A hot blast stove SCR ammonia injection optimization control method, device, equipment and medium

Abstract

The application discloses a hot blast stove SCR ammonia injection optimization control method, device, equipment and medium, relates to the intelligent control technical field of steel smelting flue gas denitration, and comprises the following steps: when the corresponding SCR denitration system of the hot blast stove runs, system key process data is collected to determine the current system state vector; the current system state vector is input into the current world model, the obtained current potential state representation is input into the policy model, the obtained current candidate ammonia injection action and the current world model are used for state deduction to obtain the current ammonia injection action and the system state prediction result of the next moment; after the current ammonia injection action is used for control, the current world model is corrected based on the new current system state vector and the system state prediction result, the action is regenerated based on the new, previous moment current potential state representation, until the termination time, the obtained ammonia injection action sequence, state sequence and multi-objective reward function are used for optimizing the policy model. The application realizes intelligent optimization control of the ammonia injection amount.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology for denitrification of flue gas in iron and steel smelting, and in particular to a method, device, equipment and medium for optimized control of SCR ammonia injection in a hot blast stove. Background Technology

[0002] In actual production, the operating conditions of hot blast stoves exhibit significant dynamic fluctuations, easily causing rapid changes in NOx (nitrogen oxides) concentration, flue gas temperature, and reaction rate at the SCR (Selective Catalytic Reduction) inlet. Furthermore, with increasing demands, the SCR denitrification system corresponding to the hot blast stove places higher requirements on NOx emission control precision, ammonia slip control level, ammonia injection economy, and system stability.

[0003] In existing implementation schemes, PID (Proportional-Integral-Derivative) control, feedforward-feedback control, and some predictive control based on neural networks or reinforcement learning are mainly used to achieve ammonia injection control. However, the following problems still exist in practical applications: 1) PID control relies on fixed parameters for adjustment, which makes it difficult to accurately reflect complex operating conditions, thus easily leading to problems such as adjustment lag, system oscillation, and insufficient control accuracy; 2) Existing control schemes are mostly biased towards single-objective control, and conservative strategies are selected on-site to avoid NOx exceeding the standard, making it difficult to simultaneously take into account NOx and ammonia escape; 3) Existing predictive models are usually trained offline based on historical data, and model drift is prone to occur once on-site operating conditions change, catalyst aging occurs, or environmental boundary conditions change; 4) These schemes mainly make action decisions based on the current state and cannot extrapolate the evolution of operating conditions at multiple future time steps before control, making it easy to have response lag in strong disturbance scenarios such as furnace replacement and load fluctuations. Summary of the Invention

[0004] In view of this, the purpose of this invention is to provide a method, apparatus, equipment, and medium for optimized control of ammonia injection in a hot blast stove SCR system. This method addresses the problems existing in current related solutions, thereby achieving intelligent optimized control of ammonia injection in the SCR denitrification system, reducing world model drift, improving control stability and adaptability under complex operating conditions, and ultimately improving nitrogen oxide removal efficiency and operational economy, while reducing ammonia slip and ammonia consumption. The specific solution is as follows: In a first aspect, this application provides a method for optimized control of ammonia injection in a hot blast stove under SCR conditions, including: During the operation of the SCR denitrification system corresponding to the hot blast stove, the current system state vector is determined by collecting key process data of the SCR denitrification system; wherein, the key process data includes nitrogen oxide concentration data, ammonia slip data, ammonia injection flow rate, flue gas operating parameters, and hot blast stove operating parameters. The current system state vector is input into the current world model to obtain the current potential state representation at the current moment; wherein, the current world model is the model used to simulate the state evolution law of the SCR denitrification system at the current moment; The current potential state representation is input into the strategy model, and the obtained current candidate ammonia injection action and the current world model are used to extrapolate the future state under the current operating conditions, so as to determine the current ammonia injection action and the system state prediction results at the next moment; the strategy model is used to generate actions based on the ammonia injection control strategy. After controlling ammonia injection based on the current ammonia injection action, a new current system state vector is acquired, and the potential dynamic representation within the current world model is corrected based on the new current system state vector and the system state prediction result to determine the corrected current world model and the new current potential state representation; the new current potential state representation is the state representation obtained by re-acquiring key process data from the field and using the corrected current world model. Based on the current potential state representation of the previous moment and the new current potential state representation, the process jumps back to the step of inputting the current potential state representation into the policy model until the preset termination condition is met to obtain the ammonia injection action sequence and the corresponding state sequence. The strategy model is optimized using the ammonia injection action sequence, the state sequence, and a preset multi-objective reward function, and then the optimized strategy model is used to re-participate in ammonia injection control.

[0005] Optionally, determining the current system state vector by collecting key process data of the SCR denitrification system includes: Key process data of the SCR denitrification system are collected to determine the current data collection results; wherein, the data collection results include inlet nitrogen oxide concentration, outlet nitrogen oxide concentration, flue gas temperature, ammonia injection flow rate, flue gas flow rate, combustion air pressure, combustion air flow rate, ammonia slip data, and ammonia injection valve opening; Outlier removal is performed on the data collection results to determine the data after removal; Missing values ​​are compensated for in the removed data to determine the compensated data; The compensated data is synchronized with time to determine the synchronized data; The synchronized data is normalized to determine the processed data; Based on the processed data and the preset state vector definition information, the state vector at the current moment is constructed to determine the current system state vector.

[0006] Optionally, the step of inputting the current system state vector into the current world model to obtain the current potential state representation at the current moment includes: The current potential state representation at the current moment is determined based on the current system state vector, the system state sequence within the historical time window, and the coding network in the current world model. The current potential state represents implicit kinetic information that cannot be directly observed. This implicit kinetic information includes catalyst activity, SCR reaction efficiency, system dynamic inertia, and operating condition change trends.

