A denitration control method, device and equipment for industrial flue gas and a storage medium

Abstract

The application discloses a denitration control method, device and equipment of industrial flue gas and a storage medium, relates to the technical field of artificial intelligence, and comprises the following steps: acquiring industrial data flow and constructing an offline training sample, and determining an input state vector based on a state vector. Accordingly, a regional reaction effectiveness evaluation matrix is constructed, and a basic reducing agent action amount is calculated. Then, a basic heating amount is determined according to a temperature window, and a process benchmark action group is formed in combination with a fuel and combustion air amount. An original action vector is determined by using a reinforcement learning strategy network in combination with the input state, the evaluation matrix and the benchmark action group, a candidate action is generated, the to-be-issued action is determined in combination with the candidate action and the state vector, an action set is acquired from a PID controller array, the policy confidence is calculated according to the equipment risk score, and the target issued action is obtained through fusion, so that the efficiency of denitration control of the industrial flue gas is improved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method, apparatus, equipment, and storage medium for denitrification control of industrial flue gas. Background Technology

[0002] Currently, the operational goals of industrial flue gas denitrification sites have gradually shifted towards "long-term stable compliance with standards, controlled NH3 escape, economical operation of the reducing agent, low-load adaptation, extended catalyst life, stable air preheater differential pressure, balanced injection branches, and traceable control processes." Given the challenges of low-load peak shaving, fuel quality fluctuations, furnace condition changes, catalyst degradation, and rising equipment maintenance costs, relying solely on a fixed NH3 / NOx molar ratio, outlet NOx lag feedback PID control, or manual ammonia injection branch balancing is no longer sufficient to meet the collaborative control requirements of denitrification systems under complex operating conditions.

[0003] Specifically, when NOx levels at the inlet suddenly increase or the unit load changes rapidly, traditional control methods often struggle to adjust the total amount and injection distribution of the reducing agent in a timely manner, leading to increased fluctuations in NOx levels at the outlet, sometimes even approaching or exceeding environmental limits. To ensure NOx compliance, a conservative strategy of excessive ammonia injection is often adopted on-site, which increases the risk of NH3 escape, ammonium bisulfate deposition, and air preheater blockage, creating a conflict between compliance control and equipment safety.

[0004] Meanwhile, under low load or low temperature conditions, the denitrification reaction temperature may deviate from the effective temperature window of the SCR catalytic reaction or SNCR injection reaction. If only independent temperature PID or ammonia injection PID is used for adjustment, it is difficult to coordinate the coupling relationship between fuel, combustion air, hot flue gas bypass, and reductant injection, which may lead to problems such as insufficient temperature window, increased CO, insufficient O2, or decreased reductant utilization. Therefore, denitrification control not only needs to determine whether to increase reductant, but also needs to comprehensively assess whether factors such as current temperature, flow rate, residence time, O2, CO, NH3 escape, and equipment pressure difference support increasing injection.

[0005] Furthermore, differences in temperature distribution, flow field distribution, residence time, injection response, and equipment health status often exist between SCR ammonia injection grid branches or SNCR spray gun layers. Traditional manual leveling methods have long adjustment cycles, rely heavily on experience, and suffer from inconsistent control strategies between different shifts, making it difficult to dynamically allocate reducing agent based on the real-time temperature window eligibility and historical response of each branch or spray gun layer. With catalyst activity decay, changes in flow field distribution, and long-term drift in operating load, fixed-ratio control and manual experience rules are also difficult to maintain optimal performance.

[0006] As can be seen from the above, how to improve the efficiency of denitrification control of industrial flue gas is an urgent problem to be solved. Summary of the Invention

[0007] In view of this, the purpose of this invention is to provide a method, apparatus, equipment, and storage medium for denitrification control of industrial flue gas, which can achieve multi-objective collaborative optimization control in the denitrification control process of industrial flue gas, thereby improving the efficiency and safety of denitrification control. The specific solution is as follows: In a first aspect, this application provides a method for denitrification control of industrial flue gas, comprising: Industrial data streams are acquired from the denitrification control system of industrial flue gas, and an offline training sample set is determined based on the industrial data streams. Based on the state vector corresponding to the offline training sample set, an input state vector is determined, including flue gas load and emission variables, temperature and combustion variables, reducing agent and injection variables, reaction and equipment risk variables and auxiliary variables. The regional reaction effectiveness evaluation matrix is ​​determined based on the input state vector, each preset evaluation function, and each preset limiting function. The basic reducing agent action amount is determined based on the current nitrogen oxide load and the reducing agent effective concentration correction term in the input state vector. The regional reaction effectiveness evaluation matrix is ​​used to determine the reducing agent dosage corresponding to the current region. The basic heating amount is determined based on the lower limit of the preset effective reaction temperature window and the current temperature. Based on the input state vector and the basic heating amount, a process benchmark action group including the basic reducing agent action amount, the basic fuel action amount and the basic combustion air action amount is determined. The original action vector is determined by using a reinforcement learning policy network based on the input state vector, the regional reaction effectiveness evaluation matrix and the process benchmark action group. Candidate actions are determined based on the original action vector. Based on the candidate actions and the input state vector, the action to be issued is determined, and the proportional-integral-derivative (PID) action set is obtained from the preset PID controller array. The policy confidence is determined based on the risk score of each device. Based on the PID action set, the policy confidence, and the action to be issued, the target action is determined, so as to perform industrial flue gas denitrification based on the target action.

[0008] Optionally, the step of acquiring industrial data streams from the industrial flue gas denitrification control system and determining an offline training sample set based on the industrial data streams includes: The first subset of collected data, including action values, feedback values, manual / automatic status and equipment interlock status, of each denitrification actuator is obtained from the industrial flue gas denitrification control system. The second subset of collected data, including inlet nitrogen oxide concentration, outlet nitrogen oxide concentration, oxygen concentration, carbon monoxide concentration, flue gas flow rate, humidity, pressure, sulfur dioxide estimate and sulfur trioxide estimate, is obtained from the continuous emission monitoring system. A third subset of data is acquired from the ammonia escape meter, including ammonia escape concentration and corresponding quality label. A fourth subset of data is acquired from the temperature measuring point and the equipment health system, including selective catalytic reduction inlet temperature, catalyst layer temperature, selective non-catalytic reduction spray layer temperature, spray gun layer temperature, branch pressure difference, catalyst pressure difference, air preheater pressure difference, valve position feedback deviation, fan current and equipment alarm information. The fifth subset of data collected from the operation log includes boiler or kiln load, load change rate, fuel type, coal quality or waste calorific value, catalyst operating time, soot blowing status, spray gun or ammonia grid maintenance records, manual intervention records, and historical proportional-integral-derivative outputs. A target data subset is constructed based on the first, second, third, fourth, and fifth data subsets, and a standardized industrial data stream is determined based on the target data subset. Data points that are abnormal in the industrial data stream are removed to obtain the first intermediate data, and each segment in the first intermediate data that exceeds the standard is marked as a safe negative sample to obtain the second intermediate data; The missing data in the second intermediate data is filled by linear interpolation, and the filled data is set with a low-quality flag that meets the preset low quality to obtain the cleaned data. The cleaned data is divided into several working condition segments, and the response time delays between the occurrence time of each action and the start time of the response are determined in each working condition segment. Based on the response time delays, each response variable and the corresponding action variable in the industrial data stream are aligned on the time axis to obtain a quintuple including a state vector, action, reward value, data quality identifier, and the state vector of the next moment after being shifted according to their respective response time delays. An offline training sample set is generated based on each quintuple.

