Thermal power plant APC advanced process control system and denitration optimization method

Abstract

The invention relates to the technical field of thermal power plant flue gas denitration, in particular to a thermal power plant APC advanced process control system and a denitration optimization method.The system comprises a data acquisition module for acquiring boiler combustion parameters, flue gas emission data and the operation state of a denitration system in real time; the multivariable predictive control module is used for optimizing the combustion efficiency and the NOx generation amount based on a dynamic matrix control algorithm; the denitration optimization decision module is used for dynamically adjusting the ammonia spraying amount through an ammonia escape feedback model; the intelligent coordination module integrates a unit load instruction and an environmental protection constraint condition and is used for cooperative control of combustion-denitration; according to the system, boiler combustion parameters, flue gas emission data and the operation state of the denitration system are obtained in real time through the data acquisition module, a dynamic matrix control algorithm of the multivariable prediction control module is combined, the NOx generation trend can be predicted in advance, the combustion efficiency can be optimized, and therefore the problem that adjustment is lagged when loads fluctuate in traditional PID control is effectively solved.

Description

Technical Field

[0001] This invention relates to the field of flue gas denitrification technology in thermal power plants, and in particular to an advanced process control system (APC) for thermal power plants and a denitrification optimization method. Background Technology

[0002] During the operation of thermal power plants, nitrogen oxides (NOx) produced by combustion are one of the main air pollutants and need to be controlled through denitrification technologies such as selective catalytic reduction (SCR). Existing thermal power plants generally adopt distributed control systems (DCS) combined with conventional PID control strategies to adjust combustion and ammonia injection to achieve NOx emission standards.

[0003] Existing control methods often suffer from problems such as adjustment lag, excessive or insufficient ammonia injection when dealing with rapid changes in unit load or fluctuations in fuel characteristics. This can lead to instantaneous exceedances of NOx emissions or increased ammonia slip. When the load suddenly increases, the combustion conditions change drastically, and traditional PID control struggles to adjust the air-coal ratio and ammonia injection in a timely manner. This can cause a sharp rise in NOx concentration at the SCR inlet, and the delayed response of the ammonia injection system may lead to a short-term decrease in denitrification efficiency, thus affecting environmental indicators. Due to the lack of dynamic compensation for catalyst activity decay, after long-term operation, there is a tendency for excessive ammonia injection but still insufficient denitrification efficiency, which increases operating costs and may accelerate catalyst deactivation.

[0004] Therefore, to address the above problems, this invention proposes an advanced process control system (APC) for thermal power plants and a denitrification optimization method. Through dynamic modeling and intelligent decision-making, it achieves advanced adjustment of the air-coal ratio and ammonia injection quantity. By combining digital twin simulation and reinforcement learning algorithms, it improves the control robustness under variable load conditions, and ultimately reduces ammonia consumption while ensuring emission compliance. Summary of the Invention

[0005] To overcome the problem of lag in traditional PID control during load fluctuations, this invention proposes an advanced process control system (APC) for thermal power plants and a denitrification optimization method.

[0006] The technical solution of this invention is: an advanced process control system (APC) for thermal power plants, comprising: The data acquisition module acquires boiler combustion parameters, flue gas emission data, and denitrification system operating status in real time; The multivariable predictive control module optimizes combustion efficiency and NOx generation based on a dynamic matrix control algorithm; The denitrification optimization decision module dynamically adjusts the ammonia injection rate through an ammonia slip feedback model. The intelligent coordination module integrates unit load commands and environmental constraints for coordinated control of combustion and denitrification.

[0007] Preferably, the multivariate predictive control module includes a boiler-side sub-model and a load response sub-model. The boiler-side sub-model establishes a dynamic mapping relationship between the air-coal ratio, the secondary damper opening, and the NOx generation, and introduces a burner sway angle compensation coefficient to correct the influence of the furnace temperature field distribution on NOx. The load response sub-model trains a flue gas flow prediction function under load step changes based on historical operating data, updates model parameters using a sliding time window, and integrates real-time oxygen feedback signals to perform closed-loop correction of the air-coal ratio setpoint.