[0007] Optionally, the step of inputting the current potential state representation into the policy model and using the obtained current candidate ammonia injection action and the current world model to extrapolate the future state under the current operating conditions includes: The current latent state representation is input into the reinforcement learning policy network in the policy model to obtain the current candidate ammonia spraying action; The current candidate ammonia injection action and the state transition network in the current world model are used to virtually simulate the working conditions at multiple future time steps in order to determine the simulation results; The inference results are input into the decoder in the current world model for operating condition prediction and decoding to determine the decoding results; the decoding results include predicted values ​​for nitrogen oxide concentration, ammonia slip, flue gas temperature, and flue gas flow rate. Action selection is performed based on the strategy model, the decoding result, and the current potential state representation to determine the current ammonia injection action and the system state prediction result at the next moment.

[0008] Optionally, after performing ammonia injection control based on the current ammonia injection action, acquiring a new current system state vector, and correcting the latent dynamic representation within the current world model based on the new current system state vector and the system state prediction result, includes: After performing ammonia injection control based on the SCR denitrification system and the ammonia injection control command corresponding to the current ammonia injection action, key process data of the SCR denitrification system are collected to determine the new current system state vector. The prediction error is determined based on the new current system state vector and the system state prediction result. The prediction error, error coding network, and feedback memory coefficient are used to construct a feedback state; the feedback state is used to record the long-term error information of the current world model, including model drift information, calorific value change information, catalyst aging information, and unmodeled perturbation information. The feedback state, the current potential state representation corresponding to the previous moment, and the current ammonia injection action are input into the current world model to correct the potential dynamic representation inside the model, and the corrected current world model is used to determine the new current potential state representation.

[0009] Optionally, the step of jumping back to the step of inputting the current potential state representation into the policy model based on the current potential state representation of the previous time step and the new current potential state representation includes: The current potential state representation from the previous moment and the new current potential state representation are input into the policy model to obtain the new current candidate ammonia injection action; Using the new current candidate ammonia injection action and the revised current world model, the future state under the current operating conditions is re-deduced to determine the new current ammonia injection action and the system state prediction results for the next moment. Based on the new current ammonia injection action, the system jumps back to the step of acquiring the new current system state vector after performing ammonia injection control according to the current ammonia injection action.

[0010] Optionally, optimizing the strategy model using the ammonia injection action sequence, the state sequence, and a preset multi-objective reward function includes: The ammonia injection action sequence, the state sequence, and a preset multi-objective reward function are used to evaluate the action and determine the action evaluation result. The preset multi-objective reward function includes a nitrogen oxide control term, an ammonia escape penalty term, an ammonia injection consumption term, and a control stability term. The reinforcement learning policy network in the policy model is adjusted based on the action evaluation results to complete the model optimization operation.

[0011] Secondly, this application provides a hot blast stove SCR ammonia injection optimization control device, comprising: The system status determination module is used to determine the current system status vector by collecting key process data of the SCR denitrification system during the operation of the SCR denitrification system corresponding to the hot blast stove; wherein, the key process data includes nitrogen oxide concentration data, ammonia slip data, ammonia injection flow rate, flue gas operating parameters, and hot blast stove operating parameters. The potential state determination module is used to input the current system state vector into the current world model to obtain the current potential state representation at the current moment; wherein, the current world model is a model used to simulate the state evolution law of the SCR denitrification system at the current moment; The ammonia injection action determination module is used to input the current potential state representation into the strategy model, and use the obtained current candidate ammonia injection actions and the current world model to extrapolate the future state under the current operating conditions, so as to determine the current ammonia injection action and the system state prediction result at the next moment; the strategy model is used to generate actions based on the ammonia injection control strategy. The model correction module is used to collect a new current system state vector after controlling ammonia injection based on the current ammonia injection action, and to correct the potential dynamic representation inside the current world model based on the new current system state vector and the system state prediction result, so as to determine the corrected current world model and the new current potential state representation; the new current potential state representation is the state representation obtained by re-collecting key process data on site and using the corrected current world model. The sequence acquisition module is used to jump back to the step of inputting the current potential state representation into the policy model based on the current potential state representation of the previous moment and the new current potential state representation, until the preset termination condition is met to obtain the ammonia injection action sequence and the corresponding state sequence. The model optimization module is used to optimize the strategy model using the ammonia injection action sequence, the state sequence, and a preset multi-objective reward function, and then use the optimized strategy model to re-participate in ammonia injection control.

[0012] 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 steps of the aforementioned hot blast stove SCR ammonia injection optimization control method.

[0013] Fourthly, this application provides a computer-readable storage medium for storing a computer program, which, when executed by a processor, implements the steps of the aforementioned hot blast stove SCR ammonia injection optimization control method.