[0009] Optionally, the determination of the input state vector based on the state vector corresponding to the offline training sample set includes variables such as flue gas load and emissions, temperature and combustion, reducing agent and injection, reaction and equipment risk, and auxiliary variables, including: First separation data, including inlet nitrogen oxide concentration, outlet nitrogen oxide concentration and nitrogen oxide emission limit, is separated from the state vector of the offline training sample set, and flue gas load and emission variables are determined based on the first separation data; Separate second separation data from the state vector, including selective catalytic reduction inlet temperature, catalyst layer temperature, selective non-catalytic reduction injection layer temperature, current temperature and target temperature, and determine temperature and combustion variables based on the second separation data; A third set of separation data is extracted from the state vector, including ammonia flow rate, urea solution flow rate, liquid ammonia regulating valve opening feedback value, reducing agent main pipe pressure, ammonia-nitrogen oxide molar ratio, basic reducing agent action amount, ammonia injection grid branch valve opening feedback value, spray gun layer valve start / stop status, branch flow rate, branch pressure difference, and historical injection volume accumulation value. The reducing agent and injection variables are determined based on the third set of separation data. The ammonia-nitrogen oxide molar ratio is the ratio between the number of ammonia moles and the number of nitrogen oxide moles. A fourth set of separation data is extracted from the state vector, including catalyst differential pressure, air preheater differential pressure, branch differential pressure, catalyst operating time, estimated catalyst activity, ammonium bisulfate deposition risk level, spray gun blockage alarm status, valve position feedback deviation, pump / fan current, and equipment interlock status. This fourth set of separation data is used to determine reaction and equipment risk variables. The estimated catalyst activity is a value determined based on the catalyst operating time and the historical decay curve of the inlet nitrogen oxide concentration. The ammonium bisulfate deposition risk level is a level determined based on ammonia slip concentration, catalyst bed temperature, estimated sulfur dioxide, and estimated sulfur trioxide. The input state vector is determined based on the flue gas load and emission variables, the temperature and combustion variables, the reducing agent and injection variables, and the reaction and equipment risk variables.

[0010] Optionally, determining the input state vector based on the flue gas load and emission variables, the temperature and combustion variables, the reducing agent and injection variables, and the reaction and equipment risk variables includes: The state vector is separated into a fifth set of data, which includes the value sequence of each key variable over several past frames as a historical window, frame change rate, each response delay feature, historical proportional-integral-derivative output value, key measurement point quality identifier, state distribution distance, prediction uncertainty, and strategy confidence. Auxiliary variables are determined based on each of the fifth set of data. Wherein, the frame change rate is the difference between the current frame and the previous frame for each key variable; the state distribution distance is the Mahalanobis distance between the current state vector and the state distribution of the training set; and the prediction uncertainty is the variance of the prediction result generated using the environmental prediction model. By concatenating the flue gas load and emission variables, the temperature and combustion variables, the reducing agent and injection variables, the reaction and equipment risk variables, and the auxiliary variables, the input state vector of the reinforcement learning strategy is obtained.

[0011] Optionally, the step of determining the regional reaction effectiveness evaluation matrix based on the input state vector, each preset evaluation function, and each preset limiting function, and determining the basic reducing agent action amount based on the current nitrogen oxide load and reducing agent effective concentration correction term in the input state vector, includes: The temperature term is determined based on the input state vector, the preset temperature weight coefficient, and the preset temperature evaluation function; the nitrogen oxide term is determined based on the preset nitrogen oxide weight coefficient and the preset local nitrogen oxide evaluation function; the flow rate term is determined based on the preset flow rate weight coefficient and the preset flow rate evaluation function; the residence time term is determined based on the preset residence time weight coefficient and the preset residence time evaluation function; and the accumulation result is determined based on the temperature term, the nitrogen oxide term, the flow rate term, and the residence time term. Based on the accumulated results, the preset ammonia escape risk weight coefficient, preset ammonia escape risk, preset differential pressure risk weight coefficient, preset branch differential pressure risk, historical response score weight coefficient, and historical response score, the intermediate quantity for qualification evaluation is determined. The intermediate quantity of the qualification evaluation is nonlinearly transformed to obtain the first transformation result. The qualification value of the temperature window is determined based on the first transformation result and the preset limiting function. The regional response effectiveness evaluation matrix is ​​determined according to the preset region number and the qualification value of each temperature window. The flue gas flow rate and inlet nitrogen oxide concentration are extracted from the input state vector to determine the current nitrogen oxide load based on the flue gas flow rate and the inlet nitrogen oxide concentration, and the target nitrogen oxide removal amount is determined based on the target removal rate and the current nitrogen oxide load. A preset safety correction coefficient is determined based on the ammonia nitrogen oxide molar ratio in the input state vector, and a reducing agent effective concentration correction term is determined based on the ammonia water flow rate, urea solution flow rate and total flow rate in the input state vector; The basic reducing agent action amount is determined based on the preset reducing agent stoichiometric coefficient, the safety correction coefficient, the target nitrogen oxide removal amount, and the reducing agent effective concentration correction term.

[0012] Optionally, the step of determining the basic heating amount based on the preset effective reaction temperature window lower limit and the current temperature, and determining a process baseline action group including the basic reducing agent action amount, the basic fuel action amount, and the basic combustion air action amount based on the input state vector and the basic heating amount, to utilize a reinforcement learning policy network and determine the original action vector based on the input state vector, the regional reaction effectiveness evaluation matrix, and the process baseline action group, and to determine candidate actions based on the original action vector, includes: The current critical temperature is determined based on the selective catalytic reduction inlet temperature, catalyst layer temperature, and selective non-catalytic reduction spray layer temperature in the input state vector. When the current critical temperature is lower than the lower limit of the preset effective reaction temperature window, the basic heating amount is determined based on the flue gas flow rate, the preset flue gas isobaric specific heat capacity, the lower limit of the preset effective reaction temperature window, and the current critical temperature in the input state vector. The basic fuel action quantity is determined based on the basic heating quantity, preset heating efficiency and preset fuel lower heating value, and the basic combustion air action quantity is determined based on the air-fuel ratio, air excess coefficient and the basic fuel action quantity in the input state vector, so as to determine the process reference action group based on the basic reducing agent action quantity, the basic fuel action quantity and the basic combustion air action quantity. Using a reinforcement learning policy network and based on the input state vector, the regional response effectiveness evaluation matrix, and the process baseline action group, the original action vector is determined, and the dimension position corresponding to the denitrification actuator in manual mode or interlock action state is extracted from the manual / automatic state and the equipment interlock state. A permission mask vector is generated based on the dimensional position, and a filtered action vector is determined based on the indicator dimension in the original action vector that corresponds to the permission mask vector. The filtered action vector is constrained by a preset action lower limit vector, a preset action upper limit vector, and a preset change rate upper limit vector to obtain a rate-constrained action vector, and candidate actions are determined based on the rate-constrained action vector.

[0013] Optionally, the step of determining the action to be issued based on the candidate action and the input state vector, obtaining a proportional-integral-derivative (PID) action set from a preset PID controller array, determining the strategy confidence based on the risk score of each device, and determining the target action to be issued based on the PID action set, the strategy confidence, and the action to be issued, so as to perform industrial flue gas denitrification based on the target action, includes: The candidate action is concatenated with the input state vector to obtain a concatenated vector. A preset environmental prediction model is used to determine the predicted values ​​for several future time steps, including the predicted values ​​for nitrogen oxides at the outlet, ammonia escape, key temperature, carbon monoxide, oxygen content, and pressure difference, based on the concatenated vector. Determine whether each predicted value meets the corresponding emission limit constraint. If all of them meet the constraint, determine the action to be issued based on the preset action weight matrix and the candidate actions. The proportional-integral-derivative (PID) actions, including temperature PID action, air-fuel ratio PID action, outlet nitrogen oxides PID action, ammonia-nitrogen molar ratio PID action, ammonia injection flow rate PID action, and branch balance PID action, are obtained from the preset proportional-integral-derivative (PID) controller array. The PID actions of each of these actions are then combined to obtain a set of proportional-integral-derivative (PID) actions. Extract unprocessed data, including state distribution distance, prediction uncertainty, data quality score, and proportional-integral-derivative takeover times, from the input state vector; extract equipment risk score from equipment alarms. Each product is determined based on the preset weight coefficients, the state distribution distance, the prediction uncertainty, the equipment risk score, the data quality score, and the number of proportional-integral-derivative takeover attempts. Based on the preset benchmark value and each of the products, the difference result is determined. The difference result is then subjected to a nonlinear transformation to obtain the policy confidence. Based on the policy confidence and the action to be issued, the weighted reinforcement learning action is determined. Based on the proportional-integral-differential action set and reinforcement learning control weights, a weighted minimum action is determined; based on the weighted reinforcement learning action and the weighted minimum action, a target action is determined; the target action is converted into control commands corresponding to each denitrification actuator, so as to perform industrial flue gas denitrification based on the control commands.