[0008] Preferably, the denitrification optimization decision module includes an ammonia slip monitoring unit and an ammonia injection valve fuzzy PID controller. The ammonia slip monitoring unit uses tunable laser absorption spectroscopy to detect the ammonia slip concentration at the SCR outlet in real time, and integrates the data with the NOx distribution field model in the DCS system to generate an ammonia injection demand index. The ammonia injection valve fuzzy PID controller divides the ammonia injection grid into multiple independent control areas according to the difference in NOx concentration distribution across the cross section of the SCR reactor. The opening degree of each area is adjusted by a fuzzy rule base. The input variables of the rule base include the catalyst temperature, flue gas velocity, and upstream NOx concentration gradient of the area.

[0009] Preferably, the system also includes a digital twin simulation unit and a dynamic constraint library. The digital twin simulation unit constructs a virtual SCR reactor through a three-dimensional CFD model to simulate the denitrification efficiency and ammonia escape trend under different ammonia injection strategies, and uses the deviation between the simulation results and the actual sensor data as an adaptive correction factor for the control strategy. The dynamic constraint library stores the SCR catalyst activity decay curve, the unit's environmental emission limits, and seasonal fuel sulfur content change data, and calls the constraint conditions in real time when the APC system generates control commands.

[0010] Preferably, the digital twin simulation unit uses a transfer learning algorithm to transfer the catalyst deactivation model parameters in the simulation environment to the actual control system. It calculates the difference between the simulation data and the real data through an online loss function. When the difference exceeds a threshold, it triggers a retraining process for the control parameters. At the same time, it uses reinforcement learning to optimize the ammonia injection sequence to minimize the weighted cost function of long-term ammonia consumption and NOx exceedance penalty.

[0011] Preferably, the data acquisition module communicates with the distributed control system through a wireless sensor network. The wireless transmission protocol uses time-synchronous channel frequency hopping technology to resist electromagnetic interference from the power plant. The data cleaning algorithm is based on isolated forest to detect outliers and combines Kalman filtering to smooth the flue gas flow signal. The cleaned data is stored in a time series database and labeled with confidence level for priority access by the multivariate predictive control module.

[0012] Preferably, the intelligent coordination module embeds feedforward-feedback composite control logic during the unit's load change phase. The feedforward branch predicts the NOx fluctuation trend based on the load command change rate and adjusts the ammonia injection benchmark value in advance. The feedback branch generates a compensation signal through the NOx concentration PID control at the SCR outlet. The outputs of the two are superimposed and then passed through a rate limiter, while a flue gas temperature change rate constraint is introduced.

[0013] As a preferred embodiment, a denitrification optimization method for an APC (Advanced Process Control System) in a thermal power plant includes the following steps: S1 collects boiler combustion parameters and NOx concentration at the SCR inlet in real time, and generates air-coal ratio optimization instructions through a multivariate prediction model; S2, based on ammonia slip rate and catalyst temperature field distribution, calculate the dynamic correction coefficient for ammonia injection rate; S3, combined with the unit load change trend, uses rolling time domain optimization to adjust the opening of the ammonia injection valve and the zoned flow of the ammonia injection grid.

[0014] Preferably, the calculation of the dynamic correction coefficient in step S2 includes predicting the current denitrification efficiency decline rate based on the catalyst activity detection model, the operating time, flue gas temperature, and ammonium sulfate deposition, and establishing a dynamic balance equation for the ammonia-nitrogen molar ratio by combining the real-time collected NOx concentration distribution data at the SCR inlet. With the ammonia escape concentration not exceeding 3 ppm as a hard constraint, the optimal ammonia injection rate for each ammonia injection zone is solved in reverse using the least squares method with boundary constraints. At the same time, a flue gas flow fluctuation compensation factor is introduced to smooth the calculation results to avoid ammonia injection rate oscillations caused by sudden load changes. Before being output to the actuator, the uniformity of the catalyst temperature field needs to be verified. If a local low temperature area is detected, the upper limit of the ammonia injection rate for the corresponding zone is automatically reduced.

[0015] As a preferred approach, a reinforcement learning algorithm is introduced to generate ammonia injection control strategies during the unit start-up and shutdown phases. First, a Q-learning model is constructed, with the action sequence of the ammonia injection valve as the decision variable and the weighted sum of ammonia consumption and NOx exceedance time as the reward function. In the initial stage of cold start-up, the ammonia injection amount is controlled by an initial strategy trained offline. After the flue gas temperature at the SCR inlet reaches the catalyst activation threshold, the model is switched to online learning mode. The real-time collected NOx removal rate and ammonia escape data are used to update the Q-value table. At the same time, an action exploration rate decay mechanism is set to make the control strategy gradually converge to the steady-state optimal solution. For the shutdown process, a preset ammonia injection decay curve is used, and the residual ammonia purging time is adjusted according to the shutdown duration and the interval of the next start-up.