[0014] As can be seen, in this application, during the operation of the SCR denitrification system corresponding to the hot blast stove, the current system state vector is determined by collecting key process data of the SCR denitrification system; wherein, the key process data includes nitrogen oxide concentration data, ammonia slip data, ammonia injection flow rate, flue gas operating parameters, and hot blast stove operating parameters; the current system state vector is input into the current world model to obtain the current potential state representation corresponding to the current moment; wherein, the current world model is a model used to simulate the state evolution law of the SCR denitrification system at the current moment; the current potential state representation is input into the strategy model, and the obtained current candidate ammonia injection action and the current world model are used to extrapolate the future state under the current operating conditions to determine the current ammonia injection action and the system state prediction result at the next moment; the strategy model is used to generate actions based on the ammonia injection control strategy; and according to the current After the ammonia injection action performs ammonia injection control, a new current system state vector is acquired. Based on the new current system state vector and the system state prediction result, the potential dynamic representation within the current world model is corrected to determine the corrected current world model and the new current potential state representation. The new current potential state representation is the state representation obtained by re-acquiring key process data from the field and using the corrected current world model. Based on the current potential state representation at the previous moment and the new current potential state representation, the process jumps back to the step of inputting the current potential state representation into the strategy model until a preset termination condition is met to obtain the ammonia injection action sequence and the corresponding state sequence. The strategy model is optimized using the ammonia injection action sequence, the state sequence, and a preset multi-objective reward function, and the optimized strategy model is used to re-participate in ammonia injection control. In other words, during the operation of the SCR denitrification system corresponding to the hot blast stove, this application first collects key process data of the SCR denitrification system to determine the current system state vector, and inputs the current system state vector into the current world model to obtain the current potential state representation. Then, actions are generated based on the current potential state representation and the strategy model, and future operating conditions are extrapolated using the current candidate ammonia injection actions and the current world model to determine the current ammonia injection action and the system state prediction result at the next moment. Then, after ammonia injection control is performed based on the current ammonia injection action, the potential dynamic state cognition of the current world model is corrected based on the new current system state vector and the system state prediction result to determine the corrected current world model and the new current potential state representation. After that, based on the current potential state representation at the previous moment and the new current potential state representation, the process jumps back to the step of inputting the current potential state representation into the strategy model until termination, when the obtained ammonia injection action sequence and the corresponding state sequence, as well as the preset multi-objective reward function, are used to optimize the strategy model.This approach addresses the problems existing in current solutions, enabling intelligent optimization control of ammonia injection in the SCR denitrification system. It also reduces world model drift, improves control stability and adaptability under complex operating conditions, thereby enhancing nitrogen oxide removal efficiency and operational economy, and reducing ammonia slip and ammonia consumption. Attached Figure Description

[0015] 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.

[0016] Figure 1 A flowchart of an optimized control method for SCR ammonia injection in a hot blast stove is provided in this application; Figure 2 A flowchart of a specific SCR ammonia injection optimization control method for hot blast stoves is provided in this application; Figure 3 A schematic diagram of the structure of an optimized control device for SCR ammonia injection in a hot blast stove provided in this application; Figure 4 This application provides a structural diagram of an electronic device. Detailed Implementation

[0017] 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.

[0018] In existing implementation schemes, PID (Proportional-Integral-Derivative) control, feedforward-feedback control, and some predictive control based on neural networks or reinforcement learning are mainly used to achieve ammonia injection control. However, the following problems still exist in practical applications: 1) PID control relies on fixed parameters for adjustment, which makes it difficult to accurately reflect complex operating conditions, thus easily leading to problems such as adjustment lag, system oscillation, and insufficient control accuracy; 2) Existing control schemes are mostly biased towards single-objective control, and conservative strategies are selected on-site to avoid NOx exceeding the standard, making it difficult to simultaneously take into account NOx and ammonia escape; 3) Existing predictive models are usually trained offline based on historical data, and model drift is prone to occur once on-site operating conditions change, catalyst aging occurs, or environmental boundary conditions change; 4) These schemes mainly make action decisions based on the current state and cannot extrapolate the evolution of operating conditions at multiple future time steps before control, making it easy to have response lag in strong disturbance scenarios such as furnace replacement and load fluctuations.

[0019] To address this, this application provides an optimized control scheme for ammonia injection in a hot blast stove SCR system, which solves the problems existing in the current related schemes. This enables intelligent optimization control of ammonia injection in the SCR denitrification system, reduces the drift of the world model, improves the control stability and adaptability under complex operating conditions, thereby improving the nitrogen oxide removal efficiency and operating economy, and reducing ammonia slip and ammonia consumption.

[0020] See Figure 1 As shown in the figure, an embodiment of the present invention discloses an optimized control method for SCR ammonia injection in a hot blast stove, comprising: Step S11: During the operation of the SCR denitrification system corresponding to the hot blast stove, the current system state vector is determined by collecting key process data of the SCR denitrification system; wherein, the key process data includes nitrogen oxide concentration data, ammonia slip data, ammonia injection flow rate, flue gas operating parameters, and hot blast stove operating parameters.

[0021] Combination Figure 2As shown in this embodiment, during the process of optimizing the ammonia injection control of the SCR denitrification system corresponding to the hot blast stove, it is necessary to monitor the running SCR denitrification system in real time to collect data and construct a state vector. That is, key process data of the SCR denitrification system are collected to determine the current data collection results. The data collection results include inlet nitrogen oxide concentration, outlet nitrogen oxide concentration, flue gas temperature, ammonia injection flow rate, flue gas flow rate, combustion air pressure, combustion air flow rate, ammonia escape data, and ammonia injection valve opening. Outlier values ​​are removed from the data collection results to determine the data after removal. Missing values ​​are compensated for in the removed data to determine the compensated data. The compensated data is synchronized with time to determine the synchronized data. The synchronized data is normalized to determine the processed data. Based on the processed data and the preset state vector definition information, the state vector at the current moment is constructed to determine the current system state vector. It is understood that in this embodiment, after the data is collected, the data is preprocessed before the state vector is generated. The preprocessing includes, but is not limited to, outlier removal, missing value compensation, time synchronization and normalization, in order to reduce the impact of on-site noise on model training or practical application.