[0014] Secondly, this application provides a denitrification control device for industrial flue gas, comprising: The offline training sample set determination module is used to acquire industrial data streams from the denitrification control system of industrial flue gas and determine the offline training sample set based on the industrial data streams. The input state vector generation module is used to determine the input state vector, including flue gas load and emission variables, temperature and combustion variables, reducing agent and injection variables, reaction and equipment risk variables and auxiliary variables, based on the state vector corresponding to the offline training sample set. The basic reducing agent action amount determination module is used to determine the regional reaction effectiveness evaluation matrix based on the input state vector, each preset evaluation function, and each preset limiting function, and to determine the basic reducing agent action amount based on the current nitrogen oxide load and the reducing agent effective concentration correction term in the input state vector; the regional reaction effectiveness evaluation matrix is ​​used to determine the reducing dose corresponding to the current region; The candidate action generation module is used to determine the basic heating amount based on the lower limit of the preset effective reaction temperature window and the current temperature, and to determine the process benchmark action group including the basic reducing agent action amount, the basic fuel action amount and the basic combustion air action amount based on the input state vector and the basic heating amount. Then, it uses a reinforcement learning policy network and the input state vector, the regional reaction effectiveness evaluation matrix and the process benchmark action group to determine the original action vector, and determines the candidate action based on the original action vector. The action generation module is used to determine the action to be issued based on the candidate actions and the input state vector, obtain the proportional-integral-derivative (PID) action set from the preset PID controller array, determine the strategy confidence based on the risk score of each device, and determine the target action based on the PID action set, the strategy confidence and the action to be issued, so as to perform industrial flue gas denitrification based on the target action.

[0015] 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 industrial flue gas denitrification control method.

[0016] 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 industrial flue gas denitrification control method.

[0017] As can be seen from the above, before performing denitrification control of industrial flue gas, this application needs to obtain industrial data streams from the denitrification control system of industrial flue gas and determine the offline training sample set. Then, based on the state vectors corresponding to the offline training sample set, an input state vector including multiple types of variables is constructed. Subsequently, the regional reaction effectiveness evaluation matrix, the basic reducing agent action quantity and the process baseline action group are determined. Candidate actions are generated using a reinforcement learning policy network. Finally, the target action is determined by combining the policy confidence and the PID minimum action to perform industrial flue gas denitrification.

[0018] Therefore, this application first constructs a high-quality offline training sample set through multi-source data acquisition and processing, providing a reliable data foundation for the training of reinforcement learning strategies. Second, by constructing a comprehensive input state vector including flue gas load and emissions, temperature and combustion, reducing agent and injection, reaction and equipment risks, and auxiliary variables, the strategy network can comprehensively perceive the operating status of the denitrification system. Third, it achieves refined allocation of zoned injections through a regional reaction effectiveness evaluation matrix, and provides a process-level safety benchmark through basic reducing agent action quantities and process baseline action groups. Finally, through a collaborative arbitration mechanism of strategy confidence assessment and PID backup action, it achieves soft switching and safety fallback between reinforcement learning control and traditional control. In this way, the efficiency of industrial flue gas denitrification control is improved, while ensuring the safety and reliability of the control process, thereby enhancing the user experience. Attached Figure Description

[0019] 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.

[0020] Figure 1 This is a flowchart of a denitrification control method for industrial flue gas disclosed in this application; Figure 2This is a schematic diagram of the structure of a denitrification control device for industrial flue gas disclosed in this application; Figure 3 This is a structural diagram of an electronic device disclosed in this application. Detailed Implementation

[0021] 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.

[0022] The denitrification control of industrial flue gas faces complex characteristics such as multivariate coupling, large time delay, nonlinearity, and multi-objective conflict. Traditional PID control and manual empirical rules are difficult to meet modern environmental protection requirements. To address this, this application provides a denitrification control method for industrial flue gas that can achieve multi-objective collaborative optimization control of the denitrification system based on offline reinforcement learning.

[0023] See Figure 1 As shown in the figure, an embodiment of the present invention discloses a method for denitrification control of industrial flue gas, comprising: Step S11: Obtain industrial data stream from the denitrification control system of industrial flue gas, and determine the offline training sample set based on the industrial data stream.

[0024] In this embodiment, during the denitrification control of industrial flue gas, the aim of this application is to construct a high-quality set of real industrial offline training samples to provide a data foundation for the subsequent training of reinforcement learning strategies. Specifically, this application requires collecting industrial data streams from multiple data sources of the denitrification control system. After data cleaning, anomaly removal, and time-delay alignment, an offline training sample set in the form of a five-tuple containing a state vector, action, reward, next-moment state, and data quality identifier is generated.

[0025] In one specific implementation, an industrial data stream is acquired from the industrial flue gas denitrification control system, and an offline training sample set is determined based on the industrial data stream. This specifically includes the following steps: First, multi-source data acquisition is conducted. The denitrification system is divided into several subsystems, including flue gas load subsystem, temperature and combustion heating subsystem, reducing agent supply subsystem, zoned injection subsystem, catalytic or in-furnace reaction subsystem, equipment health subsystem, and emission monitoring subsystem. Data is collected from the corresponding data sources for each subsystem.

[0026] Specifically, a first subset of data, including action values, feedback values, manual / automatic status, and equipment interlock status, is acquired from the denitrification control system of industrial flue gas for each denitrification actuator. This means reading the action values, feedback values, manual / automatic status, and equipment interlock status of fuel valves, combustion air valves, hot flue gas bypass valves, heat exchange regulating valves, ammonia pumps, liquid ammonia regulating valves, urea pumps, ammonia injection grid branch valves, spray gun layer valves, atomizing air valves, dilution water valves, and dilution fans from a DCS (Distributed Control System) or PLC (Programmable Logic Controller).

[0027] A second subset of collected data was obtained from the Continuous Emission Monitoring System (CEMS), including inlet nitrogen oxide concentration, outlet nitrogen oxide concentration, oxygen concentration, carbon monoxide concentration, flue gas flow rate, humidity, pressure, sulfur dioxide estimate, and sulfur trioxide estimate.

[0028] A third subset of data is acquired from the ammonia escape meter, including ammonia escape concentration and corresponding quality label. A fourth subset of data is acquired from the temperature measuring point and the equipment health system, including selective catalytic reduction (SCR) inlet temperature, catalyst bed temperature, selective non-catalytic reduction (SNCR) injection layer temperature, spray gun layer temperature, branch pressure difference, catalyst pressure difference, air preheater pressure difference, valve position feedback deviation, fan current, and equipment alarm information.

[0029] The fifth subset of data collected from the operation log includes boiler or kiln load, load change rate, fuel type, coal quality or waste calorific value, catalyst operating time, soot blowing status, spray gun or ammonia grid maintenance records, manual intervention records, and historical proportional-integral-derivative (PID) output.

[0030] Then, a target data subset is constructed based on the first, second, third, fourth, and fifth data subsets, and a standardized industrial data stream is determined based on the target data subset. Specifically, a unified timestamp, sampling period, unit, range, data quality identifier, availability identifier, and reliability identifier are generated for each variable to form a standardized industrial data stream.

[0031] Next, abnormal data points in the industrial data stream are removed to obtain the first intermediate data. Specifically, bad points are identified in the historical industrial data, and abnormal segments such as those related to furnace shutdown, maintenance, instrument drift, communication interruption, prolonged manual forcing, and unexplained abnormal segments are removed.

