An ammonia injection control method, device and equipment of an SCR denitration system and a storage medium

Abstract

The application discloses an ammonia injection control method, device and equipment of an SCR denitration system and a storage medium, relates to the technical field of computers, and comprises the following steps: in a current control period, a target current process state is constructed based on a current process variable of an operation process of the SCR denitration system, an initial ammonia injection control amount, a current inlet disturbance variable and a historical sequence; a learning prior model is used to determine a prior probability distribution of a future ammonia injection action sequence and determine a candidate ammonia injection action sequence; an environment prediction model is used to determine a future process prediction trajectory; a trajectory cost value is determined based on a generated NOx control amount and by using a preset trajectory cost function, a posterior weight is determined by using a Boltzmann weight, the candidate ammonia injection action sequence is weighted, a target ammonia injection action is determined based on a obtained posterior action distribution, and ammonia injection control operation is performed on the SCR denitration system. In the premise of ensuring NOx standard emission and safe and smooth ammonia injection action, the real-time performance of ammonia injection control is improved.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and in particular to a method, apparatus, equipment and storage medium for controlling ammonia injection in an SCR denitrification system. Background Technology

[0002] Selective Catalytic Reduction (SCR) is a widely used flue gas denitrification process in industrial settings such as steel sintering, waste incineration, coal-fired boilers, and cement kilns. In SCR denitrification systems, the direct goal of ammonia injection control is to ensure that outlet NOx levels remain consistently below the emission limit, while avoiding problems such as ammonia escape, catalyst blockage, and increased operating costs caused by excessive ammonia injection.

[0003] Actual industrial sites place the following demands on SCR denitrification ammonia injection control: First, inlet NOx, flue gas load, and oxygen content fluctuate frequently, requiring the control system to be forward-looking; second, the impact of ammonia injection on outlet NOx has a significant lag, requiring the control system to avoid repeated ammonia additions due to a lack of immediate results; third, the site desires to achieve intelligent recommendation or closed-loop control of ammonia injection valve positions through a DCS (Distributed Control System) / MQTT (Message Queuing Telemetry Transport) gateway via a host computer, edge computing device, or industrial server without large-scale hardware modifications; fourth, the control system needs to have safety protection capabilities under abnormal operating conditions and cannot rely solely on black-box learning models for direct control.

[0004] Existing SCR denitrification ammonia injection control methods mainly include manual experience control, fixed rule control, PID (Proportional-Integral-Derivative) control, model predictive control, and data-driven control, but they have the following disadvantages: (1) Manual experience control and fixed rule control rely on operational experience and are difficult to adapt to complex operating conditions such as rapid fluctuations in inlet NOx, changes in flue gas load, and catalyst activity decay; (2) Traditional PID control mainly adjusts ammonia injection actions based on the current or short-term outlet NOx deviation, lacks the ability to predict future process trajectories, and is prone to adjustment lag, overshoot, and frequent oscillations; (3) Although traditional model predictive control can make multi-step predictions based on environmental models, it usually needs to search for action sequences in a large action space, resulting in low sampling efficiency, large online computation, and difficulty in adapting the initial sampling distribution to different operating conditions; (4) Although pure reinforcement learning or pure data-driven controllers have fast inference speed, the learning strategy usually directly outputs actions and lacks explicit constraint inference processes. When no operating conditions are seen, data is abnormal, model extrapolation occurs, or outlet NOx is close to the upper limit, unsafe actions may be output.

[0005] As can be seen from the above, improving the real-time performance of ammonia injection control while ensuring NOx emissions meet standards and ammonia injection is a pressing issue that needs to be addressed. Summary of the Invention

[0006] In view of this, the purpose of this invention is to provide an ammonia injection control method, apparatus, equipment, and storage medium for an SCR denitrification system, which can improve the real-time performance of ammonia injection control while ensuring NOx emission compliance and safe and stable ammonia injection operation. The specific solution is as follows: In a first aspect, this application provides a method for controlling ammonia injection in an SCR denitrification system, including: In the current control cycle, the target current process state is constructed based on the collected current process variables, initial ammonia injection control quantity, and current inlet disturbance variable of the SCR denitrification system during operation, as well as the historical sequences corresponding to the current process variables, the initial ammonia injection control quantity, and the current inlet disturbance variable. The current process variables include outlet converted NOx, outlet measured NOx, outlet oxygen, ammonia water header flow rate, flue gas temperature, flue gas flow rate, and catalyst bed temperature. The prior probability distribution of the ammonia injection action sequence in the future prediction time domain is determined by using the current process state of the target and a learned prior model. Based on the prior probability distribution, each candidate ammonia injection action sequence is determined. The learned prior model is a model determined based on a neural network, a time series model, an attention model, or a diffusion generation model. Based on the candidate ammonia injection action sequences and using an environmental prediction model for rolling prediction, the future process prediction trajectory of each candidate ammonia injection action sequence within the future prediction time domain is obtained; the environmental prediction model is a prediction model constructed based on a neural network model, a state-space model, or a gray box model. Based on the predicted trajectory of the future process, a corresponding NOx control quantity is generated. Based on the NOx control quantity and using a preset trajectory cost function, the trajectory cost value corresponding to each candidate ammonia injection action sequence is determined. Based on the trajectory cost value and using Boltzmann weights, a posterior weight is determined. The candidate ammonia injection action sequences are then weighted using the posterior weights to obtain a posterior action distribution. Based on the candidate ammonia injection action sequences in the posterior action distribution and using a preset projection operator, the target ammonia injection action of the current control cycle is determined, so as to perform ammonia injection control operation on the SCR denitrification system based on the target ammonia injection action.

[0007] Optionally, in the current control cycle, constructing the target current process state based on the collected current process variables, initial ammonia injection control quantity, and current inlet disturbance variable of the SCR denitrification system during operation, as well as the historical sequences corresponding to the current process variables, the initial ammonia injection control quantity, and the current inlet disturbance variable, includes: In the current control cycle, the current process state is obtained by splicing together the current process variables, initial ammonia injection control quantity, and current inlet disturbance variables of the SCR denitrification system during operation, as well as the historical sequences corresponding to the current process variables, the initial ammonia injection control quantity, and the current inlet disturbance variables; the current inlet disturbance variables include inlet NOx, inlet flue gas flow rate, inlet oxygen, and load change. The current process state is preprocessed to obtain the preprocessed current process state; the preprocessing includes timestamp alignment, outlier removal, missing value handling, data smoothing, unit unification, normalization or standardization, and historical sequence pruning. The data validity of the preprocessed current process state is judged. If the judgment result indicates that the preprocessed current process state is valid, the target current process state is determined based on the preprocessed current process state.