[0022] It should be noted that in this embodiment, key process parameters are collected in real time and a system state vector is constructed during the operation of the hot blast stove SCR denitrification system. The collected data includes, but is not limited to: SCR inlet NOx concentration, SCR outlet NOx concentration, flue gas flow rate, flue gas temperature, ammonia slip data, ammonia injection flow rate, ammonia injection valve opening, coal gas flow rate, coal gas calorific value, combustion air pressure, combustion air flow rate, furnace changing status, and induced draft fan operating status. The SCR inlet is generally located after the hot blast stove outlet.

[0023] Meanwhile, the observation vector (i.e., state vector) of the SCR denitrification system at time t. The definition is as follows: ; in, Let be the NOx concentration at the SCR inlet at time t; Let be the NOx concentration at the SCR outlet at time t; Let t be the flue gas temperature at time t; Let be the flue gas flow rate at time t; Let be the amount of ammonia injected at time t; The combustion air pressure at time t; Let t be the combustion air flow rate.

[0024] Step S12: Input the current system state vector into the current world model to obtain the current potential state representation at the current moment; wherein, the current world model is the model used to simulate the state evolution law of the SCR denitrification system at the current moment.

[0025] In this embodiment, after obtaining the state vector at the current moment, a potential state can be generated based on the state vector and the world model. That is, the current potential state representation corresponding to the current moment is determined based on the current system state vector, the system state sequence within the historical time window, and the encoding network in the current world model. The current potential state representation is used to characterize implicit kinetic information that cannot be directly observed. The implicit kinetic information includes catalyst activity, SCR reaction efficiency, system dynamic inertia, and operating condition change trends.

[0026] It is important to understand that, regarding the generation of latent states, this embodiment achieves this through world model encoding, specifically by using an encoder to construct the system's latent states: In the formula, For encoding networks; This refers to the historical memory state corresponding to the previous moment. The potential state of a system is used to characterize implicit kinetic information that cannot be directly observed, such as catalyst activity, SCR reaction efficiency, system dynamic inertia, and trends in operating conditions.

[0027] Understandably, the world model in this embodiment was pre-trained using historical operating data of the SCR denitrification system to learn the state evolution law of the SCR denitrification system corresponding to the hot blast stove. During the training process, the input of the world model includes the current state of the system, the current ammonia injection action, and the state sequence within the historical time window. The output includes the predicted state value at the next moment, the predicted NOx value, the predicted ammonia slip value, the predicted flue gas temperature value, and the potential state representation corresponding to the current operating condition.

[0028] Step S13: Input the current potential state representation into the strategy model, and use the obtained current candidate ammonia injection action and the current world model to extrapolate the future state under the current operating conditions, so as to determine the current ammonia injection action and the system state prediction results at the next moment; the strategy model is used to generate actions based on the ammonia injection control strategy.

[0029] In this embodiment, after obtaining the system's potential state, the potential state is input into the current policy model to generate actions. Then, the world model is used to dynamically extrapolate the SCR operating conditions for multiple future time steps. For different candidate ammonia injection actions, the world model predicts future changes in NOx, ammonia slip, flue gas temperature, and system stability, respectively, and uses these future evolution results as the basis for policy evaluation in the policy model. That is: the current potential state representation is input into the reinforcement learning policy network in the policy model to obtain the current candidate ammonia injection action; the current candidate ammonia injection action and the state transition network in the current world model are used to virtually extrapolate the operating conditions for multiple future time steps to determine the extrapolation results; the extrapolation results are input into the decoder in the current world model for operating condition prediction decoding to determine the decoding results; the decoding results include predicted values ​​for nitrogen oxide concentration, ammonia slip, flue gas temperature, and flue gas flow rate; based on the policy model, the decoding results, and the current potential state representation, action selection is performed to determine the current ammonia injection action and the system state prediction results for the next time step. It is understandable that the reinforcement learning policy network in the policy model does not make decisions solely based on the current state, but rather selects the ammonia injection action based on the simulated future operating conditions, thus giving the control policy a forward-looking nature. The reinforcement learning algorithm in this policy network can employ the PPO algorithm (Proximal Policy Optimization).

[0030] It should be understood that, regarding the process of determining the current ammonia injection action and the system state prediction result at the next moment, this embodiment first generates the ammonia injection action, and uses a reinforcement learning policy network to generate the ammonia injection action based on the potential state: In the formula, This is the current ammonia injection control level. To strengthen the learning policy network, a world model is then used for future projection, in which the world model learns the latent state transition rules: ;in, For world model state transition networks, This provides the operating condition information for the next moment (i.e., the system state prediction result). Further multi-step future projections are then performed: This is used to simulate future operating conditions such as changes in the calorific value of coal gas, fluctuations in flue gas flow rate, and changes in catalyst performance. After performing future projections based on this, operating condition prediction decoding is performed, using the decoder to map the potential states to future operating conditions. ;in, For the decoder in the world model, This provides operating condition information for the next k steps (i.e., system state prediction results), including key process parameters such as future NOx concentration, future ammonia slip, future flue gas temperature, and future flue gas flow rate. Then, a policy network and... Determine the current optimal ammonia injection action, i.e. the current ammonia injection action, and output the optimal ammonia injection control command corresponding to the current ammonia injection action.

[0031] Step S14: After controlling ammonia injection according to the current ammonia injection action, a new current system state vector is collected, and the potential dynamic representation inside the current world model is corrected based on the new current system state vector and the system state prediction result, so as to determine the corrected current world model and the new current potential state representation.