[0032] Each segment before exceeding the limit in the first intermediate data is marked as a safe negative sample to obtain the second intermediate data. Specifically, segments before NOx critical exceedance, NH3 escape increase, temperature exceedance, CO increase, O2 deficiency, pressure difference increase, and equipment interlock are retained as safe negative samples.

[0033] Missing data in the second intermediate data is imputed using linear interpolation, and the imputed data is then labeled with a low-quality flag to meet a preset low-quality standard, resulting in cleaned data. Specifically, short-term missing data is imputed using a finite window, and segments that are missing for a long time or whose key measurement points are unavailable are labeled with a low-quality flag and are not used as high-quality samples for strategy training.

[0034] After cleaning, the data is divided into several operating condition segments. Specifically, the operating conditions are segmented according to the type of denitrification (SCR, SNCR, or combined denitrification), load range, inlet NOx level, temperature range, NH3 slip level, catalyst activity, air preheater pressure differential, and injection branch health status.

[0035] Determine the response time delays between the occurrence time and the start time of each action within each operating condition segment. Specifically, calculate the time window from the total amount of reducing agent action to the outlet NOx and NH3 escape response, forming ammonia or urea injection response time delay labels; calculate the time window from the action of fuel valve, combustion air valve, or hot flue gas bypass to the SCR inlet temperature, catalyst bed temperature, O2, and CO response, forming combustion heating response time delay labels; calculate the response time window from the action of SCR ammonia injection grid branch or SNCR spray gun layer to the outlet grid NOx, local NH3 escape, and branch pressure difference change, forming zoned injection response time delay labels.

[0036] Align each response variable and its corresponding action variable in the industrial data stream on the time axis based on the response time delay to obtain a quintuple including state vector, action, reward value, data quality identifier, and the state vector of the next time step after being shifted according to their respective response time delays. Then, generate an offline training sample set based on each quintuple.

[0037] The offline training sample set can be represented as: ； in, This is a real-world industrial offline dataset; ( )for Time-state vector; ( ) refers to historical actions performed; ( ) represents the reward value; ( +1) represents the state at the next moment; ( This is an identifier for data quality or device status. N The total number of samples.

[0038] Through the above-mentioned multi-source data acquisition, cleaning, and time-delay alignment processes, a high-quality offline training sample set reflecting the dynamic characteristics of real industrial processes can be constructed, laying a data foundation for the subsequent training of reinforcement learning strategies.

[0039] Step S12: Determine the input state vector based on the state vector corresponding to the offline training sample set, including flue gas load and emission variables, temperature and combustion variables, reducing agent and injection variables, reaction and equipment risk variables and auxiliary variables.

[0040] In this embodiment, the purpose of this application is to construct the input state vector of the reinforcement learning strategy. To avoid interpreting all data collection points on-site as reinforcement learning state variables, this invention divides the variables into strategy input state variables, temperature window eligibility variables, candidate action variables, safety boundary variables, PID safety net variables, and operation record variables according to their control applications. Specifically, the strategy input state variables serve as the input to the reinforcement learning strategy model; the temperature window eligibility variables determine the selection of zoned spraying or spray gun layers; the candidate action variables represent the actions or action increments that the model intends to issue in the current cycle; the safety boundary variables are used for action barriers, safety projections, and takeover judgments; the PID safety net variables are used for takeover under abnormal or low-confidence conditions; and the operation record variables are used for tracing and offline retraining. These variables can come from the same on-site data stream, but are not required to all enter the reinforcement learning strategy network.

[0041] In one specific implementation, an input state vector is determined based on the state vector corresponding to the offline training sample set, including variables such as flue gas load and emissions, temperature and combustion, reducing agent and injection, reaction and equipment risk, and auxiliary variables. Specifically, this includes: First-stage separation data, including inlet NOx concentration, outlet NOx concentration, and NOx emission limits, is extracted from the state vector of the offline training sample set. Based on this first-stage separation data, flue gas load and emission variables are determined. Specifically, flue gas load and emission variables include inlet NOx concentration, outlet NOx concentration, NOx emission margin, NH3 slip concentration, flue gas flow rate, O2 concentration, and CO concentration; optionally, they also include inlet NOx change rate, outlet NOx change rate, NH3 slip change rate, flue gas flow rate change rate, humidity, pressure, and CEMS data quality identifiers. These variables are used to determine denitrification load, emission risk, ammonia slip risk, and whether reducing reductant is permissible.

[0042] A second set of separation data is extracted from the state vector, including the selective catalytic reduction (SCR) inlet temperature, catalyst bed temperature, selective non-catalytic reduction (SNCR) injection bed temperature, and the current and target temperatures. Temperature and combustion variables are then determined based on this second set of separation data. Specifically, these variables include the SCR inlet temperature, catalyst bed temperature, SNCR injection bed temperature, temperature deviation, temperature change rate, fuel valve feedback, combustion air valve feedback, air-fuel ratio, hot flue gas bypass opening, CO deviation, and O2 deviation. These variables are used to determine whether the temperature is within the effective reaction window and to prevent excessive CO or insufficient O2 caused by attempts to raise the temperature.

[0043] A third set of separation data is extracted from the state vector, including ammonia flow rate, urea solution flow rate, liquid ammonia regulating valve opening feedback value, reducing agent main pipe pressure, ammonia-NOx molar ratio, basic reducing agent action amount, ammonia injection grid branch valve opening feedback value, spray gun layer valve start / stop status, branch flow rate, branch pressure difference, and historical cumulative injection volume. This third set of separation data is used to determine the reducing agent and injection variables. The ammonia-NOx molar ratio is the ratio between the number of moles of ammonia and the number of moles of NOx. This type of variable is used to characterize the reducing agent supply capacity, injection intensity, and zoned injection status.

[0044] A fourth set of separation data is extracted from the state vector, including catalyst differential pressure, air preheater differential pressure, branch differential pressure, catalyst operating time, estimated catalyst activity, ammonium bisulfate deposition risk level, nozzle blockage alarm status, valve position feedback deviation, pump / fan current, and equipment interlock status. This fourth set of separation data is used to determine reaction and equipment risk variables. The estimated catalyst activity is a value determined based on the catalyst operating time and the historical decay curve of inlet nitrogen oxide concentration. The ammonium bisulfate deposition risk level is determined based on ammonia slip concentration, catalyst bed temperature, and estimated sulfur dioxide and sulfur trioxide concentrations. These variables are used to determine whether candidate control actions may cause deposition, blockage, catalyst aging, or equipment interlock risks.

[0045] A fifth set of separation data is extracted from the state vector, including the value sequence of each key variable over several past frames as a historical window, frame change rate, various response time delay characteristics, historical proportional-integral-derivative output values, key measurement point quality indicators, state distribution distance, prediction uncertainty, and policy confidence. Auxiliary variables are determined based on this fifth set of separation data. Specifically, the frame change rate is the difference between the current frame and the previous frame for each key variable; the state distribution distance is the Mahalanobis distance between the current state vector and the state distribution of the training set; and the prediction uncertainty is the variance of the prediction results generated using the environmental prediction model. The historical window covers CEMS detection lag, denitrification reaction lag, combustion thermal inertia, and slow changes in catalyst state; the PID output reflects the control trend of the original control system under the current operating conditions; and the policy confidence determines whether reinforcement learning actions are allowed to enter the final control.

[0046] Finally, by concatenating the flue gas load and emission variables, temperature and combustion variables, reducing agent and injection variables, reaction and equipment risk variables, and auxiliary variables, the input state vector of the reinforcement learning strategy is obtained.

[0047] The input state vector can be represented as: s ( t )=[ xgas ( t ), xT ( t ), xred ( t ), xrisk ( t ), xaux ( t )]; in, ( ) represents the flue gas load and emission variables; ( ) represents temperature and combustion variables; ( ) represents the reducing agent and the injection variable; ( ) represents the risk variables related to reaction and equipment; ( ) are auxiliary variables such as historical window, rate of change, PID output, quality label and confidence level.

[0048] It is worth mentioning that, through the construction of the above five types of core variables, the embodiments of this application not only retain the key coupling information necessary for denitrification control, but also avoid piling up irrelevant points as model inputs, so that the strategy network can comprehensively and efficiently perceive the operating status of the denitrification system.