[0008] Optionally, the step of performing rolling prediction based on the candidate ammonia injection action sequences and using an environmental prediction model to obtain the future process prediction trajectory of each candidate ammonia injection action sequence within the future prediction time domain includes: Based on the candidate ammonia injection action sequence and the current process state of the target, and using an environmental prediction model for rolling prediction, the future process prediction trajectory of each candidate ammonia injection action sequence within the future prediction time domain is obtained; the future process prediction trajectory includes the outlet NOx prediction trajectory, the outlet oxygen prediction trajectory, and the ammonia water flow prediction trajectory.

[0009] Optionally, the step of generating corresponding NOx control variables based on the predicted trajectory of the future process, and determining the trajectory cost value corresponding to each of the candidate ammonia injection action sequences based on the NOx control variables and using a preset trajectory cost function, includes: Based on the predicted NOx value at the outlet, the predicted NOx value at the outlet, the predicted oxygen value at the outlet, and the measurement confidence factor in the predicted trajectory of the future process, the NOx control quantity is determined using a preset control quantity fusion function. Based on the deviation between the NOx control quantity and the NOx target value, the excess amount of the NOx control quantity relative to the preset safety upper limit, the action amplitude and action change of the candidate ammonia injection action sequence, and the out-of-bounds amount of auxiliary variables in the future process prediction trajectory, the trajectory cost value corresponding to each candidate ammonia injection action sequence is determined using the preset trajectory cost function; the auxiliary variables include outlet oxygen, ammonia water header flow rate, and flue gas temperature.

[0010] Optionally, determining the target ammonia injection action for the current control cycle based on each of the candidate ammonia injection action sequences in the posterior action distribution and using a preset projection operator includes: The weighted action is obtained by weighting the first step of the candidate ammonia spraying action sequence in the posterior action distribution and the corresponding posterior weight. The weighted action is projected using the preset projection operator to obtain the target ammonia injection action for the current control cycle.

[0011] Optional, also includes: The process involves determining a new initial ammonia injection control quantity using the target ammonia injection action of the current control cycle, designating the next control cycle as the new current control cycle, and then jumping to the step of constructing the target current process state based on the collected current process variables, initial ammonia injection control quantity, current inlet disturbance variable, and the historical sequences corresponding to the current process variables, initial ammonia injection control quantity, and current inlet disturbance variable during the operation of the SCR denitrification system. Based on the obtained target ammonia injection actions for each control cycle, the SCR denitrification system is subjected to rolling control operation of the ammonia injection process.

[0012] Optionally, after determining the target ammonia injection action for the current control cycle based on each of the candidate ammonia injection action sequences in the posterior action distribution and using a preset projection operator, the method further includes: The target current process state, the prior probability distribution, each candidate ammonia injection action sequence, the future process prediction trajectory, the trajectory cost, the posterior weight, the posterior action distribution, and the target ammonia injection action of the current control cycle are stored in the posterior distillation sample library. The distillation loss is determined based on the KL divergence between the posterior action distribution and the prior probability distribution in the posterior distillation sample library. The model parameters of the learned prior model are updated using the distillation loss to obtain the updated learned prior model. The updated learning prior model is validated based on evaluation metrics. If the validation results meet the target safety requirements, the updated learning prior model is used to replace the learning prior model. The evaluation metrics include NOx exceedance ratio, average ammonia injection amount, motion fluctuation degree, average trajectory cost, effective sampling number, and output stability.

[0013] Secondly, this application provides an ammonia injection control device for an SCR denitrification system, comprising: The target state construction module is used to construct the target current process state in the current control cycle based on the collected current process variables, initial ammonia injection control quantity, current inlet disturbance variable, and the historical sequences corresponding to the current process variables, the initial ammonia injection control quantity, and the current inlet disturbance variable during the operation of the SCR denitrification system. The current process variables include outlet converted NOx, outlet measured NOx, outlet oxygen, ammonia water header flow rate, flue gas temperature, flue gas flow rate, and catalyst bed temperature. The candidate sequence determination module is used to determine the prior probability distribution of the ammonia injection action sequence in the future prediction time domain by utilizing the current process state of the target and using a learned prior model, and to determine each candidate ammonia injection action sequence based on the prior probability distribution; the learned prior model is a model determined based on a neural network, a time series model, an attention model, or a diffusion generation model; The prediction trajectory determination module is used to perform rolling prediction based on the candidate ammonia injection action sequences and using an environmental prediction model to obtain the future process prediction trajectory of each candidate ammonia injection action sequence in the future prediction time domain; the environmental prediction model is a prediction model constructed based on a neural network model, a state-space model, or a gray box model. The candidate sequence weighting module is used to generate corresponding NOx control quantities based on the predicted trajectory of the future process, determine the trajectory cost value corresponding to each candidate ammonia injection action sequence based on the NOx control quantities and using a preset trajectory cost function, determine the posterior weight based on the trajectory cost value and using Boltzmann weights, and weight the candidate ammonia injection action sequences using the posterior weights to obtain the posterior action distribution. The target action determination module is used to determine the target ammonia injection action of the current control cycle based on each of the candidate ammonia injection action sequences in the posterior action distribution and using a preset projection operator, so as to perform ammonia injection control operation on the SCR denitrification system based on the target ammonia injection action.

[0014] Thirdly, this application provides an electronic device, comprising: Memory, used to store computer programs; A processor is used to execute the computer program to implement the aforementioned ammonia injection control method for the SCR denitrification system.

[0015] Fourthly, this application provides a computer-readable storage medium for storing a computer program, wherein the computer program, when executed by a processor, implements the aforementioned ammonia injection control method for an SCR denitrification system.

[0016] This application constructs a target current process state in the current control cycle based on the collected current process variables, initial ammonia injection control quantity, and current inlet disturbance variable of the SCR denitrification system during operation, as well as the historical sequences corresponding to the current process variables, the initial ammonia injection control quantity, and the current inlet disturbance variable. The current process variables include outlet converted NOx, outlet measured NOx, outlet oxygen, ammonia water header flow rate, flue gas temperature, flue gas flow rate, and catalyst bed temperature. Using the target current process state and a learned prior model, a prior probability distribution of the ammonia injection action sequence in the future prediction time domain is determined. Based on the prior probability distribution, each candidate ammonia injection action sequence is determined. The learned prior model is a model determined based on a neural network, time series model, attention model, or diffusion generation model. Based on the candidate ammonia injection action sequences and an environmental prediction model... The system performs rolling predictions to obtain the future process prediction trajectory of each candidate ammonia injection action sequence within the future prediction time domain. The environmental prediction model is a prediction model constructed based on a neural network model, a state-space model, or a gray-box model. A corresponding NOx control quantity is generated based on the future process prediction trajectory. Based on the NOx control quantity and a preset trajectory cost function, the trajectory cost value corresponding to each candidate ammonia injection action sequence is determined. Based on the trajectory cost value and Boltzmann weights, a posterior weight is determined. The candidate ammonia injection action sequences are then weighted using the posterior weights to obtain a posterior action distribution. Based on each candidate ammonia injection action sequence in the posterior action distribution and a preset projection operator, the target ammonia injection action for the current control cycle is determined. Based on the target ammonia injection action, the SCR denitrification system is subjected to ammonia injection control operations.