[0016] The beneficial effects of this invention are: The data acquisition module acquires boiler combustion parameters, flue gas emission data, and the operating status of the denitrification system in real time. Combined with the dynamic matrix control algorithm of the multivariate predictive control module, it can predict NOx generation trends in advance and optimize combustion efficiency, thereby effectively solving the problem of adjustment lag in traditional PID control when the load fluctuates. At the same time, the denitrification optimization decision module dynamically adjusts the ammonia injection amount based on the ammonia slip feedback model, avoiding excessive or insufficient ammonia injection that leads to excessive emissions or increased ammonia slip. The intelligent coordination module integrates unit load commands and environmental constraints to achieve coordinated optimization control of combustion and denitrification, thereby ensuring that the unit can still stably meet environmental requirements under variable load conditions and significantly reducing ammonia consumption. Attached Figure Description

[0017] Figure 1 The diagram shown is a schematic representation of the system framework of this invention. Detailed Implementation

[0018] 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 with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of the present invention, but 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.

[0019] Please see Figure 1 This invention provides an embodiment: an advanced process control (APC) system for thermal power plants, comprising: The data acquisition module acquires boiler combustion parameters, flue gas emission data, and denitrification system operating status in real time; The multivariable predictive control module optimizes combustion efficiency and NOx generation based on a dynamic matrix control algorithm; The denitrification optimization decision module dynamically adjusts the ammonia injection rate through an ammonia slip feedback model. The intelligent coordination module integrates unit load commands and environmental constraints for coordinated control of combustion and denitrification.

[0020] Furthermore, the data acquisition module acquires boiler combustion parameters, flue gas emission data, and the operating status of the denitrification system in real time, and transmits the data to the multivariate predictive control module. This module analyzes the dynamic relationship between combustion efficiency and NOx generation based on a dynamic matrix control algorithm, and generates optimized control commands. At the same time, the denitrification optimization decision module receives ammonia slip data from the data acquisition module and dynamically adjusts the ammonia injection amount through a feedback model. The intelligent coordination module integrates unit load commands and environmental constraints to coordinate the operation of combustion control and the denitrification system, achieving optimized control throughout the entire process.

[0021] This invention significantly improves the control response speed under load fluctuation conditions through real-time data acquisition and multivariate predictive control, solving the problem of adjustment lag in traditional systems. The dynamic adjustment of ammonia injection based on ammonia escape feedback effectively avoids excessive or insufficient ammonia injection, ensuring emission compliance while reducing ammonia consumption. Intelligent coordinated control realizes the synergistic optimization of combustion and denitrification, enabling the system to improve overall operating efficiency while meeting environmental protection requirements.

[0022] The multivariate predictive control module includes a boiler-side sub-model and a load response sub-model. The boiler-side sub-model establishes a dynamic mapping relationship between the air-coal ratio, secondary damper opening, and NOx generation, and introduces a burner tilt angle compensation coefficient to correct the influence of the furnace temperature field distribution on NOx. The load response sub-model trains a flue gas flow prediction function under load step changes based on historical operating data, updates model parameters using a sliding time window, and integrates real-time oxygen feedback signals to perform closed-loop correction of the air-coal ratio setpoint.

[0023] Furthermore, the boiler-side sub-model uses real-time collected data on the air-coal ratio, secondary damper opening, and furnace temperature to establish a predictive model for NOx generation using a dynamic matrix control algorithm. This algorithm first constructs a dynamic matrix containing historical operating data, solves the optimal control law using the least squares method, and continuously optimizes the air-coal ratio setpoint for multiple future sampling times within each control cycle. The load response sub-model, based on a sliding time window mechanism, uses historical load change data of the unit to train an ARIMA time series prediction model, predicts the flue gas flow rate change trend in real time, and performs closed-loop correction of the air-coal ratio based on feedback from a real-time oxygen sensor. This invention significantly improves the response speed of the boiler combustion system to load changes through the multi-step prediction and rolling optimization characteristics of the dynamic matrix control algorithm, reducing NOx emission fluctuations by more than 30%. The use of the ARIMA model to predict flue gas flow rate combined with closed-loop oxygen correction improves the air-coal ratio control accuracy by 20%, effectively solving the problem of adjustment lag in traditional PID control under variable load conditions. The collaborative work of the two sub-models realizes feedforward-feedback composite control of the combustion process, significantly reducing the original NOx generation while ensuring boiler thermal efficiency.