[0032] Combination Figure 2 As shown, in this embodiment, the optimal ammonia injection control command is executed through the SCR denitrification system to perform the current ammonia injection action and complete the ammonia injection control. Then, field feedback data is collected, and the system enters the next round of state update and strategy optimization. Specifically: after ammonia injection control is performed based on the SCR denitrification system and the ammonia injection control command corresponding to the current ammonia injection action, key process data of the SCR denitrification system is collected to determine a new current system state vector; the prediction error is determined based on the new current system state vector and the system state prediction result; a feedback state is constructed using the prediction error, error coding network, and feedback memory coefficient; the feedback state is used to record long-term error information of the current world model, including model drift information, calorific value change information, catalyst aging information, and unmodeled disturbance information; the feedback state, along with the current potential state representation corresponding to the previous moment and the current ammonia injection action, are input into the current world model to correct the potential kinetic representation within the model, and the corrected current world model is used to determine a new current potential state representation.

[0033] It is important to understand that in this embodiment, after the SCR denitrification system performs ammonia injection, it acquires new SCR operating status, actual NOx data, and ammonia slip data in real time, and then re-inputs the actual feedback into the world model for correction. This feedback correction does not directly compensate for the ammonia injection controller output, but rather performs online correction on the potential state and state transition representation of the closed-loop world model. The deviation between the measured state and the state projected by the world model is encoded into a feedback signal using the SCR denitrification system, and this feedback signal is input into the state update unit of the world model to correct the model's internal understanding of SCR kinetics. This allows for subsequent re-projection of future operating conditions using the updated potential state, and generation of the next round of ammonia injection optimization strategy. Therefore, this embodiment achieves closed-loop correction at the model's cognitive layer, rather than error compensation at the controller parameter layer.

[0034] Regarding the aforementioned feedback correction process, specifically, in this embodiment, the closed-loop feedback error is first calculated to obtain the actual field observation value at time t+1. Calculate the prediction error: ,in, The model prediction error is then used. Next, the feedback state is constructed: ,in, Let be the feedback state at time t. This represents the feedback state at time t+1. To provide feedback on the memory coefficient, As an error encoding network, the feedback state is used to record long-term error information such as model drift, catalyst aging, calorific value changes, and unmodeled perturbations. Then, a closed-loop feedback world model correction is performed, re-injecting the feedback state into the world model. or ,in, This is a feedback correction network. The object of correction here is not the controller parameters or the predicted output, but rather the dynamic state representation within the world model. In this way, during the control operation phase, the system compares the actual field detection values ​​with the world model's predicted values, using the deviation information to correct the potential state online, forming a continuously self-correcting closed-loop feedback world model. The corrected world model will then be used to virtually simulate the operating conditions at multiple future time steps, thus providing a more reliable basis for selecting ammonia injection actions based on future states.

[0035] Step S15: Based on the current potential state representation of the previous moment and the new current potential state representation, jump back to the step of inputting the current potential state representation into the policy model until the preset termination condition is met to obtain the ammonia injection action sequence and the corresponding state sequence.

[0036] In this embodiment, after the world model feedback correction is completed, the new current potential state representation obtained from the corrected current world model is used, and together with the current potential state representation from the previous moment, it is input into the policy model to regenerate actions and perform a new round of ammonia injection control. That is: the current potential state representation from the previous moment and the new current potential state representation are input into the policy model to obtain a new current candidate ammonia injection action; the new current candidate ammonia injection action and the corrected current world model are used to re-determine the future state under the current operating conditions to determine the new current ammonia injection action and the system state prediction result for the next moment; based on the new current ammonia injection action, the process jumps back to the step of collecting the new current system state vector after performing ammonia injection control according to the current ammonia injection action. In other words, after the world model is corrected, actions are regenerated and future operating conditions are deduced to complete a new round of ammonia injection control.

[0037] It is understood that the preset termination condition in this embodiment can be a pre-configured time period or a threshold for the number of cycles of ammonia injection control. When the preset termination condition is reached, all executed ammonia injection actions are organized to obtain an ammonia injection action sequence, and the potential state representations corresponding to each action in the sequence are organized to obtain a state sequence.

[0038] Step S16: Optimize the strategy model using the ammonia injection action sequence, the state sequence, and the preset multi-objective reward function, and then re-participate in ammonia injection control using the optimized strategy model.

[0039] In this embodiment, to achieve synergistic optimization of denitrification efficiency and economy, a multi-objective reward function is constructed. The reward items include at least: a NOx control item, an ammonia slip penalty item, an ammonia injection consumption item, and a control stability item. This reward mechanism ensures that the system meets environmental emission requirements while also considering ammonia injection economy and control stability, avoiding increased ammonia slip and equipment operation risks caused by excessive ammonia injection. Specifically, the action evaluation is performed using the ammonia injection action sequence, the state sequence, and the preset multi-objective reward function to determine the action evaluation result. The preset multi-objective reward function includes a nitrogen oxide control item, an ammonia slip penalty item, an ammonia injection consumption item, and a control stability item. Based on the action evaluation result, the reinforcement learning policy network in the strategy model is adjusted to complete the model optimization operation.

[0040] It is important to understand that, in order to achieve synergistic optimization of denitrification efficiency and economics, this scheme constructs a multi-objective reward function. The reward function includes: NOx control index, ammonia slip index (NH3), ammonia injection consumption index (U), and control stability index (the fluctuation of U). Reward Function Defined as: In this formula, the first term is used for NOx emission control, the second term is used for ammonia slip penalty, the third term is used for ammonia injection cost penalty, and the fourth term is used for control stability constraint. Through the above reward mechanism, the system can reduce ammonia injection consumption and control oscillations while meeting environmental emission requirements.

[0041] Meanwhile, the policy model also uses the PPO algorithm to iteratively optimize the reinforcement learning policy network, mainly by updating the policy parameters. The optimization objective is: ; In the formula, As a discount factor, The policy network parameters are continuously updated using the PPO algorithm. To obtain the optimal ammonia injection control strategy. Combined with Figure 2 As shown, it is understandable that the optimized strategy model will continue to be used for ammonia injection control, thus enabling the model to be optimized while control is being implemented.