[0049] Step S13: Determine the regional reaction effectiveness evaluation matrix based on the input state vector, each preset evaluation function and each preset limiting function, and determine the basic reducing agent action amount based on the current nitrogen oxide load and reducing agent effective concentration correction term in the input state vector; the regional reaction effectiveness evaluation matrix is ​​used to determine the reducing dose corresponding to the current region.

[0050] In this embodiment, the present application aims to achieve two core functions: first, to construct a regional reaction effectiveness evaluation matrix (i.e., a temperature window qualification value matrix) to determine whether each SCR ammonia injection branch or SNCR spray gun layer is suitable for adding reducing agent under the current operating conditions; second, to calculate the basic reducing agent action amount as a process benchmark for reinforcement learning strategies.

[0051] In one specific implementation, a regional reaction effectiveness evaluation matrix is ​​determined based on the input state vector, each preset evaluation function, and each preset limiting function. The basic reducing agent action amount is then determined based on the current nitrogen oxide load and the reducing agent effective concentration correction term in the input state vector. Specifically, this includes: First, calculate the temperature window qualification value. The temperature window qualification value is used to describe the... i Whether an SCR ammonia injection branch area or an SNCR spray gun layer is suitable for adding reducing agent under current operating conditions is determined by factors beyond just temperature. These factors include flow rate, residence time, local NOx and O2 levels, historical response, NH3 escape risk, branch pressure differential, and branch health status. A higher qualifying value indicates that the area is more suitable for adding reducing agent; a lower qualifying value indicates that spraying should be limited, reduced, or frozen.

[0052] Specifically, the temperature term is determined based on the input state vector, the preset temperature weight coefficient, and the preset temperature evaluation function; the nitrogen oxide term is determined based on the preset nitrogen oxide weight coefficient and the preset local nitrogen oxide evaluation function; the flow rate term is determined based on the preset flow rate weight coefficient and the preset flow rate evaluation function; the residence time term is determined based on the preset residence time weight coefficient and the preset residence time evaluation function; and the cumulative result is determined based on the temperature term, the nitrogen oxide term, the flow rate term, and the residence time term.

[0053] Based on the cumulative results, preset weight coefficients for ammonia escape risk, preset ammonia escape risk, preset differential pressure risk weight coefficients, preset branch differential pressure risk, weight coefficients for historical response scores, and historical response scores, intermediate quantities for qualification evaluation are determined.

[0054] Among them, the intermediate quantity of qualification evaluation can be expressed as: ; in, For the qualification evaluation intermediate quantity of the i-th ammonia injection branch or spray gun layer; , , and These are evaluation functions for temperature, local NOx, flow rate, and residence time, respectively. Risk of ammonia escape; For differential pressure risk; ( (This refers to a historical response score; each) The weighting coefficients are used. A nonlinear transformation is performed on the intermediate quantities of the qualification evaluation to obtain the first transformation result. Based on the first transformation result and the preset limiting function, the qualification value of the temperature window is determined. The regional response effectiveness evaluation matrix is ​​determined according to the preset region number and the qualification value of each temperature window.

[0055] The temperature window qualification value can be expressed as: Ei ( t )= clip ( σ ( Zi ( t ),0,1); in, Ei ( t ) is the first i Temperature window qualification value for each branch or spray gun layer; σ ( () is the Sigmoid function; clip ( ,0,1) means restricting the eligibility value to between 0 and 1.

[0056] Then, the basic reducing agent action amount is calculated. The flue gas flow rate and inlet nitrogen oxide concentration are extracted from the input state vector to determine the current nitrogen oxide load based on the flue gas flow rate and inlet nitrogen oxide concentration, and the target nitrogen oxide removal amount is determined based on the target removal rate and the current nitrogen oxide load.

[0057] The formula for calculating NOx load is: ; in, for The nitrogen oxide load constantly enters the denitrification system; This refers to the flue gas flow rate; This represents the inlet nitrogen oxide concentration.

[0058] The formula for calculating the target removal amount is: ; in, The target NOx removal amount; The target removal rate can be determined based on emission limits, NOx margin at the outlet, load change rate, and operating strategy.

[0059] The preset safety correction coefficient is determined based on the ammonia-nitrogen oxide molar ratio in the input state vector, and the effective concentration correction term of the reducing agent is determined based on the ammonia water flow rate, urea solution flow rate and total flow rate in the input state vector.

[0060] The basic reducing agent action amount is determined based on the reducing agent stoichiometric coefficient, safety correction coefficient, target nitrogen oxide removal amount, and reducing agent effective concentration correction term.

[0061] The formula for calculating the basic reduction dose is: ; in, Basal reducing dose; This represents the stoichiometric coefficient of the reducing agent reacting with NOx. This is a safety correction factor; This is a correction term for the effective concentration of ammonia water, urea solution, or liquid ammonia. ε To prevent small positive numbers with a denominator of zero.

[0062] In addition, the zoned injection ratio and zoned injection volume can be calculated. The formula for calculating the zoned injection ratio is: ; in, ( ) is the first The spray ratio of each branch or spray gun layer; ( (This refers to the partition preference score output by the reinforcement learning strategy.) This refers to the number of branch paths or spray gun layers.

[0063] The formula for calculating the zoned injection volume is: ; in, ( ) is the first The spray volume of each branch or spray gun layer; , ( The total amount of reducing agent after safety barrier verification is considered.

[0064] By constructing the regional reaction effectiveness evaluation matrix and calculating the basic reducing agent action amount, the present invention can achieve fine allocation at the zoned spray level, while providing a process-level safety benchmark for reinforcement learning strategies.

[0065] Step S14: Determine the basic heating amount based on the preset effective reaction temperature window lower limit and the current temperature, and determine the process benchmark action group including the basic reducing agent action amount, the basic fuel action amount and the basic combustion air action amount based on the input state vector and the basic heating amount. Then, use a reinforcement learning strategy network and the input state vector, the regional reaction effectiveness evaluation matrix and the process benchmark action group to determine the original action vector, and determine the candidate action based on the original action vector.

[0066] In this embodiment, the purpose of this application is to calculate the process baseline action set (including the basic reducing agent action amount, the basic fuel action amount, and the basic combustion air action amount) and generate candidate actions using a reinforcement learning policy network.

[0067] In one specific implementation, a basic heating amount is determined based on a preset effective reaction temperature window lower limit and the current temperature. A process baseline action set, including a basic reducing agent action amount, a basic fuel action amount, and a basic combustion air action amount, is determined based on the input state vector and the basic heating amount. Then, an original action vector is determined using a reinforcement learning policy network based on the input state vector, the regional reaction effectiveness evaluation matrix, and the process baseline action set. Finally, candidate actions are determined based on the original action vector, specifically including: First, the current critical temperature is determined based on the selective catalytic reduction inlet temperature, catalyst layer temperature, and selective non-catalytic reduction spray layer temperature in the input state vector.

[0068] When the current critical temperature is lower than the lower limit of the preset effective reaction temperature window, the basic heating amount is determined based on the flue gas flow rate, the preset flue gas isobaric specific heat capacity, the lower limit of the preset effective reaction temperature window, and the current critical temperature in the input state vector.

[0069] The formula for calculating the basic heating amount is: ; in, Basic heating quantity; The specific heat capacity of flue gas at constant pressure; This represents the lower limit of the effective reaction temperature window. This refers to the critical temperature at the SCR inlet, catalyst layer, or SNCR injection zone.

[0070] The basic fuel action quantity is determined based on the basic heating quantity, the preset heating efficiency, and the preset lower heating value of the fuel.

[0071] The basic fuel quantity calculation formula is: ; in, Basic fuel quantity; For heating efficiency; LHVIt is the lower heating value of fuel.

[0072] The basic combustion air action quantity is determined based on the air-fuel ratio, excess air coefficient, and basic fuel action quantity in the input state vector.