[0017] As can be seen from the above, this application integrates current process variables, ammonia injection control quantity, inlet disturbance, and their corresponding historical sequences to construct the current process state. It outputs the prior probability distribution of the ammonia injection action sequence through a prior model learning, and then generates candidate ammonia injection action sequences based on the prior probability distribution. These candidate action sequences are input into an environmental prediction model, which continuously outputs the future process prediction trajectory within the future prediction time domain. The environmental prediction model performs forward extrapolation of the future impact of each group of candidate actions, allowing for advance knowledge of the changing trends of outlet NOx, oxygen content, and ammonia flow rate under different ammonia injection strategies. Based on the predicted trajectory, a NOx control quantity is generated, and the trajectory cost is calculated using a preset cost function. Then, the prior distribution is corrected to a posterior action distribution using Boltzmann weights. Based on the posterior action distribution and a preset projection operator, the target ammonia injection action is determined. In this way, by controlling the ammonia injection process of the SCR denitrification system based on the target ammonia injection action, the real-time performance, economy, interpretability, and multi-condition adaptability of ammonia injection control are improved while ensuring NOx compliance and safe and stable ammonia injection action. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0019] Figure 1 This is a flowchart of an ammonia injection control method for an SCR denitrification system disclosed in this application; Figure 2 This is a schematic diagram of the ammonia injection control device of an SCR denitrification system disclosed in this application; Figure 3 This is a structural diagram of an electronic device disclosed in this application. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Currently, existing SCR denitrification ammonia injection control methods mainly include manual experience control, fixed rule control, PID control, model predictive control, and data-driven control. However, manual experience control and fixed rule control rely on operational experience and are difficult to adapt to complex operating conditions such as rapid fluctuations in inlet NOx, changes in flue gas load, and catalyst activity decay. Traditional PID control mainly adjusts ammonia injection based on current or short-term outlet NOx deviations, lacking the ability to predict future process trajectories, and is prone to regulation lag, overshoot, and frequent oscillations. Although traditional model predictive control can make multi-step predictions based on environmental models, it usually needs to search for action sequences within a large action space, resulting in low sampling efficiency, large online computation, and difficulty in adapting the initial sampling distribution to different operating conditions. Although pure reinforcement learning or pure data-driven controllers have fast inference speeds, the learning strategy usually directly outputs actions, lacking explicit constraints on the inference process. When no operating conditions are observed, data is abnormal, model extrapolation occurs, or outlet NOx is close to the upper limit, unsafe actions may be output. Therefore, this application provides an ammonia injection control method for an SCR denitrification system, which controls the ammonia injection process of the SCR denitrification system based on the target ammonia injection action, thereby improving the real-time performance, economy, interpretability and multi-condition adaptability of ammonia injection control while ensuring NOx compliance and safe and stable ammonia injection action.

[0022] See Figure 1As shown in the figure, an embodiment of the present invention discloses a method for controlling ammonia injection in an SCR denitrification system, comprising: Step S11: In the current control cycle, construct the target current process state based on the collected current process variables, initial ammonia injection control quantity, current inlet disturbance variable, and the historical sequences corresponding to the current process variables, the initial ammonia injection control quantity, and the current inlet disturbance variable during the operation of the SCR denitrification system; the current process variables include outlet converted NOx, outlet measured NOx, outlet oxygen, ammonia water header flow rate, flue gas temperature, flue gas flow rate, and catalyst bed temperature.

[0023] In this embodiment, the current process variables, control variables, inlet disturbance variables, and their historical sequences of the SCR denitrification process are collected. The current process variables may include outlet converted NOx, outlet measured NOx, outlet oxygen, ammonia header flow rate, flue gas temperature, flue gas flow rate, catalyst bed temperature, etc.; control variables may include ammonia injection valve position, ammonia injection flow rate, ammonia pump frequency, etc.; inlet disturbance variables may include inlet NOx, inlet flue gas flow rate, inlet oxygen, load changes, etc. During the control period k, the current process variables can be denoted as... Record the control quantities executed in the previous cycle as The current ingress disturbance variable is denoted as The historical process variable sequence, the historical ammonia injection control quantity sequence (i.e., the historical control action sequence), and the historical inlet disturbance sequence are respectively denoted as... , and The current process state is constructed based on the above variables. Historical sequences are used to characterize the large inertia, large hysteresis, and temporal coupling characteristics of the SCR denitrification ammonia injection process, enabling the controller to make decisions not only based on the current NOx deviation, but also considering the historical trends of inlet disturbances, ammonia injection actions, and outlet responses.

[0024] It should be noted that the initial ammonia injection control quantity in the embodiments of the present invention is the same as that in the above formula. In the first control cycle, the initial ammonia injection control quantity is taken from the current actual ammonia injection valve position, the feedback value of the ammonia injection actuator, or the manually set value. In subsequent control cycles, the initial ammonia injection control quantity is taken from the target ammonia injection action determined and written into the ammonia injection actuator in the previous control cycle, or from the action feedback value of the ammonia injection actuator for the target ammonia injection action.

[0025] It is understood that the current process state is preprocessed, including timestamp alignment, outlier removal, missing value handling, data smoothing, unit unification, normalization or standardization, historical sequence pruning, and data validity assessment. If key measurement points are abnormal, communication is interrupted, actuator feedback is abnormal, or historical sequences are insufficient, the SCR denitrification system enters a safety protection mode; if the data validity assessment result indicates that the data is valid, the target current process state is determined based on the preprocessed data.

[0026] Specifically, in the current control cycle, constructing the target current process state based on the collected current process variables, initial ammonia injection control quantity, current inlet disturbance variable, and the historical sequences corresponding to the current process variables, initial ammonia injection control quantity, and current inlet disturbance variable during the operation of the SCR denitrification system includes: in the current control cycle, constructing the target current process state based on the collected current process variables, initial ammonia injection control quantity, current inlet disturbance variable, and the historical sequences corresponding to the current process variables, initial ammonia injection control quantity, and current inlet disturbance variable during the operation of the SCR denitrification system. The data is spliced ​​together to obtain the current process state; the current inlet disturbance variables include inlet NOx, inlet flue gas flow rate, inlet oxygen, and load change; the current process state is preprocessed to obtain the preprocessed current process state; the preprocessing includes timestamp alignment, outlier removal, missing value handling, data smoothing, unit unification, normalization or standardization, and historical sequence pruning; the preprocessed current process state is used to determine data validity, and if the determination result indicates that the preprocessed current process state is valid, the target current process state is determined based on the preprocessed current process state.