[0024] The denitrification optimization decision module includes an ammonia slip monitoring unit and an ammonia injection valve fuzzy PID controller. The ammonia slip monitoring unit uses tunable laser absorption spectroscopy to detect the ammonia slip concentration at the SCR outlet in real time, and integrates the data with the NOx distribution field model in the DCS system to generate ammonia injection demand indicators. The ammonia injection valve fuzzy PID controller divides the ammonia injection grid into multiple independent control areas according to the differences in NOx concentration distribution across the cross section of the SCR reactor. The opening degree of each area is adjusted by a fuzzy rule base. The input variables of the rule base include the catalyst temperature, flue gas velocity, and upstream NOx concentration gradient of the area.

[0025] Furthermore, the ammonia slip monitoring unit acquires ammonia concentration distribution data at the outlet section of the SCR reactor in real time using tunable laser absorption spectroscopy (TDLAS) technology, and transmits the measurement results to the three-dimensional ammonia slip concentration field constructed by the central processor. The fuzzy PID controller of the ammonia injection valve first divides the cross section of the SCR reactor into grids. Each section establishes a local control model based on the upstream NOx concentration distribution, catalyst temperature field, and flue gas velocity field data. The input variables are processed by a fuzzy inference system (FIS), including the real-time NOx concentration gradient, temperature deviation, and flue gas residence time of the section. The adjustment coefficient of the ammonia injection amount of each section is calculated through a preset fuzzy rule base, and then the final control signal is output to the ammonia injection valve actuator of the section through a PID algorithm with dead zone.

[0026] The application of TDLAS technology enables real-time and accurate measurement of escaped ammonia concentration, with a measurement accuracy of ±0.5ppm, which is 5 times higher than that of traditional electrochemical sensors. The partitioned fuzzy PID control strategy improves the matching degree between ammonia injection rate and NOx distribution field by 40%, and reduces the peak concentration of ammonia escape to below 2ppm. Through multi-parameter coordinated control of temperature field and flue gas velocity field, the problem of local ammonium salt deposition is effectively avoided, and the catalyst life is extended by 20%. The entire system achieves dynamic and precise control of ammonia injection rate.

[0027] The system also includes a digital twin simulation unit and a dynamic constraint library. The digital twin simulation unit constructs a virtual SCR reactor through a three-dimensional CFD model to simulate the denitrification efficiency and ammonia escape trend under different ammonia injection strategies. The deviation between the simulation results and the actual sensor data is used as an adaptive correction factor for the control strategy. The dynamic constraint library stores the SCR catalyst activity decay curve, the unit's environmental emission limits, and seasonal fuel sulfur content change data. The constraint conditions are called in real time when the APC system generates control commands.

[0028] Furthermore, the digital twin simulation unit constructs a virtual three-dimensional power plant model using real-time collected DCS system operation data. It employs computational fluid dynamics (CFD) to simulate flue gas flow, ammonia diffusion, and chemical reaction processes within the SCR reactor. The simulation step size is set to 0.1 seconds. In each control cycle, the simulated NOx removal efficiency and ammonia slip concentration distribution are compared with actual sensor data. When the deviation exceeds 5%, a control strategy correction mechanism is triggered. The dynamic constraint library stores and updates SCR catalyst activity data in real time, including cumulative operating time, historical temperature curves, and ammonium sulfate deposition. Combined with environmental emission limits, dynamic constraints are generated. These constraints are then embedded into the optimization objective function of the APC system using a quadratic programming algorithm.

[0029] The digital twin simulation unit uses a transfer learning algorithm to transfer the catalyst deactivation model parameters from the simulation environment to the actual control system. It calculates the difference between the simulation data and the real data through an online loss function. When the difference exceeds a threshold, it triggers a retraining process for the control parameters. At the same time, it uses reinforcement learning to optimize the ammonia injection sequence to minimize the weighted cost function of long-term ammonia consumption and NOx exceedance penalties.