[0073] The basic formula for calculating the amount of combustion air is: ; in, Basic combustion air volume; This refers to the excess air coefficient. This is the theoretical air-fuel ratio.

[0074] The process baseline action group is determined based on the basic reducing agent action quantity, the basic fuel action quantity, and the basic combustion air action quantity.

[0075] Then, the original action vector is determined using a reinforcement learning policy network based on the input state vector, the regional response effectiveness evaluation matrix, and the process baseline action set.

[0076] For SCR systems, the action vector may include the total ammonia injection volume, the increment of the ammonia injection grid branch opening, the increment of the dilution fan frequency, the increment of the hot flue gas bypass opening, and the increments of the fuel valve and combustion air valve openings. For SNCR systems, the action vector may include the increment of the urea pump or ammonia water pump frequency, the action of the spray gun layer selection, the action of the injection ratio of each layer, the increment of the atomizing air valve, and the increment of the dilution water valve. For combined denitrification, the action vector may simultaneously include the front-end SNCR injection action and the back-end SCR fine denitrification action.

[0077] Extract the dimensional position of the denitrification actuator that is in manual mode or interlocking action state from the manual / automatic state and the equipment interlocking state.

[0078] An access control mask vector is generated based on the dimensional position, and the filtered action vector is determined based on the indicator dimension corresponding to the access control mask vector in the original action vector. Specifically, unavailable branches, spray gun layers, valves, pumps, and fans are filtered out by the action access control mask to form candidate actions.

[0079] The filtered action vectors are constrained by the preset action lower limit vector, preset action upper limit vector, and preset change rate upper limit vector to obtain the rate-constrained action vectors.

[0080] The formula for calculating candidate actions is: ; in, Candidate actions; For action permission filtering and boundary projection operators; The reinforcement learning strategy obtained through offline training; For the qualification matrix; For process reference actions; , and These are the lower limit of the action, the upper limit of the action, and the upper limit of the rate of change, respectively.

[0081] Candidate actions are determined based on the action vector after rate constraints.

[0082] By using the above-mentioned process benchmark action group calculation and reinforcement learning policy network to generate candidate actions, this invention combines process knowledge with reinforcement learning decision-making, which not only ensures the process rationality of the actions, but also gives full play to the advantages of reinforcement learning in multi-objective optimization.

[0083] Step S15: Determine the action to be issued based on the candidate action and the input state vector, obtain the proportional-integral-derivative action set from the preset proportional-integral-derivative controller array, determine the strategy confidence based on the risk score of each device, and determine the target action to be issued based on the proportional-integral-derivative action set, the strategy confidence and the action to be issued, so as to perform industrial flue gas denitrification based on the target action.

[0084] In this embodiment, the purpose of this application is to achieve safe verification of actions, policy confidence assessment, and collaborative arbitration with PID backup actions, so as to ultimately determine safe and executable target actions to be issued.

[0085] In one specific embodiment of this application, the step of determining the action to be issued based on the candidate action and the input state vector, obtaining a proportional-integral-derivative (PID) action set from a preset PID controller array, determining the strategy confidence based on the risk score of each device, and determining the target action based on the PID action set, the strategy confidence, and the action to be issued, so as to perform industrial flue gas denitrification based on the target action, specifically includes: First, the candidate action is concatenated with the input state vector to obtain a concatenated vector. Then, a preset environmental prediction model is used to determine the predicted values ​​for several future time steps, including predicted values ​​for nitrogen oxides at the outlet, ammonia escape, key temperature, carbon monoxide, oxygen content, and pressure difference, based on the concatenated vector.

[0086] Determine whether each predicted value meets the corresponding emission limit constraint. If all of them meet the constraint, determine the action to be issued based on the preset action weight matrix and the candidate actions.

[0087] Specifically, the safety conditions include: NOx safety conditions: ; in, The predicted NOx export value under the influence of candidate actions; These are NOx emission limits or warning thresholds.

[0088] NH3 escape safety conditions: ; in, This is a predicted value for future NH3 escape. This is the upper limit or warning threshold for NH3 escape.

[0089] Temperature safety conditions: ; in, For the first Predicted future temperature of key temperature measurement points or injection areas; and To effectively reflect the lower and upper limits of the temperature window.

[0090] Combustion safety conditions: ; ; in, To predict future CO concentrations; The upper limit of CO; This is a predicted value for future oxygen content; This is the lower limit of oxygen content.

[0091] Equipment differential pressure safety conditions: ; in, This refers to the predicted pressure differential of the catalyst, air preheater, or injection branch. This is the upper limit of the allowable pressure difference.

[0092] The safe and feasible domain is defined as: ; in, for t The safe and feasible domain is defined at all times; each subset represents NOx, NH3 escape, temperature, combustion safety, pressure difference, and actuator action constraints, respectively.

[0093] When a candidate action slightly goes out of bounds but can be corrected, perform a safe projection: ; in, The executable actions after secure projection; W Let be the action weight matrix. This formula is used to preserve the optimization direction when a candidate action slightly deviates from the bounds, while avoiding violations of hard boundaries.

[0094] Then, proportional-integral-derivative (PID) actions, including temperature PID action, air-fuel ratio PID action, outlet nitrogen oxides PID action, ammonia-nitrogen molar ratio PID action, ammonia injection flow rate PID action, and branch balance PID action, are obtained from the preset proportional-integral-derivative (PID) controller array. All of the PID actions are then combined to obtain a set of PID actions.

[0095] The PID control module operates synchronously with the reinforcement learning controller, generating directly executable PID actions for each control cycle without stopping due to the reinforcement learning module's operation. Temperature PID controls the SCR inlet temperature, catalyst bed temperature, or SNCR injection bed temperature; air-fuel ratio PID or O2 correction PID controls combustion air during heating and combustion to prevent CO spikes or oxygen deficiency; outlet NOx PID controls the total reductant volume when NOx risk increases; NH3 / NOx molar ratio PID maintains the basic reductant ratio to prevent excessive ammonia injection; ammonia injection flow rate PID, urea flow rate PID, branch balance PID, or spray gun layer balance PID ensure stable operation of the main valve, branch valves, spray gun layer, and ammonia injection grid.

[0096] Next, unprocessed data including state distribution distance, prediction uncertainty, data quality score, and proportional-integral-derivative takeover times are extracted from the input state vector, and equipment risk score is extracted from equipment alarms.

[0097] Each product is determined based on the preset weight coefficients, the state distribution distance, the prediction uncertainty, the equipment risk score, the data quality score, and the number of proportional-integral-derivative takeovers.

[0098] Based on the preset benchmark value and each of the products, the difference results are determined, and the difference results are subjected to a nonlinear transformation to obtain the strategy confidence level.

[0099] The formula for calculating policy confidence is: ; in, For policy confidence; The distance outside the state distribution; To predict uncertainty; Assess equipment risk. Rate the data quality; This represents the number of times the PID has been taken over recently. to These are the weighting coefficients.

[0100] The weighted reinforcement learning action is determined based on the policy confidence and the action to be issued.

[0101] The weighted minimum action is determined based on the proportional-integral-differential action set and the reinforcement learning control weights.

[0102] The final motion calculation formula is: ; in, This is the final action to be issued; This is a PID control action. To reinforce learning control weights, it is used to achieve soft switching from PID to reinforcement learning control or from reinforcement learning to PID.

[0103] The target action is determined based on the weighted reinforcement learning action and the weighted minimum action.

[0104] The target action is converted into control commands for each denitrification actuator, so as to perform industrial flue gas denitrification based on the control commands.

[0105] The following arbitration rules must also be followed during the action issuance process: When the temperature exceeds the window, the action barrier will restrict injection in unsuitable areas, prioritize adjusting fuel and air or switching the spray gun layer, and the PID protection action will be controlled by temperature PID, air-fuel ratio PID or O2 correction PID.

[0106] When there is a risk of NOx exceeding the standard, the action barrier is to prohibit excessive negative ammonia injection, and if necessary, increase the base reduction dose. The PID protection action is to connect the outlet NOx PID, NH3 / NOx feedforward PID, or ammonia injection flow PID.