[0027] Step S12: Utilize the current process state of the target and the learned prior model to determine the prior probability distribution of the ammonia injection action sequence in the future prediction time domain, and determine each candidate ammonia injection action sequence based on the prior probability distribution; the learned prior model is a model determined based on a neural network, a time series model, an attention model, or a diffusion generation model.

[0028] In this embodiment, the role of the prior model is to output the prior probability distribution of the future ammonia injection action sequence based on the current process state. The future ammonia injection action sequence is defined as follows: This refers to the ammonia injection control amount at time step T from the current cycle. The learned prior model can output a single-peak distribution or a multi-modal mixed distribution based on the current operating conditions to represent various reasonable operating modes such as maintaining ammonia injection, slow ammonia addition, rapid ammonia replenishment, and slow ammonia reduction. Its general form is: ; in, In the current process state Future ammonia injection sequence under the conditions The prior probability distribution; These are the model parameters of the prior learning model; Number of action patterns; Let m be the weight of the m-th action pattern; , These are the mean and covariance of the action sequence distribution under this action pattern, respectively. When When the learning prior model is 1, it degenerates into a unimodal action prior distribution. It should be noted that the learning prior model can be a neural network, a temporal model, an attention model, a diffusion generation model, or other models capable of outputting probability distribution parameters. This learning prior model ensures that candidate action sequences preferentially concentrate in regions considered reasonable based on historical experience.

[0029] Step S13: Based on the candidate ammonia injection action sequence and using the environmental prediction model, perform rolling prediction to obtain the future process prediction trajectory of each candidate ammonia injection action sequence in the future prediction time domain; the environmental prediction model is a prediction model constructed based on a neural network model, a state-space model, or a gray box model.

[0030] In this embodiment, rolling prediction is performed based on each candidate ammonia injection action sequence and by calling the environmental prediction model to obtain the predicted trajectory of the future process corresponding to the action sequence. The corresponding formula is as follows: ; in, Predict the trajectory for the future process; The environmental prediction model; This represents the current process state; This is the sequence of candidate ammonia injection actions for the i-th group; This represents the total number of candidate ammonia injection action sequences. The environmental prediction model may employ a neural network model, a state-space model, a mechanism-data fusion model, a gray box model, or other prediction models, without limiting the specific model structure.

[0031] Specifically, the step of performing rolling predictions based on the candidate ammonia injection action sequences and using an environmental prediction model to obtain the future process prediction trajectory of each candidate ammonia injection action sequence within the future prediction time domain includes: performing rolling predictions based on the candidate ammonia injection action sequences and the target current process state, and using an environmental prediction model to obtain the future process prediction trajectory of each candidate ammonia injection action sequence within the future prediction time domain; the future process prediction trajectory includes the outlet NOx prediction trajectory, the outlet oxygen prediction trajectory, and the ammonia water flow prediction trajectory.

[0032] Step S14: Generate corresponding NOx control quantities based on the predicted trajectory of the future process; determine the trajectory cost value corresponding to each candidate ammonia injection action sequence based on the NOx control quantities and using a preset trajectory cost function; determine the posterior weight based on the trajectory cost value and using Boltzmann weights; and weight the candidate ammonia injection action sequences using the posterior weights to obtain the posterior action distribution.

[0033] In this embodiment, the NOx control quantity for cost evaluation is determined based on the predicted future process trajectory. The NOx control quantity can be comprehensively determined by the export-converted NOx, the export-measured NOx and its reliability, the safety margin, the degree of signal divergence, or the operating condition type, as shown in the following formula: ; in, The NOx control quantity at the h-th prediction time step for the future process prediction trajectory of the i-th candidate ammonia injection action sequence; This is a function for fusion of control quantities; Calculate the NOx prediction value at the outlet of the i-th candidate ammonia injection action sequence at the h-th prediction time step for the future process prediction trajectory; The predicted NOx value at the outlet measured at the h-th prediction time step is the future process prediction trajectory of the i-th candidate ammonia injection action sequence. The predicted value of outlet oxygen at the h-th prediction time step is given by the future process prediction trajectory of the i-th candidate ammonia injection action sequence. This is used to measure the reliability factor or operating condition risk factor. In one specific implementation, the NOx control quantity can be selected as the one with higher risk between the converted NOx and the measured NOx.

[0034] It is understood that the corresponding trajectory cost is determined based on the predicted future process trajectory corresponding to each candidate ammonia injection action sequence. The trajectory cost simultaneously considers NOx compliance, ammonia injection economy, action stability, and auxiliary variable constraints, and the corresponding formula is as follows: ; in, Let the trajectory value be the corresponding value of the i-th candidate ammonia injection action sequence; The prediction weight is the prediction weight for the h-th prediction time step; The weighting coefficients for the NOx target deviation term; The deviation between the NOx control quantity and the target value at the h-th prediction time step for the future process prediction trajectory of the i-th candidate ammonia injection action sequence. , For NOx control; The target value; The weighting coefficient for the NOx over-limit penalty term; The safe limit for NOx; This indicates that only the portion exceeding the limit will be penalized; The weighting coefficients for the out-of-bounds penalty term for auxiliary variables; For the auxiliary variable out-of-bounds penalty function, The auxiliary variables for the future process prediction trajectory of the i-th candidate ammonia injection action sequence at the h-th prediction time step include outlet oxygen, ammonia water header flow rate, flue gas temperature, etc. The weighting coefficient for the ammonia injection amplitude penalty term; Used to suppress long-term excessive ammonia spraying This is the sequence of candidate ammonia injection actions for the i-th group; The weighting coefficient for the penalty item for action change; Used to suppress action mutations and action sequence oscillations Let be the change in action of the i-th candidate ammonia injection action sequence.

[0035] Furthermore, the aforementioned preset trajectory cost function not only considers the NOx target deviation at the outlet, but also simultaneously considers penalties for NOx exceeding limits, ammonia injection amplitude, action variation, action sequence smoothing, and out-of-bounds penalties for auxiliary variables such as outlet oxygen, ammonia water header flow rate, and actuator status. Under low NOx conditions, the cost function can increase the penalties for ammonia injection amplitude and action variation to prevent over-injection; when NOx approaches the emission limit, the cost function can increase the weight of the exceeding penalty to prioritize ensuring compliance with emission standards; when inlet disturbances are severe or model reliability decreases, the action smoothing penalty can be increased to put the controller into a conservative control state.

[0036] Specifically, the step of generating corresponding NOx control quantities based on the predicted future process trajectory, and determining the trajectory cost value corresponding to each candidate ammonia injection action sequence based on the NOx control quantities and using a preset trajectory cost function, includes: determining the NOx control quantity based on the outlet converted NOx prediction value, the outlet measured NOx prediction value, the outlet oxygen prediction value, and the measurement confidence factor in the predicted future process trajectory, and using a preset control quantity fusion function; determining the trajectory cost value corresponding to each candidate ammonia injection action sequence based on the deviation between the NOx control quantity and the NOx target value, the excess amount of the NOx control quantity relative to the preset safety upper limit, the action amplitude and action change amount of the candidate ammonia injection action sequence, and the out-of-bounds amount of auxiliary variables in the predicted future process trajectory; the auxiliary variables include outlet oxygen, ammonia water header flow rate, and flue gas temperature.

