flue gas environmental protection island, characterized in that, The control method includes establishing a denitrification control model, a dust removal control model, a desulfurization control model, and an induced draft fan control model; The denitrification control model includes a urea pyrolysis control sub-model, an ammonia injection regulation control sub-model, a denitrification reactor temperature field control sub-model, and a dilution fan regulation sub-model. The urea pyrolysis control sub-model is used to optimize the urea injection rate and burner power online, recover waste heat from flue gas, and preheat the urea solution. The ammonia injection regulation and control sub-model is used to predict and regulate the ammonia injection amount through time-series trend analysis. The temperature field control sub-model of the denitrification reactor is used to predict the temperature field distribution, outlet NOx concentration and ammonia slip of the denitrification reactor in future periods. The dilution fan control sub-model is used to generate the target fan speed and output the optimal dilution air volume under the current operating conditions.

17. The intelligent control method for flue gas environmental protection island according to claim 16, characterized in that, The dust removal control model includes a bag filter cleaning control sub-model and / or an electrostatic field impedance matching sub-model. The bag cleaning control sub-model is used to solve the optimal pulse-jet cleaning strategy for the bag dust collector in future time periods through multi-objective optimization. The electrostatic field impedance matching sub-model is used to dynamically correct the current and voltage values ​​of the electrostatic dust removal equipment and adjust the high-frequency power supply output.

18. The intelligent control method for flue gas environmental protection island according to claim 16, characterized in that, The desulfurization control model includes a gypsum dewatering control sub-model, a slurry pH control sub-model, a circulating pump optimization control sub-model, and a slurry supply control model. The gypsum dehydration control sub-model is used to generate the optimal filter cake moisture content and the optimal control command for the vacuum belt filter. The pH optimization control sub-model for the slurry is used to dynamically track the optimal pH setpoint of the absorption tower. The circulating pump optimization control sub-model is used to generate the circulating pump start-up, shutdown, and frequency conversion strategies for the current operating conditions; The slurry supply control sub-model is used to solve for the optimal control increment sequence of limestone slurry through multi-objective optimization.

19. The intelligent control method for flue gas environmental protection island according to claim 16, characterized in that, The induced draft fan control model is used to generate the optimal induced draft fan speed and blade angle.

20. The intelligent control method for flue gas environmental protection island according to claim 16, characterized in that, The urea pyrolysis control sub-model includes: S201 deploys a multimodal sensor network inside the pyrolysis furnace to collect multi-source heterogeneous data in real time; S202, after the collected data is synchronized by the PLC, it is input into the edge computing node to construct the CFD-DEM digital twin synchronization data and output the prediction residual, and expand the effective working condition data based on the adversarial generative network; S203, based on the online optimization of urea injection quantity and burner power, continuously calculate the optimal control command of the first decision module, and switch the control mode of the first decision module according to abnormal conditions; S204 is used to receive control commands from the first decision module, adjust the spray gun and burner based on the control commands and dynamic zoning strategy, and actively intervene to remove crystals; S205 recovers waste heat from flue gas and preheats urea solution through the ORC unit, achieving a closed-loop energy-saving system.

21. The intelligent control method for flue gas environmental protection island according to claim 16, characterized in that, The ammonia injection regulation control sub-model includes: S211, Deploy a multimodal sensor network to collect raw signals from the denitrification reactor; S212, preprocesses the raw signal to form a standardized data stream; S213 performs weighted smoothing on historical standardized data to capture the overall trend of signal changes and short-term multi-stage predictions. S214, Calculate the ammonia injection adjustment amount based on the deviation between the predicted value and the target set value; S215, compares the predicted value with the actual measured value in real time and corrects the ammonia injection amount; S216, control the ammonia injection actuator to complete ammonia injection based on the corrected ammonia injection quantity; S217 allows for online updates and optimization of model parameters.

22. The intelligent control method for flue gas environmental protection island according to claim 16, characterized in that, The temperature field control sub-model of the denitrification reactor includes: S221, Collect sparse temperature measurement data of the denitrification reactor and perform data preprocessing; S222 performs spatial interpolation reconstruction on sparse temperature measurement data, and outputs the reconstructed temperature field, boundary gradient, and near-wall mask features. S223, based on the output of spatial interpolation reconstruction, combines working condition data, coal type information and actuator status data as input features to construct a spatiotemporal prediction model based on directional ConvLSTM. The model is trained with a boundary-first masking strategy and outputs the temperature field distribution, outlet NOx concentration and ammonia escape amount for multiple future steps. S224, based on the output of the prediction model, performs multi-objective model predictive control, constructs a multi-objective cost function, and optimizes the control quantity of the actuator of the denitrification reactor online; S225 performs time delay compensation on the optimized control quantity for the denitrification reactor actuator and issues it, sets back conditions, and automatically reverts to the original control strategy when the back conditions are met and continues to collect data.