[0107] When there is a risk of NH3 escape, the action barrier is to restrict the injection of relevant zones, reduce the total reducing agent action, trigger ammonia escape constraint, and the PID protection action is to take over the ammonia escape constraint PID, ammonia injection flow PID, or branch balance PID.

[0108] In case of abnormal combustion, the action barrier will prevent the addition of fuel alone, prioritize the addition of safety air, reduce the heating load or switch the heat source, and the PID protection action will be controlled by the temperature PID and air-fuel ratio PID to manage the fuel and air action.

[0109] When a branch or equipment malfunctions, the action barrier process freezes the abnormal branch action, restricting actions that could lead to a risk of accumulation. The PID backup action is branch balancing PID, DCS interlocking, or manual confirmation logic takeover.

[0110] When the policy confidence is insufficient, the action barrier is to clear the reinforcement learning action permissions and maintain shadow calculation, and the PID backup action is to switch to the verified PID control.

[0111] When transitioning from PID control to reinforcement learning control, a soft-switching approach is used to gradually increase the speed. ( To avoid step changes in the ammonia injection valve, spray gun layer valve, fuel valve, combustion air valve, or hot flue gas bypass.

[0112] Each candidate action, safety projection, final action, switching reason, prediction result, actual feedback, and manual takeover is recorded in the risk ledger to form a traceable record.

[0113] In this way, through the aforementioned safety verification, policy confidence assessment, and PID collaborative arbitration mechanism, the embodiments of this application can fully leverage the optimization capabilities of reinforcement learning while ensuring the safety and reliability of the control process.

[0114] In a first specific embodiment, the denitrification system is first divided into a flue gas load subsystem, a temperature and combustion heating subsystem, a reducing agent supply subsystem, a zoned injection subsystem, a catalytic or in-furnace reaction subsystem, an equipment health subsystem, and an emission monitoring subsystem.

[0115] Secondly, the system reads the action values, feedback values, manual / automatic status, and equipment interlock status of each actuator from the DCS or PLC; it reads process data such as inlet NOx, outlet NOx, O2, CO, flue gas flow rate, humidity, pressure, estimated SO2 / SO3, and particulate matter from the CEMS; it reads the NH3 escape concentration and its quality label from the ammonia escape meter; it reads the SCR inlet temperature, catalyst bed temperature, SNCR injection bed temperature, spray gun bed temperature, branch pressure difference, catalyst pressure difference, air preheater pressure difference, valve position feedback deviation, pump / fan current, and equipment alarms from temperature measuring points and the equipment health system; and it reads the boiler or kiln load, load change rate, fuel type, coal quality or waste calorific value, catalyst operating hours, soot blowing status, spray gun or ammonia injection grid maintenance records, manual intervention records, and historical PID outputs from the operation log.

[0116] Furthermore, a standardized industrial data stream is formed by generating a unified timestamp, sampling period, unit, range, data quality identifier, availability identifier, and credibility identifier for each variable.

[0117] Subsequently, the historical industrial data was used to identify bad points and remove abnormal segments; segments that exceeded the standard were marked as safe negative samples; missing data were interpolated and filled in, and low-quality labels were set; and the operating conditions were segmented according to the type of operating condition.

[0118] Finally, various response delays are calculated, and the states, actions, rewards, next states, and data quality are realigned based on the response delays to form a real industrial offline training sample set.

[0119] In the second specific implementation, the embodiments of this application first need to separate the flue gas load and emission variables, temperature and combustion variables, reducing agent and injection variables, reaction and equipment risk variables, and auxiliary variables from the state vector of the offline training sample set, and then concatenate them to obtain the input state vector of the reinforcement learning strategy.

[0120] Secondly, based on the input state vector and each preset evaluation function, the temperature window qualification value of each branch or spray gun layer is calculated to form a regional response effectiveness evaluation matrix.

[0121] Then, the NOx load is calculated based on the flue gas flow rate and inlet NOx concentration in the input state vector, and the basic reducing agent action amount is calculated by combining the target removal rate and the effective concentration correction term of the reducing agent.

[0122] Furthermore, when the critical temperature is below the lower limit of the effective reaction temperature window, the basic heating amount is calculated, and then the basic fuel action amount and basic combustion air action amount are determined to form the process baseline action group.

[0123] Finally, the input state vector, the regional response effectiveness evaluation matrix, and the process baseline action group are input into the reinforcement learning policy network, and the original action vector is output. After being filtered by permission mask and rate constraint, candidate actions are obtained.

[0124] In the third specific implementation, the embodiments of this application first need to splice candidate actions with input state vectors, and use environmental prediction models to predict key variables such as future NOx, NH3 escape, temperature, CO, O2 and pressure difference.

[0125] Secondly, determine whether each predicted value meets the emission limit constraints and safety boundary conditions. If it does, determine the action to be issued; if it slightly exceeds the limits, perform a safety projection.

[0126] Furthermore, baseline actions such as temperature PID, air-fuel ratio PID, outlet NOx PID, NH3 / NOx molar ratio PID, ammonia injection flow rate PID, and branch balance PID are obtained from the PID controller array.

[0127] Then, the strategy confidence is calculated based on the state distribution distance, prediction uncertainty, equipment risk score, data quality score, and PID takeover frequency.

[0128] Finally, based on the policy confidence and reinforcement learning control weights, the reinforcement learning action and the PID backup action are weighted and fused to determine the final action to be issued and converted into a control command to be issued.

[0129] As can be seen from the above, through the above-described action issuance and PID collaborative process, this application realizes the collaborative arbitration and soft switching of reinforcement learning control and PID control, which ensures the safety of the control process while leveraging the optimization capabilities of reinforcement learning.

[0130] Accordingly, see Figure 2 As shown, this application also provides a denitrification control device for industrial flue gas, comprising: The offline training sample set determination module 11 is used to acquire industrial data streams from the denitrification control system of industrial flue gas and determine the offline training sample set based on the industrial data streams. The input state vector generation module 12 is used to determine the input state vector, including flue gas load and emission variables, temperature and combustion variables, reducing agent and injection variables, reaction and equipment risk variables and auxiliary variables, based on the state vector corresponding to the offline training sample set. The basic reducing agent action amount determination module 13 is used to determine the regional reaction effectiveness evaluation matrix based on the input state vector, each preset evaluation function and each preset limiting function, and to determine the basic reducing agent action amount based on the current nitrogen oxide load and the reducing agent effective concentration correction term in the input state vector; the regional reaction effectiveness evaluation matrix is ​​used to determine the reducing dose corresponding to the current region; The candidate action generation module 14 is used to determine the basic heating amount based on the lower limit of the preset effective reaction temperature window and the current temperature, and to determine the process benchmark action group including the basic reducing agent action amount, the basic fuel action amount and the basic combustion air action amount based on the input state vector and the basic heating amount, so as to use a reinforcement learning policy network and the input state vector, the regional reaction effectiveness evaluation matrix and the process benchmark action group to determine the original action vector, and to determine the candidate action based on the original action vector; The action generation module 15 is used to determine the action to be issued based on the candidate action and the input state vector, obtain the proportional integral differential action set from the preset proportional integral differential controller array, determine the strategy confidence based on the risk score of each device, and determine the target action based on the proportional integral differential action set, the strategy confidence and the action to be issued, so as to perform industrial flue gas denitrification based on the target action.