[0037] In this embodiment, the posterior weight is determined based on the trajectory cost value corresponding to each candidate ammonia injection action sequence. Boltzmann weights can be used here to give low-cost action sequences a higher probability and high-cost action sequences a lower probability. The corresponding formula is as follows: ; in, Let be the posterior weight of the i-th candidate ammonia injection action sequence; Let the trajectory value be the corresponding value of the i-th candidate ammonia injection action sequence; Let the trajectory value be the j-th candidate ammonia injection action sequence; The minimum trajectory cost among the candidate action sequences; For temperature parameters, When the posterior distribution is smaller, it is more concentrated on low-cost actions, and the control is more aggressive. When the posterior distribution is larger, it is closer to the learned prior, resulting in more stable control. The total number of candidate ammonia injection action sequences.

[0038] Understandably, the prior probability distribution is updated to the posterior action distribution based on the obtained posterior weights, and the corresponding formula is as follows: ; in, Current process state Future ammonia injection sequence under the conditions posterior action distribution; Current process state Future ammonia injection sequence under the conditions The trajectory of value; For temperature parameters; In the current process state Future ammonia injection sequence under the conditions The prior probability distribution; These are the model parameters of the prior learning model; It indicates that they are directly proportional.

[0039] Step S15: Based on each of the candidate ammonia injection action sequences in the posterior action distribution and using a preset projection operator, determine the target ammonia injection action for the current control cycle, and perform ammonia injection control operation on the SCR denitrification system based on the target ammonia injection action.

[0040] In this embodiment, under the sampling approximation condition, the posterior action distribution can be expressed as a weighted empirical distribution of the candidate ammonia spraying action sequence, with the corresponding formula as follows: ; in, Current process state Future ammonia injection sequence under the conditions posterior action distribution; Let be the posterior weight of the i-th candidate ammonia injection action sequence; It is the Dirac function; This is the sequence of candidate ammonia injection actions for the i-th group; The total number of candidate ammonia injection action sequences.

[0041] It is understandable that the final target ammonia injection action is generated by the posterior action distribution, taking the weighted average of the first action of each candidate ammonia injection action sequence, and then subjecting it to safe projection or amplitude limiting processing, as shown in the following formula: ; in, To control the target ammonia injection action in cycle k; The safety projection operator is used to ensure that the final action meets the upper and lower limits of the ammonia injection actuator, the limit of the range of action change, and the on-site safety protection conditions. Let be the posterior weight of the i-th candidate ammonia injection action sequence; This is the first action of the i-th candidate ammonia injection action sequence; The total number of candidate ammonia injection action sequences.

[0042] Specifically, determining the target ammonia injection action for the current control period based on each candidate ammonia injection action sequence in the posterior action distribution and using a preset projection operator includes: weighting the first step action of the candidate ammonia injection action sequence in the posterior action distribution and the corresponding posterior weight to obtain a weighted action; and projecting the weighted action using the preset projection operator to obtain the target ammonia injection action for the current control period.

[0043] It is understood that the target ammonia injection action is written into the field actuator; in the next control cycle, a new initial ammonia injection control quantity is determined based on the target ammonia injection action, the next control cycle is determined as the new current control cycle, field data is collected again and the above process is repeated, thereby realizing the rolling control operation of the ammonia injection process of the SCR denitrification system based on the target ammonia injection action of each control cycle. Specifically, it also includes: using the target ammonia injection action of the current control cycle to determine a new initial ammonia injection control quantity, determining the next control cycle as the new current control cycle, and then jumping to the step of constructing the target current process state based on the collected current process variables, initial ammonia injection control quantity, current inlet disturbance variable, and the historical sequences corresponding to the current process variables, initial ammonia injection control quantity, and current inlet disturbance variable during the operation of the SCR denitrification system, and performing rolling control operation of the ammonia injection process of the SCR denitrification system based on the obtained target ammonia injection actions of each control cycle.

[0044] In this embodiment, to avoid the posterior weights being too concentrated or too dispersed, the effective sample number can be calculated to evaluate the contribution of the candidate action sequence to the posterior action distribution. The corresponding formula is as follows: ; in, The number of valid samples; Let be the posterior weight of the i-th candidate ammonia injection action sequence; This represents the total number of candidate ammonia injection action sequences. When the number of effective samples is too low, it indicates that a few candidate ammonia injection actions occupy the majority of the weight. In this case, the temperature parameter can be increased, the variance of the prior probability distribution can be expanded, the number of candidate samples can be increased, or a conservative control mode can be entered. The temperature parameter can be adaptively adjusted according to the model error, NOx safety margin, and sampling degradation degree, and the corresponding formula is as follows: ; in, The temperature parameter is used to control the cycle k; This is a limiting function; These are the initial temperature parameters; , , This is the adjustment coefficient; To control the uncertainty or error level of the model prediction for period k; To control the NOx safety margin in period k; To control the degree of sampling degradation in period k; , These represent the lower and upper limits of the temperature parameter, respectively. This mechanism prioritizes cost optimization when the risk of failing to meet the target is high, and emphasizes smooth operation and prior experience when the system is running smoothly, thereby improving the controller's adaptability under different risk levels.

[0045] Understandably, at the end of each control cycle, the system stores the current process state, prior probability distribution, candidate action sequence, environmental predicted trajectory, trajectory cost, posterior weight, posterior action distribution, target execution action, and on-site feedback results into the posterior distillation sample library. Samples with sensor malfunctions, manual intervention, communication interruptions, or actuator malfunctions are not included in training. The goal of posterior distillation is to make the prior probability distribution output by the learned prior model gradually approximate the posterior action distribution obtained by the probabilistic inference model predicting the controller. The corresponding formula is as follows: ; in, This is due to distillation losses; These are the model parameters of the prior learning model; Let KL divergence be a metric. Current process state Future ammonia injection sequence under the conditions posterior action distribution; Let be the prior probability distribution. In the sampling form, a weighted negative log-likelihood loss can also be used, with the corresponding formula as follows: ; in, This is due to distillation losses; Let be the posterior weight of the i-th candidate ammonia injection action sequence; This is the sequence of candidate ammonia injection actions for the i-th group; This represents the current process state; The total number of candidate ammonia injection action sequences; To learn the probability density of the prior model output for the i-th candidate action sequence; For safety regularization, it is used to penalize samples in the output distribution of the prior learning model that may lead to actions going out of bounds, action mutations, continued large-scale addition of ammonia when NOx is low, or large-scale reduction of ammonia when NOx is close to the upper limit. This refers to the weights of the safety regularization term. Before the new version of the learning prior model goes online, it needs to be evaluated on the validation dataset for NOx exceedance rate, average ammonia injection rate, action fluctuation degree, average trajectory cost, effective sampling number, and output stability. Only when it meets the safety requirements and is superior to the current version will the online model be replaced. In addition, the online running data can be classified and stored according to types such as low NOx operating conditions, normal operating conditions, near-upper limit operating conditions, high NOx operating conditions, inlet NOx fluctuation operating conditions, and oxygen fluctuation operating conditions, and used for posterior distillation training, while retaining the model fallback mechanism.