23. The intelligent control method for flue gas environmental protection island according to claim 16, characterized in that, The dilution fan control sub-model includes: S231 collects and transmits denitrification operating parameters in real time, and preprocesses the denitrification operating parameters; S232 employs an improved dual-layer LSTM neural network to predict the outlet NOx concentration and optimal dilution air volume in future periods based on the denitrification operating parameters, and optimizes them through a process feature weighting mechanism, a cross-layer attention mechanism, and a multi-objective loss function. S233 employs a deep deterministic gradient reinforcement learning algorithm to output the target speed of the dilution fan based on the prediction results and the current operating conditions. S234 employs a fuzzy PID controller, which automatically takes over control when the decision module outputs an abnormality. S235, The fourth prediction module and the fourth decision module are simulated and verified under full working conditions based on the digital twin platform.

24. The intelligent control method for flue gas environmental protection island according to claim 17, characterized in that, The bag filter cleaning control sub-model includes: S301, collects and monitors the operating parameters of the bag filter in real time, including inlet and outlet pressure difference, flue gas flow rate, dust concentration, temperature and humidity; S302 is a mathematical model constructed based on the dynamic equation of filter bag pressure difference, used to describe the influence of dust accumulation process and pulse-jet cleaning behavior on the pressure difference change of internal filter bags in bag dust collectors. S303, based on a mathematical model and the current operating parameters, predicts the operating status of the bag filter over a period of time in the future; S304, Under the premise of satisfying the preset constraints, the optimal blowing strategy for a future period of time is solved by a multi-objective optimization algorithm. The blowing strategy includes blowing pressure, pulse width, blowing interval and timing arrangement. S305, based on the rolling time domain strategy, performs pulse-jet cleaning on the bag filter according to the instructions generated by the optimization decision module, and collects feedback information for optimization in the next cycle; S306 automatically corrects mathematical model parameters based on long-term operating data.

25. The intelligent control method for flue gas environmental protection island according to claim 17, characterized in that, The electrostatic field impedance matching sub-model includes: S311, real-time acquisition of operating parameters and slurry quality data; S312, perform data cleaning, feature extraction and operating condition prediction, and generate initial current and voltage setpoints; specifically, clean the collected data and extract a subset of features strongly correlated with current and voltage; based on the feature subset, predict the initial current and voltage setpoints through a generative adversarial network model. S313, The initial set value is dynamically corrected based on the slurry quality index model; S314 adjusts the high-frequency power supply output according to the corrected current and voltage settings.

26. The intelligent control method for flue gas environmental protection island according to claim 18, characterized in that, The gypsum dehydration control sub-model includes: S401, real-time acquisition of key operating parameters during the gypsum dehydration process, and preprocessing of the key operating parameters; S402, based on the preprocessed data, runs the mechanism model and the data-driven model in parallel, and dynamically calculates the fusion weight based on the real-time prediction error of the two models, and obtains the final predicted value of filter cake moisture content by weighted fusion. S403 constructs a weighted multi-objective optimization function based on constraints, uses model predictive control for real-time optimization, and employs a periodic evolutionary algorithm for global multi-objective optimization. Based on the fusion strategy, it generates the final control command and issues it to the corresponding actuator. S404: The actuator executes the issued gypsum dehydration control command, collects feedback data, performs real-time deviation correction, and forms a closed-loop control. S405 performs online parameter fine-tuning and periodic offline retraining of the data-driven model based on feedback data, and constructs anomaly detection and response strategies.

27. The intelligent control method for flue gas environmental protection island according to claim 18, characterized in that, The slurry pH control sub-model includes: S411, Collect and preprocess the absorption tower operation data; S412 uses gradient boosting trees and Gaussian process regression to construct a hybrid prediction model, outputting prediction results with uncertainty assessment; S413 uses a hybrid prediction model as the optimization objective function and employs a multi-objective Bayesian optimization algorithm to search for the Pareto optimal frontier. S414, the Pareto optimal frontier input adaptive model predictive control execution dynamically tracks the optimal pH setpoint and solves the optimal control sequence in a rolling manner; S415 sets safety constraints and a backoff mechanism. If the safety constraints are not met, the optimal control sequence output by the lower-level MPC module is rejected. If the safety constraints are met, the first element of the optimal control sequence is selected and sent to the corresponding actuator. If the backoff mechanism is not met, the control mode is automatically switched based on a conservative control strategy to ensure the safe and stable operation of the system.