[0030] Furthermore, the transfer learning algorithm steps are as follows: First, a high-fidelity SCR reactor CFD simulation model is established in the offline stage. A training dataset covering different loads, fuel characteristics, and catalyst states is generated through millions of operating condition simulations. A catalyst deactivation prediction model is constructed using a deep neural network. In the online operation stage, the actual operating data is mapped to the feature space of the simulation model through the feature extraction layer. The difference index between the simulation prediction value and the actual sensor data is calculated in real time. When the difference exceeds the preset threshold, an online update mechanism for model parameters based on incremental learning is triggered. Only the weights of the last three layers of the network need to be updated to achieve rapid adaptation. At the same time, the dual-delay deep deterministic strategy gradient (TD3) algorithm is used to optimize the ammonia injection control strategy, using the long-term operating cost as the reward function, and storing the optimal control sequence through an experience replay buffer.

[0031] The data acquisition module communicates with the distributed control system through a wireless sensor network. The wireless transmission protocol uses time-synchronous channel frequency hopping technology to resist electromagnetic interference from the power plant. The data cleaning algorithm is based on isolated forest to detect outliers and combines Kalman filtering to smooth the flue gas flow signal. The cleaned data is stored in a time series database and labeled with confidence level for priority access by the multivariate predictive control module.

[0032] Furthermore, a wireless sensor node network deployed at key locations in the boiler flue and SCR reactor employs Time Synchronous Channel Frequency Hopping (TSCH) technology based on the IEEE 802.15.4e protocol for data acquisition and transmission. Each node is equipped with four types of sensors: temperature, pressure, NOx, and O2. Data is collected synchronously at a period of 250ms and transmitted to the gateway node via a frequency hopping sequence of six preset channels. Before entering the distributed control system (DCS), the raw data undergoes a three-level data processing flow. First, an improved isolated forest algorithm is used to detect outliers. By setting dynamic thresholds, abrupt changes in data caused by electromagnetic interference are automatically identified and removed. Second, a recursive least squares method with a forgetting factor is used to smooth and filter the flue gas flow signal. Finally, time series alignment is used to unify data from different sampling frequencies to a 100ms time base.

[0033] The intelligent coordination module embeds feedforward-feedback composite control logic during the unit's load change phase. The feedforward branch predicts the NOx fluctuation trend based on the load command change rate and adjusts the ammonia injection benchmark value in advance. The feedback branch generates a compensation signal through the PID control of the NOx concentration at the SCR outlet. The outputs of the two are superimposed and then passed through a rate limiter, while the flue gas temperature change rate constraint is introduced.

[0034] Furthermore, when a load change command signal is received, the feedforward branch immediately activates a load change rate prediction algorithm based on a dynamic differentiator. It uses a third-order Taylor expansion to predict the NOx generation trend within the next 30 seconds and generates a corresponding feedforward reference value for ammonia injection. Simultaneously, the feedback branch uses an improved PID algorithm with a Smith predictor. By comparing the deviation between the NOx concentration at the SCR outlet and the set value in real time, and combining it with the catalyst temperature change rate compensation factor, a feedback correction is generated. After the feedforward and feedback signals are weighted and fused, the final control command is output through a rate limiter. This limiter dynamically adjusts the maximum change rate according to the current operating status of the unit, and controls the ammonia injection valve action rate within a safe range of 0.5%-2% / second when the load changes drastically.

[0035] Furthermore, the denitrification optimization method of the present invention will be described in detail: By collecting boiler combustion parameters and SCR inlet NOx concentration in real time, a nonlinear relationship between air-coal ratio, oxygen content and NOx concentration is established through the boiler-side sub-model in the multivariate predictive control module. Based on the dynamic matrix control algorithm, an air-coal ratio optimization command is generated. At the same time, the load response sub-model is used to train the prediction function of load change on flue gas temperature through historical data, thereby realizing closed-loop optimization control of combustion conditions.

[0036] The ammonia slip monitoring unit uses tunable laser absorption spectroscopy to detect the ammonia slip concentration at the SCR outlet in real time. Combined with the virtual SCR reactor model established by the digital twin simulation unit, the current catalyst activity is predicted. The dynamic correction coefficient of each ammonia injection zone is calculated by the least squares method with boundary constraints. The optimal ammonia injection rate is solved in reverse with the constraint that the ammonia slip concentration does not exceed 3 ppm. After verification of the uniformity of the catalyst temperature field, the control command is output.