[0046] Specifically, after determining the target ammonia injection action for the current control period based on the candidate ammonia injection action sequences in the posterior action distribution and using a preset projection operator, the method further includes: storing the target current process state, the prior probability distribution, the candidate ammonia injection action sequences, the future process prediction trajectory, the trajectory cost, the posterior weight, the posterior action distribution, and the target ammonia injection action in a posterior distillation sample library; determining the distillation loss based on the KL divergence between the posterior action distribution and the prior probability distribution in the posterior distillation sample library, and updating the model parameters of the learned prior model using the distillation loss to obtain an updated learned prior model; verifying the updated learned prior model based on evaluation metrics, and if the verification results meet the target safety requirements, replacing the learned prior model with the updated learned prior model; the evaluation metrics include NOx exceedance ratio, average ammonia injection amount, action fluctuation degree, average trajectory cost, effective sampling number, and output stability.

[0047] Furthermore, this method supports recommended and automatic modes. In recommended mode, the system only outputs recommended actions, predicted trajectories, and control reasons, which are then executed after confirmation by the operator. In automatic mode, the system automatically writes the final ammonia injection action when the data is valid, the model is reliable, and the candidate actions meet the constraints. Furthermore, normal optimization mode, conservative control mode, and safety protection mode can be set. In normal optimization mode, the system executes the final action generated by the posterior distribution of probabilistic inference. In conservative control mode, when the inlet disturbance is severe, the model prediction error increases, the posterior distribution uncertainty is high, or key variables are close to the boundary, the system improves the smoothness of the action and limits the action change. In safety protection mode, when key measuring points are abnormal, communication is interrupted, the actuator is abnormal, the candidate action is infeasible, or the model output is abnormal, the system retains the action of the previous cycle, reverts to the manually set value, outputs a preset safety action, or stops automatic writing. This solution can be deployed in industrial servers, edge computing devices, or host computers, reading real-time data from DCS, PLC, data gateways, or industrial databases through communication interfaces, and outputting ammonia injection valve position, ammonia injection flow setpoint, ammonia pump frequency, or other ammonia injection actuator control quantities. The system can first run in the recommended mode, which can be confirmed by the operator before execution; or it can switch to automatic mode after the safety conditions are met. It does not require large-scale modification of the on-site hardware and is easy to implement in the existing SCR denitrification system.

[0048] As can be seen from the above, this application integrates current process variables, ammonia injection control quantity, inlet disturbance, and their corresponding historical sequences to construct the current process state. It outputs the prior probability distribution of the ammonia injection action sequence through a prior model learning, and then generates candidate ammonia injection action sequences based on the prior probability distribution. These candidate action sequences are input into an environmental prediction model, which continuously outputs the future process prediction trajectory within the future prediction time domain. The environmental prediction model performs forward extrapolation of the future impact of each group of candidate actions, allowing for advance knowledge of the changing trends of outlet NOx, oxygen content, and ammonia flow rate under different ammonia injection strategies. Based on the predicted trajectory, a NOx control quantity is generated, and the trajectory cost is calculated using a preset cost function. Then, the prior distribution is corrected to a posterior action distribution using Boltzmann weights. Based on the posterior action distribution and a preset projection operator, the target ammonia injection action is determined. In this way, by controlling the ammonia injection process of the SCR denitrification system based on the target ammonia injection action, the real-time performance, economy, interpretability, and multi-condition adaptability of ammonia injection control are improved while ensuring NOx compliance and safe and stable ammonia injection action.

[0049] Accordingly, see Figure 2 As shown, this application also provides an ammonia injection control device for an SCR denitrification system, comprising: The target state construction module 11 is used to construct the target current process state in the current control cycle based on the collected current process variables, initial ammonia injection control quantity, current inlet disturbance variable, and the historical sequences corresponding to the current process variables, the initial ammonia injection control quantity, and the current inlet disturbance variable during the operation of the SCR denitrification system. The current process variables include outlet converted NOx, outlet measured NOx, outlet oxygen, ammonia water header flow rate, flue gas temperature, flue gas flow rate, and catalyst bed temperature. The candidate sequence determination module 12 is used to determine the prior probability distribution of the ammonia injection action sequence in the future prediction time domain by using the current process state of the target and the learning prior model, and to determine each candidate ammonia injection action sequence based on the prior probability distribution; the learning prior model is a model determined based on a neural network, a time series model, an attention model or a diffusion generation model; The prediction trajectory determination module 13 is used to perform rolling prediction based on the candidate ammonia injection action sequence and using an environmental prediction model to obtain the future process prediction trajectory of each candidate ammonia injection action sequence in the future prediction time domain; the environmental prediction model is a prediction model constructed based on a neural network model, a state-space model, or a gray box model. The candidate sequence weighting module 14 is used to generate corresponding NOx control quantities based on the predicted trajectory of the future process, determine the trajectory cost value corresponding to each candidate ammonia injection action sequence based on the NOx control quantities and using a preset trajectory cost function, determine the posterior weight based on the trajectory cost value and using Boltzmann weights, and weight the candidate ammonia injection action sequences using the posterior weights to obtain the posterior action distribution. The target action determination module 15 is used to determine the target ammonia injection action of the current control cycle based on each of the candidate ammonia injection action sequences in the posterior action distribution and using a preset projection operator, so as to perform ammonia injection control operation on the SCR denitrification system based on the target ammonia injection action.

[0050] In some specific embodiments, the target state construction module 11 may specifically include: The sequence splicing unit is used, in the current control cycle, to splice together the current process variables, initial ammonia injection control quantity, current inlet disturbance variable, and the historical sequences corresponding to the current process variables, initial ammonia injection control quantity, and current inlet disturbance variable during the operation of the SCR denitrification system, so as to obtain the current process state; the current inlet disturbance variable includes inlet NOx, inlet flue gas flow rate, inlet oxygen, and load change; A state preprocessing unit is used to preprocess the current process state to obtain the preprocessed current process state; the preprocessing includes timestamp alignment, outlier removal, missing value handling, data smoothing, unit unification, normalization or standardization, and historical sequence pruning. The validity judgment unit is used to judge the validity of the preprocessed current process state. If the judgment result indicates that the preprocessed current process state is valid, the target current process state is determined based on the preprocessed current process state.