28. The intelligent control method for flue gas environmental protection island according to claim 18, characterized in that, The circulating pump optimization control sub-model includes: S421, Collect absorption tower operation data and perform preprocessing; S422, based on a long short-term memory network, constructs an outlet SO2 concentration prediction model, using the system state within a sliding window as input, to make rolling predictions of SO2 concentration in future periods; S423, based on the predicted SO2 concentration at the outlet and the environmental emission standards, calculates the desulfurization efficiency and reverses it to calculate the minimum slurry flow rate required to meet the desulfurization efficiency requirements in the future period. S424 uses slurry flow rate demand as a hard constraint and power consumption of circulating pump group, limestone consumption and number of start-stop switching as optimization objectives to form a comprehensive cost function. It uses a genetic algorithm to quickly search the frequency-start-stop combination space and generate the most economical pump start-stop and frequency conversion strategy under the current working conditions in real time. S425 distributes the most economical pump start-up and shutdown and frequency conversion strategies to the distributed control system, continuously executes the data acquisition and preprocessing modules to the multi-objective optimization module, and sets a rollback mechanism.

29. The intelligent control method for flue gas environmental protection island according to claim 18, characterized in that, The slurry supply control sub-model includes: S431: Collect multimodal raw data and preprocess it to generate process variables, derivative features and multimodal features; S432 performs soft measurement and health estimation on multimodal characteristics, and transforms them into key process parameters and equipment status indicators. S433 uses a rolling prediction sequence and sensitivity of the absorber slurry pH, outlet sulfur dioxide concentration and supersaturation in the future time domain using an absorber mechanism model. S434, establish the actuator model based on the current actuator state, including the slurry flow rate model, pump power model and Jacobian matrix, and define the actuator constraints; S435, solve for the current control deviation, generate the feedforward control quantity based on the sensitivity mapping, and introduce oversaturation suppression and healthy amplitude limiting bias; S436, construct a multi-objective optimization function with slurry acidity / alkalinity and outlet sulfur dioxide concentration as the core, while taking into account energy consumption, material consumption, supersaturation and equipment health status, and solve the model predictive control based on actuator constraints to obtain the optimal control increment sequence; S437 maps the optimal control increment sequence or the feedforward reference quantity when optimization fails to the current control command and applies protective constraints based on oversaturation and equipment health score before issuing the command. S438 performs closed-loop feedback based on the execution results, records operation events, performs scene memory storage, and stores the current control scene and effect in the memory bank.

30. The intelligent control method for flue gas environmental protection island according to claim 19, characterized in that, The induced draft fan control model includes: S501, real-time acquisition of operating parameters of boiler and other furnace structure systems; S502, preprocesses the collected operating parameters; S503 uses a parallel prediction model to make multi-step predictions of furnace negative pressure and flue gas oxygen content based on preprocessed operating parameters. S504, based on the operation of the nitrogen oxide online meter, performs soft measurement of nitrogen oxide concentration and trend prediction of nitrogen oxide concentration to obtain the predicted value of nitrogen oxide concentration; S505, based on the static characteristic model of the fan, calculates the estimated air volume, the Jacobian matrix of air volume to actuator, and the estimated fan power according to the current speed and blade angle; S506, solves the predicted deviation between furnace negative pressure and flue gas oxygen content, is used to calculate the engineering feedforward component, and calculates the engineering feedforward air volume based on the engineering feedforward component; S507 constructs a quadratic programming problem with negative pressure deviation, oxygen content deviation, nitrogen oxide deviation and fan power as optimization objectives, and actuator limit, speed and surge risk as constraints. S508, solve the quadratic programming problem to obtain the optimal control increment, including the optimal control increment sequence for the induced draft fan speed and blade angle; S509, if the quadratic programming problem is solved successfully, the first term of the control increment sequence is taken as the current control command; if the solution fails or timeout occurs, the backoff control command is obtained by combining the engineering feedforward air volume value with the Jacobian matrix. S510 performs a security check on the current control command and then sends it to the induced draft fan actuator.

Images