[0037] The intelligent coordination module integrates unit load commands and environmental constraints. When the load changes, the feedforward branch predicts the NOx fluctuation trend based on the dynamic differential algorithm and adjusts the ammonia injection standard value in advance. The feedback branch generates a compensation signal through PID control with a Smith predictor, and outputs the opening command of the zoned ammonia injection valve through the adaptive rate limiter, so as to realize precise control of the ammonia injection grid flow under variable load conditions.

[0038] During the unit start-up and shutdown phases, a Q-learning model with the ammonia injection valve action sequence as the decision variable is constructed using a reinforcement learning algorithm. The initial strategy is trained offline, and the control strategy is updated based on real-time NOx removal rate and ammonia escape data. The shutdown process is performed by using a preset ammonia injection rate decay curve and dynamically adjusting the residual ammonia purging time to optimize ammonia consumption under special operating conditions.

[0039] Furthermore, the present invention provides an embodiment: After implementing the APC system of the present invention in a 1000MW ultra-supercritical unit, the air-coal ratio is dynamically optimized through the multivariate predictive control module, so that the original NOx emission concentration of the boiler is stabilized at 180-220mg / Nm³ within the 70%-100% load range, which reduces the fluctuation range by 25% compared with the original PID control. At the same time, the denitrification optimization decision module improves the ammonia injection rate adjustment accuracy to ±1.5%, and controls the ammonia escape concentration below 2.1ppm, saving 86 tons of liquid ammonia consumption per year.

[0040] Furthermore, the present invention provides an embodiment: a power plant's peak-shaving unit uses the intelligent coordination module of the present invention to handle load changes of ±8% / minute. The feedforward-feedback composite control reduces the NOx concentration fluctuation range at the SCR outlet from ±25mg / Nm³ to ±8mg / Nm³, and the peak ammonia escape concentration decreases from 4.3ppm to 2.7ppm. While fulfilling the grid AGC command, the average monthly ammonia consumption is reduced by 9.2 tons, and the environmental assessment penalty is reduced to zero.

Claims

1. An advanced process control (APC) system for thermal power plants, characterized in that, Including: The data acquisition module acquires boiler combustion parameters, flue gas emission data, and denitrification system operating status in real time; The multivariable predictive control module optimizes combustion efficiency and NOx generation based on a dynamic matrix control algorithm; The denitrification optimization decision module dynamically adjusts the ammonia injection rate through an ammonia slip feedback model. The intelligent coordination module integrates unit load commands and environmental constraints for coordinated control of combustion and denitrification.

2. The advanced process control system (APC) for thermal power plants according to claim 1, characterized in that: The multivariate predictive control module includes a boiler-side sub-model and a load response sub-model. The boiler-side sub-model establishes a dynamic mapping relationship between the air-coal ratio, the secondary damper opening, and the NOx generation, and introduces a burner sway angle compensation coefficient to correct the influence of the furnace temperature field distribution on NOx. The load response sub-model trains a flue gas flow prediction function under load step changes based on historical operating data, updates model parameters using a sliding time window, and integrates real-time oxygen feedback signals to perform closed-loop correction of the air-coal ratio setpoint.

3. The advanced process control system (APC) for thermal power plants according to claim 1, characterized in that: The denitrification optimization decision module includes an ammonia slip monitoring unit and an ammonia injection valve fuzzy PID controller. The ammonia slip monitoring unit uses tunable laser absorption spectroscopy to detect the ammonia slip concentration at the SCR outlet in real time, and integrates the data with the NOx distribution field model in the DCS system to generate ammonia injection demand indicators. The ammonia injection valve fuzzy PID controller divides the ammonia injection grid into multiple independent control areas according to the differences in NOx concentration distribution across the cross section of the SCR reactor. The opening degree of each area is adjusted by a fuzzy rule base. The input variables of the rule base include the catalyst temperature, flue gas velocity, and upstream NOx concentration gradient of the area.