[0051] In some specific embodiments, the predicted trajectory determination module 13 may specifically include: The trajectory prediction unit is used to perform rolling prediction based on the candidate ammonia injection action sequence and the current process state of the target, using an environmental prediction model, to obtain the future process prediction trajectory of each candidate ammonia injection action sequence in the future prediction time domain; the future process prediction trajectory includes the outlet NOx prediction trajectory, the outlet oxygen prediction trajectory, and the ammonia water flow prediction trajectory.

[0052] In some specific embodiments, the candidate sequence weighting module 14 may specifically include: The control quantity determination unit is used to calculate the NOx prediction value at the outlet, the measured NOx prediction value at the outlet, the predicted oxygen value at the outlet, and the measurement confidence factor based on the predicted trajectory of the future process, and to determine the NOx control quantity using a preset control quantity fusion function. The cost-value determination unit is used to determine the trajectory cost value corresponding to each of the candidate ammonia injection action sequences based on the deviation between the NOx control quantity and the NOx target value, the excess amount of the NOx control quantity relative to the preset safety upper limit, the action amplitude and action change amount of the candidate ammonia injection action sequence, and the out-of-bounds amount of auxiliary variables in the future process prediction trajectory, and using the preset trajectory cost function; the auxiliary variables include outlet oxygen, ammonia water header flow rate, and flue gas temperature.

[0053] In some specific embodiments, the target action determination module 15 may specifically include: An action weighting unit is used to weight the first step of the candidate ammonia spraying action sequence in the posterior action distribution and the corresponding posterior weight to obtain a weighted action. An action projection unit is used to project the weighted action using the preset projection operator to obtain the target ammonia injection action of the current control cycle.

[0054] In some specific embodiments, the ammonia injection control device of the SCR denitrification system may further include: The rolling control unit is used to determine a new initial ammonia injection control quantity using the target ammonia injection action of the current control cycle, determine the next control cycle as the new current control cycle, and then jump to the step of constructing the target current process state based on the collected current process variables, initial ammonia injection control quantity, current inlet disturbance variable, and the historical sequences corresponding to the current process variables, initial ammonia injection control quantity, and current inlet disturbance variable during the operation of the SCR denitrification system, and performs rolling control operation of the ammonia injection process on the SCR denitrification system based on the obtained target ammonia injection actions of each control cycle.

[0055] In some specific embodiments, the ammonia injection control device of the SCR denitrification system may further include: An action storage unit is used to store the target current process state, the prior probability distribution, each candidate ammonia injection action sequence, the future process prediction trajectory, the trajectory cost, the posterior weight, the posterior action distribution, and the target ammonia injection action in the posterior distillation sample library for the current control cycle. The parameter update unit is used to determine the distillation loss based on the KL divergence between the posterior action distribution and the prior probability distribution in the posterior distillation sample library, and to update the model parameters of the learned prior model using the distillation loss to obtain the updated learned prior model. The model replacement unit is used to verify the updated learning prior model based on evaluation indicators. If the verification result meets the target safety requirements, the updated learning prior model is used to replace the learning prior model. The evaluation indicators include NOx exceedance ratio, average ammonia injection amount, motion fluctuation degree, average trajectory cost, effective sampling number, and output stability.

[0056] Furthermore, embodiments of this application also disclose an electronic device, Figure 3 This is a structural diagram of an electronic device 20 according to an exemplary embodiment. The content of the diagram should not be construed as limiting the scope of this application. Specifically, the electronic device 20 may include: at least one processor 21, at least one memory 22, a power supply 23, a communication interface 24, an input / output interface 25, and a communication bus 26. The memory 22 stores a computer program, which is loaded and executed by the processor 21 to implement the relevant steps in the ammonia injection control method of the SCR denitrification system disclosed in any of the foregoing embodiments. Furthermore, the electronic device 20 in this embodiment may specifically be an electronic computer.

[0057] In this embodiment, the power supply 23 is used to provide operating voltage for each hardware device on the electronic device 20; the communication interface 24 can create a data transmission channel between the electronic device 20 and external devices, and the communication protocol it follows can be any communication protocol applicable to the technical solution of this application, and is not specifically limited here; the input / output interface 25 is used to acquire external input data or output data to the outside world, and its specific interface type can be selected according to specific application needs, and is not specifically limited here.

[0058] In addition, the memory 22, as a carrier for resource storage, can be a read-only memory, random access memory, disk or optical disk, etc. The resources stored thereon can include operating system 221, computer program 222, etc., and the storage method can be temporary storage or permanent storage.

[0059] The operating system 221 is used to manage and control the various hardware devices on the electronic device 20 and the computer program 222, which may be Windows Server, Netware, Unix, Linux, etc. In addition to including a computer program capable of performing the ammonia injection control method of the SCR denitrification system executed by the electronic device 20 as disclosed in any of the foregoing embodiments, the computer program 222 may further include computer programs capable of performing other specific tasks.

[0060] Furthermore, this application also discloses a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, it implements the aforementioned ammonia injection control method for the SCR denitrification system. Specific steps of this method can be found in the corresponding content disclosed in the foregoing embodiments, and will not be repeated here.

[0061] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section.

[0062] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0063] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented directly by hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.

[0064] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0065] The technical solutions provided in this application have been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for controlling ammonia injection in an SCR denitrification system, characterized in that, include: In the current control cycle, the target current process state is constructed based on the collected current process variables, initial ammonia injection control quantity, and current inlet disturbance variable of the SCR denitrification system during operation, as well as the historical sequences corresponding to the current process variables, the initial ammonia injection control quantity, and the current inlet disturbance variable. The current process variables include outlet converted NOx, outlet measured NOx, outlet oxygen, ammonia water header flow rate, flue gas temperature, flue gas flow rate, and catalyst bed temperature. The prior probability distribution of the ammonia injection action sequence in the future prediction time domain is determined by using the current process state of the target and a learned prior model. Based on the prior probability distribution, each candidate ammonia injection action sequence is determined. The learned prior model is a model determined based on a neural network, a time series model, an attention model, or a diffusion generation model. Based on the candidate ammonia injection action sequences and using an environmental prediction model for rolling prediction, the future process prediction trajectory of each candidate ammonia injection action sequence within the future prediction time domain is obtained; the environmental prediction model is a prediction model constructed based on a neural network model, a state-space model, or a gray box model. Based on the predicted trajectory of the future process, a corresponding NOx control quantity is generated. Based on the NOx control quantity and using a preset trajectory cost function, the trajectory cost value corresponding to each candidate ammonia injection action sequence is determined. Based on the trajectory cost value and using Boltzmann weights, a posterior weight is determined. The candidate ammonia injection action sequences are then weighted using the posterior weights to obtain a posterior action distribution. Based on the candidate ammonia injection action sequences in the posterior action distribution and using a preset projection operator, the target ammonia injection action of the current control cycle is determined, so as to perform ammonia injection control operation on the SCR denitrification system based on the target ammonia injection action.