[0042] In summary, this embodiment provides an optimized ammonia injection control scheme for SCR in a hot blast stove based on a closed-loop feedback world model. By constructing a world model with online feedback correction capabilities, it performs dynamic state prediction and optimized ammonia injection control for the SCR denitrification process in the hot blast stove, addressing the problems of control lag, model drift, and poor control stability inherent in traditional control methods under complex dynamic conditions. Unlike traditional offline prediction models, this scheme introduces an online feedback correction mechanism during the inference phase: the system continuously compares the deviation between the world model's prediction results and the actual measured results, mapping this deviation into a feedback signal to update and dynamically compensate for the potential states within the world model online. This allows the model to continuously correct prediction errors under complex conditions such as catalyst aging, fluctuating gas calorific value, and furnace switching disturbances. The feedback here is not simply a correction of the output value, but a further correction of the world model's internal representation of system dynamics, thereby achieving online closed-loop correction at the model's cognitive level. This improves long-term prediction stability and control adaptability under complex dynamic conditions, enabling the model to maintain higher reliability in subsequent simulations.

[0043] Therefore, in the operation of the SCR denitrification system corresponding to the hot blast stove in this embodiment, the key process data of the SCR denitrification system are first collected to determine the current system state vector, and the current system state vector is input into the current world model to obtain the current potential state representation. Then, actions are generated according to the current potential state representation and the strategy model, and the current candidate ammonia injection action and the current world model are used to extrapolate future operating conditions to determine the current ammonia injection action and the system state prediction result at the next moment. Then, after ammonia injection control is performed according to the current ammonia injection action, the potential dynamic state cognition of the current world model is corrected based on the new current system state vector and the system state prediction result to determine the corrected current world model and the new current potential state representation. After that, based on the current potential state representation at the previous moment and the new current potential state representation, the process jumps back to the step of inputting the current potential state representation into the strategy model until termination. At the end, the obtained ammonia injection action sequence and the corresponding state sequence, as well as the preset multi-objective reward function, are used to optimize the strategy model. This approach addresses the problems existing in current solutions, enabling intelligent optimization control of ammonia injection in the SCR denitrification system. It also reduces world model drift, improves control stability and adaptability under complex operating conditions, thereby enhancing nitrogen oxide removal efficiency and operational economy, and reducing ammonia slip and ammonia consumption.

[0044] See Figure 3 As shown in the embodiments, this application also discloses a hot blast stove SCR ammonia injection optimization control device, including: The system status determination module 11 is used to determine the current system status vector by collecting key process data of the SCR denitrification system during the operation of the SCR denitrification system corresponding to the hot blast stove; wherein, the key process data includes nitrogen oxide concentration data, ammonia slip data, ammonia injection flow rate, flue gas operating parameters and hot blast stove operating parameters. The potential state determination module 12 is used to input the current system state vector into the current world model to obtain the current potential state representation at the current moment; wherein, the current world model is a model used to simulate the state evolution law of the SCR denitrification system at the current moment; The ammonia injection action determination module 13 is used to input the current potential state representation into the strategy model, and use the obtained current candidate ammonia injection action and the current world model to extrapolate the future state under the current working condition, so as to determine the current ammonia injection action and the system state prediction result at the next moment; the strategy model is used to generate actions based on the ammonia injection control strategy. The model correction module 14 is used to collect a new current system state vector after controlling ammonia injection based on the current ammonia injection action, and to correct the potential dynamic representation inside the current world model based on the new current system state vector and the system state prediction result, so as to determine the corrected current world model and the new current potential state representation; the new current potential state representation is the state representation obtained by re-collecting key process data on site and using the corrected current world model. The sequence acquisition module 15 is used to jump back to the step of inputting the current potential state representation into the policy model based on the current potential state representation of the previous moment and the new current potential state representation, until the preset termination condition is met to obtain the ammonia injection action sequence and the corresponding state sequence. The model optimization module 16 is used to optimize the strategy model using the ammonia injection action sequence, the state sequence, and a preset multi-objective reward function, and to re-participate in ammonia injection control using the optimized strategy model.

[0045] In some specific embodiments, the system state determination module 11 may specifically include: The data acquisition unit is used to acquire key process data of the SCR denitrification system to determine the current data acquisition results; wherein, the data acquisition results include inlet nitrogen oxide concentration, outlet nitrogen oxide concentration, flue gas temperature, ammonia injection flow rate, flue gas flow rate, combustion air pressure, combustion air flow rate, ammonia slip data, and ammonia injection valve opening. The elimination unit is used to eliminate outliers from the data acquisition results in order to determine the data after elimination. The compensation unit is used to compensate for missing values ​​in the removed data to determine the compensated data; A synchronization unit is used to synchronize the compensated data in time to determine the synchronized data. A normalization unit is used to normalize the synchronized data to determine the processed data. The vector determination unit is used to construct the current state vector based on the processed data and the preset state vector definition information, so as to determine the current system state vector.

[0046] In some specific embodiments, the potential state determination module 12 may specifically include: The potential state determination unit is used to determine the current potential state representation corresponding to the current moment based on the current system state vector, the system state sequence within the historical time window, and the encoding network in the current world model; wherein, the current potential state representation is used to characterize implicit kinetic information that cannot be directly observed, and the implicit kinetic information includes catalyst activity, SCR reaction efficiency, system dynamic inertia, and operating condition change trend.