4. The advanced process control system (APC) for thermal power plants according to claim 1, characterized in that: The system also includes a digital twin simulation unit and a dynamic constraint library. The digital twin simulation unit constructs a virtual SCR reactor through a three-dimensional CFD model to simulate the denitrification efficiency and ammonia escape trend under different ammonia injection strategies. The deviation between the simulation results and the actual sensor data is used as an adaptive correction factor for the control strategy. The dynamic constraint library stores the SCR catalyst activity decay curve, the unit's environmental emission limits, and seasonal fuel sulfur content change data. The constraint conditions are called in real time when the APC system generates control commands.

5. The advanced process control system (APC) for thermal power plants according to claim 4, characterized in that: The digital twin simulation unit uses a transfer learning algorithm to transfer the catalyst deactivation model parameters from the simulation environment to the actual control system. It calculates the difference between the simulation data and the real data through an online loss function. When the difference exceeds a threshold, it triggers a retraining process for the control parameters. At the same time, it uses reinforcement learning to optimize the ammonia injection sequence to minimize the weighted cost function of long-term ammonia consumption and NOx exceedance penalties.

6. The advanced process control system (APC) for thermal power plants according to claim 1, characterized in that: The data acquisition module communicates with the distributed control system through a wireless sensor network. The wireless transmission protocol uses time-synchronous channel frequency hopping technology to resist electromagnetic interference from the power plant. The data cleaning algorithm is based on isolated forest to detect outliers and combines Kalman filtering to smooth the flue gas flow signal. The cleaned data is stored in a time series database and labeled with confidence level for priority access by the multivariate predictive control module.

7. The advanced process control system (APC) for thermal power plants according to claim 1, characterized in that: The intelligent coordination module embeds feedforward-feedback composite control logic during the unit's load change phase. The feedforward branch predicts the NOx fluctuation trend based on the load command change rate and adjusts the ammonia injection benchmark value in advance. The feedback branch generates a compensation signal through the PID control of the NOx concentration at the SCR outlet. The outputs of the two are superimposed and then passed through a rate limiter, while the flue gas temperature change rate constraint is introduced.

8. A denitrification optimization method for an APC (Advanced Process Control System) in a thermal power plant, employing the APC system described in claims 1-7, characterized in that... It includes the following steps: S1 collects boiler combustion parameters and NOx concentration at the SCR inlet in real time, and generates air-coal ratio optimization instructions through a multivariate prediction model; S2, based on ammonia slip rate and catalyst temperature field distribution, calculate the dynamic correction coefficient for ammonia injection rate; S3, combined with the unit load change trend, uses rolling time domain optimization to adjust the opening of the ammonia injection valve and the zoned flow of the ammonia injection grid.

9. The denitrification optimization method for an APC advanced process control system in a thermal power plant according to claim 8, characterized in that: The calculation of the dynamic correction coefficient in step S2 includes predicting the current denitrification efficiency decline rate based on the catalyst activity detection model, operating time, flue gas temperature, and ammonium sulfate deposition, and establishing a dynamic balance equation for the ammonia-nitrogen molar ratio by combining real-time collected NOx concentration distribution data at the SCR inlet. With the ammonia slip concentration not exceeding 3 ppm as a hard constraint, the optimal ammonia injection rate for each ammonia injection zone is solved in reverse using the least squares method with boundary constraints. At the same time, a flue gas flow fluctuation compensation factor is introduced to smooth the calculation results and avoid ammonia injection rate oscillations caused by sudden load changes. Before being output to the actuator, the uniformity of the catalyst temperature field needs to be verified. If a local low temperature area is detected, the upper limit of the ammonia injection rate for the corresponding zone is automatically reduced.

10. The denitrification optimization method for an APC advanced process control system in a thermal power plant according to claim 8, characterized in that: During the unit start-up and shutdown phases, a reinforcement learning algorithm is introduced to generate ammonia injection control strategies. First, a Q-learning model is constructed, with the action sequence of the ammonia injection valve as the decision variable and the weighted sum of ammonia consumption and NOx exceedance time as the reward function. In the initial stage of cold start-up, the ammonia injection amount is controlled by an initial strategy trained offline. After the flue gas temperature at the SCR inlet reaches the catalyst activation threshold, the model is switched to online learning mode. The real-time collected NOx removal rate and ammonia escape data are used to update the Q-value table. At the same time, an action exploration rate decay mechanism is set to make the control strategy gradually converge to the steady-state optimal solution. For the shutdown process, a preset ammonia injection decay curve is used, and the residual ammonia purging time is adjusted according to the shutdown duration and the interval of the next start-up.

Images