2. The ammonia injection control method for the SCR denitrification system according to claim 1, characterized in that, In the current control cycle, the target current process state is constructed based on the collected current process variables, initial ammonia injection control quantity, and current inlet disturbance variable of the SCR denitrification system during operation, as well as the historical sequences corresponding to the current process variables, the initial ammonia injection control quantity, and the current inlet disturbance variable, including: In the current control cycle, the current process state is obtained by splicing together the current process variables, initial ammonia injection control quantity, and current inlet disturbance variables of the SCR denitrification system during operation, as well as the historical sequences corresponding to the current process variables, the initial ammonia injection control quantity, and the current inlet disturbance variables; the current inlet disturbance variables include inlet NOx, inlet flue gas flow rate, inlet oxygen, and load change. The current process state is preprocessed to obtain the preprocessed current process state; the preprocessing includes timestamp alignment, outlier removal, missing value handling, data smoothing, unit unification, normalization or standardization, and historical sequence pruning. The data validity of the preprocessed current process state is judged. If the judgment result indicates that the preprocessed current process state is valid, the target current process state is determined based on the preprocessed current process state.

3. The ammonia injection control method for the SCR denitrification system according to claim 1, characterized in that, The step of performing rolling predictions based on the candidate ammonia injection action sequences and using an environmental prediction model to obtain the future process prediction trajectory of each candidate ammonia injection action sequence within the future prediction time domain includes: Based on the candidate ammonia injection action sequence and the current process state of the target, and using an environmental prediction model for rolling prediction, the future process prediction trajectory of each candidate ammonia injection action sequence within the future prediction time domain is obtained; the future process prediction trajectory includes the outlet NOx prediction trajectory, the outlet oxygen prediction trajectory, and the ammonia water flow prediction trajectory.

4. The ammonia injection control method for the SCR denitrification system according to claim 1, characterized in that, The step of generating corresponding NOx control variables based on the predicted trajectory of the future process, and determining the trajectory cost value corresponding to each of the candidate ammonia injection action sequences based on the NOx control variables and using a preset trajectory cost function, includes: Based on the predicted NOx value at the outlet, the predicted NOx value at the outlet, the predicted oxygen value at the outlet, and the measurement confidence factor in the predicted trajectory of the future process, the NOx control quantity is determined using a preset control quantity fusion function. Based on the deviation between the NOx control quantity and the NOx target value, the excess amount of the NOx control quantity relative to the preset safety upper limit, the action amplitude and action change of the candidate ammonia injection action sequence, and the out-of-bounds amount of auxiliary variables in the future process prediction trajectory, the trajectory cost value corresponding to each candidate ammonia injection action sequence is determined using the preset trajectory cost function; the auxiliary variables include outlet oxygen, ammonia water header flow rate, and flue gas temperature.

5. The ammonia injection control method for the SCR denitrification system according to claim 1, characterized in that, The step of determining the target ammonia injection action for the current control cycle based on each candidate ammonia injection action sequence in the posterior action distribution and using a preset projection operator includes: The weighted action is obtained by weighting the first step of the candidate ammonia spraying action sequence in the posterior action distribution and the corresponding posterior weight. The weighted action is projected using the preset projection operator to obtain the target ammonia injection action for the current control cycle.

6. The ammonia injection control method for the SCR denitrification system according to claim 1, characterized in that, Also includes: The process involves determining a new initial ammonia injection control quantity using the target ammonia injection action of the current control cycle, designating the next control cycle as the new current control cycle, and then jumping to the step of constructing the target current process state based on the collected current process variables, initial ammonia injection control quantity, current inlet disturbance variable, and the historical sequences corresponding to the current process variables, initial ammonia injection control quantity, and current inlet disturbance variable during the operation of the SCR denitrification system. Based on the obtained target ammonia injection actions for each control cycle, the SCR denitrification system is subjected to rolling control operation of the ammonia injection process.

7. The ammonia injection control method for the SCR denitrification system according to any one of claims 1 to 6, characterized in that, After determining the target ammonia injection action for the current control cycle based on each candidate ammonia injection action sequence in the posterior action distribution and using a preset projection operator, the method further includes: The target current process state, the prior probability distribution, each candidate ammonia injection action sequence, the future process prediction trajectory, the trajectory cost, the posterior weight, the posterior action distribution, and the target ammonia injection action of the current control cycle are stored in the posterior distillation sample library. The distillation loss is determined based on the KL divergence between the posterior action distribution and the prior probability distribution in the posterior distillation sample library. The model parameters of the learned prior model are updated using the distillation loss to obtain the updated learned prior model. The updated learning prior model is validated based on evaluation metrics. If the validation results meet the target safety requirements, the updated learning prior model is used to replace the learning prior model. The evaluation metrics include NOx exceedance ratio, average ammonia injection amount, motion fluctuation degree, average trajectory cost, effective sampling number, and output stability.

8. An ammonia injection control device for an SCR denitrification system, characterized in that, include: The target state construction module is used to construct the target current process state in the current control cycle based on the collected current process variables, initial ammonia injection control quantity, current inlet disturbance variable, and the historical sequences corresponding to the current process variables, the initial ammonia injection control quantity, and the current inlet disturbance variable during the operation of the SCR denitrification system. The current process variables include outlet converted NOx, outlet measured NOx, outlet oxygen, ammonia water header flow rate, flue gas temperature, flue gas flow rate, and catalyst bed temperature. The candidate sequence determination module is used to determine the prior probability distribution of the ammonia injection action sequence in the future prediction time domain by utilizing the current process state of the target and using a learned prior model, and to determine each candidate ammonia injection action sequence based on the prior probability distribution; the learned prior model is a model determined based on a neural network, a time series model, an attention model, or a diffusion generation model; The prediction trajectory determination module is used to perform rolling prediction based on the candidate ammonia injection action sequences and using an environmental prediction model to obtain the future process prediction trajectory of each candidate ammonia injection action sequence in the future prediction time domain; the environmental prediction model is a prediction model constructed based on a neural network model, a state-space model, or a gray box model. The candidate sequence weighting module is used to generate corresponding NOx control quantities based on the predicted trajectory of the future process, determine the trajectory cost value corresponding to each candidate ammonia injection action sequence based on the NOx control quantities and using a preset trajectory cost function, determine the posterior weight based on the trajectory cost value and using Boltzmann weights, and weight the candidate ammonia injection action sequences using the posterior weights to obtain the posterior action distribution. The target action determination module is used to determine the target ammonia injection action of the current control cycle based on each of the candidate ammonia injection action sequences in the posterior action distribution and using a preset projection operator, so as to perform ammonia injection control operation on the SCR denitrification system based on the target ammonia injection action.

9. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor for executing the computer program to implement the ammonia injection control method for the SCR denitrification system as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, Used to store a computer program, wherein the computer program, when executed by a processor, implements the ammonia injection control method for the SCR denitrification system as described in any one of claims 1 to 7.

Images