[0047] In some specific embodiments, the ammonia injection action determination module 13 may specifically include: The action generation unit is used to input the current potential state representation into the reinforcement learning policy network in the policy model to obtain the current candidate ammonia spraying action; The simulation unit is used to virtually simulate the working conditions at multiple future time steps using the current candidate ammonia injection action and the state transition network in the current world model, in order to determine the simulation results. The decoding unit is used to input the inference results into the decoder in the current world model for operating condition prediction decoding to determine the decoding results; the decoding results include predicted values ​​of nitrogen oxide concentration, ammonia slip, flue gas temperature, and flue gas flow rate; An action selection unit is used to select actions based on the strategy model, the decoding result, and the current potential state representation to determine the current ammonia injection action and the system state prediction result at the next moment.

[0048] In some specific embodiments, the model correction module 14 may specifically include: The new vector determination unit is used to collect key process data of the SCR denitrification system after performing ammonia injection control based on the SCR denitrification system and the ammonia injection control command corresponding to the current ammonia injection action, so as to determine the new current system state vector. A prediction error determination unit is used to determine the prediction error based on the new current system state vector and the system state prediction result. The feedback state determination unit is used to construct a feedback state using the prediction error, the error coding network, and the feedback memory coefficient; the feedback state is used to record the long-term error information of the current world model, including model drift information, calorific value change information, catalyst aging information, and unmodeled perturbation information. The world model correction unit is used to input the feedback state, the current potential state representation corresponding to the previous moment, and the current ammonia injection action into the current world model to correct the potential dynamic representation inside the model, and to use the corrected current world model to determine a new current potential state representation.

[0049] In some specific embodiments, the sequence acquisition module 15 may specifically include: The new action generation unit is used to input the current potential state representation of the previous moment and the new current potential state representation into the policy model to obtain a new current candidate ammonia injection action. The state deduction unit is used to re-determine the future state under the current working condition using the new current candidate ammonia injection action and the corrected current world model, so as to determine the new current ammonia injection action and the system state prediction result for the new next moment. The step jump unit is used to jump back to the step of acquiring the new current system state vector after performing ammonia injection control according to the new current ammonia injection action.

[0050] In some specific embodiments, the model optimization module 16 may specifically include: An evaluation unit is used to evaluate the action using the ammonia injection action sequence, the state sequence, and a preset multi-objective reward function to determine the action evaluation result; the preset multi-objective reward function includes a nitrogen oxide control term, an ammonia escape penalty term, an ammonia injection consumption term, and a control stability term. The model optimization unit is used to adjust the reinforcement learning policy network in the policy model based on the action evaluation results, so as to complete the model optimization operation.

[0051] Furthermore, embodiments of this application also disclose an electronic device, Figure 4 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.

[0052] Figure 4This is a schematic diagram of the structure of an electronic device 20 provided in an embodiment 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 hot blast stove SCR ammonia injection optimization control method disclosed in any of the foregoing embodiments. Alternatively, the electronic device 20 in this embodiment may specifically be an electronic computer.

[0053] 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.

[0054] 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.

[0055] 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 hot blast stove SCR ammonia injection optimization control method executed by the electronic device 20 as disclosed in any of the foregoing embodiments, the computer program 222 may further include a computer program capable of performing other specific tasks.

[0056] 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 hot blast stove SCR ammonia injection optimization control method. Specific steps of this method can be found in the corresponding content disclosed in the foregoing embodiments, and will not be repeated here.

[0057] 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.

[0058] 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.

[0059] 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.

[0060] 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.

[0061] 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 optimizing the control of ammonia injection in a hot blast stove under SCR conditions, characterized in that, include: During the operation of the SCR denitrification system corresponding to the hot blast stove, the current system state vector is determined by collecting key process data of the SCR denitrification system; wherein, the key process data includes nitrogen oxide concentration data, ammonia slip data, ammonia injection flow rate, flue gas operating parameters, and hot blast stove operating parameters. The current system state vector is input into the current world model to obtain the current potential state representation at the current moment; wherein, the current world model is the model used to simulate the state evolution law of the SCR denitrification system at the current moment; The current potential state representation is input into the strategy model, and the obtained current candidate ammonia injection action and the current world model are used to extrapolate the future state under the current operating conditions, so as to determine the current ammonia injection action and the system state prediction results at the next moment; the strategy model is used to generate actions based on the ammonia injection control strategy. After controlling ammonia injection based on the current ammonia injection action, a new current system state vector is acquired, and the potential dynamic representation within the current world model is corrected based on the new current system state vector and the system state prediction result to determine the corrected current world model and the new current potential state representation; the new current potential state representation is the state representation obtained by re-acquiring key process data from the field and using the corrected current world model. Based on the current potential state representation of the previous moment and the new current potential state representation, the process jumps back to the step of inputting the current potential state representation into the policy model until the preset termination condition is met to obtain the ammonia injection action sequence and the corresponding state sequence. The strategy model is optimized using the ammonia injection action sequence, the state sequence, and a preset multi-objective reward function, and then the optimized strategy model is used to re-participate in ammonia injection control.

2. The optimized control method for SCR ammonia injection in a hot blast stove according to claim 1, characterized in that, The process of determining the current system state vector by collecting key process data from the SCR denitrification system includes: Key process data of the SCR denitrification system are collected to determine the current data collection results; wherein, the data collection results include inlet nitrogen oxide concentration, outlet nitrogen oxide concentration, flue gas temperature, ammonia injection flow rate, flue gas flow rate, combustion air pressure, combustion air flow rate, ammonia slip data, and ammonia injection valve opening; Outlier removal is performed on the data collection results to determine the data after removal; Missing values ​​are compensated for in the removed data to determine the compensated data; The compensated data is synchronized in time to determine the synchronized data. The synchronized data is normalized to determine the processed data; Based on the processed data and the preset state vector definition information, the state vector at the current moment is constructed to determine the current system state vector.

3. The optimized control method for SCR ammonia injection in a hot blast stove according to claim 1, characterized in that, The step of inputting the current system state vector into the current world model to obtain the current potential state representation at the current moment includes: The current potential state representation at the current moment is determined based on the current system state vector, the system state sequence within the historical time window, and the coding network in the current world model. The current potential state represents implicit kinetic information that cannot be directly observed. This implicit kinetic information includes catalyst activity, SCR reaction efficiency, system dynamic inertia, and operating condition change trends.

4. The optimized control method for SCR ammonia injection in a hot blast stove according to claim 1, characterized in that, The step of inputting the current potential state representation into the policy model and using the obtained current candidate ammonia injection action and the current world model to extrapolate the future state under the current operating conditions includes: The current latent state representation is input into the reinforcement learning policy network in the policy model to obtain the current candidate ammonia spraying action; The current candidate ammonia injection action and the state transition network in the current world model are used to virtually simulate the working conditions at multiple future time steps in order to determine the simulation results; The inference results are input into the decoder in the current world model for operating condition prediction and decoding to determine the decoding results; the decoding results include predicted values ​​for nitrogen oxide concentration, ammonia slip, flue gas temperature, and flue gas flow rate. Action selection is performed based on the strategy model, the decoding result, and the current potential state representation to determine the current ammonia injection action and the system state prediction result at the next moment.

5. The optimized control method for SCR ammonia injection in a hot blast stove according to claim 1, characterized in that, After performing ammonia injection control based on the current ammonia injection action, a new current system state vector is acquired, and the latent dynamic representation within the current world model is corrected based on the new current system state vector and the system state prediction result, including: After performing ammonia injection control based on the SCR denitrification system and the ammonia injection control command corresponding to the current ammonia injection action, key process data of the SCR denitrification system are collected to determine the new current system state vector. The prediction error is determined based on the new current system state vector and the system state prediction result. The prediction error, error coding network, and feedback memory coefficient are used to construct a feedback state; the feedback state is used to record the long-term error information of the current world model, including model drift information, calorific value change information, catalyst aging information, and unmodeled perturbation information. The feedback state, the current potential state representation corresponding to the previous moment, and the current ammonia injection action are input into the current world model to correct the potential dynamic representation inside the model, and the corrected current world model is used to determine the new current potential state representation.

6. The method for optimized control of SCR ammonia injection in a hot blast stove according to any one of claims 1 to 5, characterized in that, The step of jumping back to the step of inputting the current potential state representation into the policy model based on the current potential state representation of the previous time step and the new current potential state representation includes: The current potential state representation from the previous moment and the new current potential state representation are input into the policy model to obtain the new current candidate ammonia injection action; Using the new current candidate ammonia injection action and the revised current world model, the future state under the current operating conditions is re-deduced to determine the new current ammonia injection action and the system state prediction results for the next moment. Based on the new current ammonia injection action, the system jumps back to the step of acquiring the new current system state vector after performing ammonia injection control according to the current ammonia injection action.

7. The optimized control method for SCR ammonia injection in a hot blast stove according to claim 1, characterized in that, The optimization of the strategy model using the ammonia injection action sequence, the state sequence, and a preset multi-objective reward function includes: The ammonia injection action sequence, the state sequence, and a preset multi-objective reward function are used to evaluate the action and determine the action evaluation result. The preset multi-objective reward function includes a nitrogen oxide control term, an ammonia escape penalty term, an ammonia injection consumption term, and a control stability term. The reinforcement learning policy network in the policy model is adjusted based on the action evaluation results to complete the model optimization operation.

8. A hot blast stove SCR ammonia injection optimization control device, characterized in that, include: The system status determination module is used to determine the current system status vector by collecting key process data of the SCR denitrification system during the operation of the SCR denitrification system corresponding to the hot blast stove; wherein, the key process data includes nitrogen oxide concentration data, ammonia slip data, ammonia injection flow rate, flue gas operating parameters, and hot blast stove operating parameters. The potential state determination module is used to input the current system state vector into the current world model to obtain the current potential state representation at the current moment; wherein, the current world model is a model used to simulate the state evolution law of the SCR denitrification system at the current moment; The ammonia injection action determination module is used to input the current potential state representation into the strategy model, and use the obtained current candidate ammonia injection actions and the current world model to extrapolate the future state under the current operating conditions, so as to determine the current ammonia injection action and the system state prediction result at the next moment; the strategy model is used to generate actions based on the ammonia injection control strategy. The model correction module is used to collect a new current system state vector after controlling ammonia injection based on the current ammonia injection action, and to correct the potential dynamic representation inside the current world model based on the new current system state vector and the system state prediction result, so as to determine the corrected current world model and the new current potential state representation; the new current potential state representation is the state representation obtained by re-collecting key process data on site and using the corrected current world model. The sequence acquisition module is used to jump back to the step of inputting the current potential state representation into the policy model based on the current potential state representation of the previous moment and the new current potential state representation, until the preset termination condition is met to obtain the ammonia injection action sequence and the corresponding state sequence. The model optimization module is used to optimize the strategy model using the ammonia injection action sequence, the state sequence, and a preset multi-objective reward function, and then use the optimized strategy model to re-participate in ammonia injection control.

9. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor for executing the computer program to implement the hot blast stove SCR ammonia injection optimization control method 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, which, when executed by a processor, implements the hot blast stove SCR ammonia injection optimization control method as described in any one of claims 1 to 7.

Images