[0131] In some specific embodiments, the offline training sample set determination module 11 may specifically include: The first data subset acquisition unit is used to acquire a first data subset including action value, feedback value, manual / automatic status and equipment interlock status of each denitrification actuator from the industrial flue gas denitrification control system, and to acquire a second data subset including inlet nitrogen oxide concentration, outlet nitrogen oxide concentration, oxygen concentration, carbon monoxide concentration, flue gas flow rate, humidity, pressure, sulfur dioxide estimate and sulfur trioxide estimate from the continuous emission monitoring system. The second data subset acquisition unit is used to acquire a third data subset including ammonia escape concentration and corresponding quality label from the ammonia escape meter, and to acquire a fourth data subset including selective catalytic reduction inlet temperature, catalyst layer temperature, selective non-catalytic reduction spray layer temperature, spray gun layer temperature, branch pressure difference, catalyst pressure difference, air preheater pressure difference, valve position feedback deviation, fan current and equipment alarm information from temperature measuring points and equipment health system. The third data subset acquisition unit is used to acquire the fifth data subset from the operation log, including boiler or kiln load, load change rate, fuel type, coal quality or waste calorific value, catalyst operating time, soot blowing status, spray gun or ammonia grid maintenance records, manual intervention records, and historical proportional-integral-derivative outputs. An industrial data flow determination unit is used to construct a target data subset based on the first acquired data subset, the second acquired data subset, the third acquired data subset, the fourth acquired data subset, and the fifth acquired data subset, and to determine a standardized industrial data flow based on the target data subset; An intermediate data determination unit is used to remove abnormal data points in the industrial data stream to obtain first intermediate data, and to mark each segment of the first intermediate data before exceeding the standard as a safe negative sample to obtain second intermediate data. The data generation unit after cleaning is used to perform linear interpolation to fill in the missing data in the second intermediate data, and to set a low-quality identifier that meets the preset low-quality requirement for the filled data, so as to obtain the cleaned data. An offline training sample set generation unit is used to divide the cleaned data into several working condition segments, and determine the response time delays between the occurrence time of each action and the start response time in each working condition segment. Based on the response time delays, each response variable and the corresponding action variable in the industrial data stream are aligned on the time axis to obtain a quintuple including a state vector, action, reward value, data quality identifier, and the state vector of the next moment after being shifted according to their respective response time delays. An offline training sample set is generated based on each quintuple.

[0132] In some specific embodiments, the input state vector generation module 12 may specifically include: The flue gas load and emission variable determination unit is used to separate first separation data, including inlet nitrogen oxide concentration, outlet nitrogen oxide concentration and nitrogen oxide emission limit, from the state vector of the offline training sample set, and determine the flue gas load and emission variables based on the first separation data; The temperature and combustion variable determination unit is used to separate second separation data from the state vector, including selective catalytic reduction inlet temperature, catalyst layer temperature, selective non-catalytic reduction injection layer temperature, current temperature and target temperature, and to determine temperature and combustion variables based on the second separation data. The reducing agent and injection variable determination unit is used to separate third separation data from the state vector, including ammonia water flow rate, urea solution flow rate, liquid ammonia regulating valve opening feedback value, reducing agent main pipe pressure, ammonia-nitrogen oxide molar ratio, basic reducing agent action amount, ammonia injection grid branch valve opening feedback value, spray gun layer valve start / stop status, branch flow rate, branch pressure difference, and historical injection quantity cumulative value, so as to determine the reducing agent and injection variables based on the third separation data; the ammonia-nitrogen oxide molar ratio is the ratio between the number of ammonia moles and the number of nitrogen oxide moles; The reaction and equipment risk variable determination unit is used to separate a fourth set of data from the state vector, including catalyst differential pressure, air preheater differential pressure, branch differential pressure, catalyst running time, estimated catalyst activity, ammonium bisulfate deposition risk level, spray gun blockage alarm status, valve position feedback deviation, pump / fan current, and the equipment interlock status, to determine reaction and equipment risk variables based on the fourth set of data; the estimated catalyst activity is a value determined based on the catalyst running time and the historical decay curve of inlet nitrogen oxide concentration; the ammonium bisulfate deposition risk level is a level determined based on ammonia slip concentration, catalyst bed temperature, estimated sulfur dioxide, and estimated sulfur trioxide; The input state vector generation subunit is used to determine the input state vector based on the flue gas load and emission variables, the temperature and combustion variables, the reducing agent and injection variables, and the reaction and equipment risk variables.

[0133] In some specific embodiments, the basic reducing agent action amount determination module 13 may specifically include: The cumulative result calculation unit is used to determine the temperature item based on the input state vector, the preset temperature weight coefficient and the preset temperature evaluation function, the nitrogen oxide item based on the preset nitrogen oxide weight coefficient and the preset local nitrogen oxide evaluation function, the flow rate item based on the preset flow rate weight coefficient and the preset flow rate evaluation function, the residence time item based on the preset residence time weight coefficient and the preset residence time evaluation function, and to determine the cumulative result based on the temperature item, the nitrogen oxide item, the flow rate item and the residence time item; The qualification evaluation intermediate quantity determination unit is used to determine the qualification evaluation intermediate quantity based on the accumulated result, the preset ammonia escape risk weight coefficient, the preset ammonia escape risk, the preset differential pressure risk weight coefficient, the preset branch differential pressure risk, the historical response score weight coefficient, and the historical response score. The regional response effectiveness evaluation matrix determination unit is used to perform nonlinear transformation on the qualification evaluation intermediate quantity to obtain a first transformation result, and determine the temperature window qualification value based on the first transformation result and a preset limiting function, and determine the regional response effectiveness evaluation matrix according to each preset region number and each temperature window qualification value; The target nitrogen oxide removal amount determination unit is used to extract flue gas flow rate and inlet nitrogen oxide concentration from the input state vector, to determine the current nitrogen oxide load based on the flue gas flow rate and the inlet nitrogen oxide concentration, and to determine the target nitrogen oxide removal amount based on the target removal rate and the current nitrogen oxide load; The reducing agent effective concentration correction term generation unit is used to determine a preset safety correction coefficient based on the ammonia nitrogen oxide molar ratio in the input state vector, and to determine the reducing agent effective concentration correction term based on the ammonia water flow rate, urea solution flow rate and total flow rate in the input state vector; The basic reducing agent action amount determination subunit is used to determine the basic reducing agent action amount based on the preset reducing agent stoichiometric coefficient, the safety correction coefficient, the target nitrogen oxide removal amount, and the reducing agent effective concentration correction term.

[0134] In some specific embodiments, the candidate action generation module 14 may specifically include: The current critical temperature determination unit is used to determine the current critical temperature based on the selective catalytic reduction inlet temperature, catalyst layer temperature, and selective non-catalytic reduction spray layer temperature in the input state vector. A baseline heating quantity generation unit is used to determine the baseline heating quantity based on the flue gas flow rate, the preset flue gas isobaric specific heat capacity, the preset effective reaction temperature window lower limit, and the current critical temperature in the input state vector when the current critical temperature is lower than the preset effective reaction temperature window lower limit. The process reference action group construction unit is used to determine the basic fuel action amount based on the basic heating amount, the preset heating efficiency and the preset fuel lower heating value, and to determine the basic combustion air action amount based on the air-fuel ratio, the excess air coefficient and the basic fuel action amount in the input state vector, so as to determine the process reference action group based on the basic reducing agent action amount, the basic fuel action amount and the basic combustion air action amount. The original action vector determination unit is used to determine the original action vector by utilizing a reinforcement learning policy network and based on the input state vector, the regional response effectiveness evaluation matrix, and the process baseline action group, and to extract the dimension position corresponding to the denitrification actuator in manual mode or interlock action state from the manual / automatic state and the equipment interlock state. The filtered action vector generation unit is used to generate a permission mask vector based on the dimensional position, and to determine the filtered action vector based on the indicator dimension in the original action vector that corresponds to the permission mask vector. The candidate action generation subunit is used to constrain the filtered action vector based on a preset action lower limit vector, a preset action upper limit vector, and a preset change rate upper limit vector to obtain a rate-constrained action vector, and to determine candidate actions based on the rate-constrained action vector.

[0135] In some specific embodiments, the action generation module 15 may specifically include: The prediction value generation unit is used to splice the candidate action with the input state vector to obtain the spliced ​​vector, and use a preset environmental prediction model and the spliced ​​vector to determine the predicted values ​​of outlet nitrogen oxides, ammonia escape, key temperature, carbon monoxide, oxygen content and pressure difference for several future time steps. The action to be issued determination unit is used to determine whether each of the predicted values ​​meets the corresponding emission limit constraints. If all of them meet the constraints, the action to be issued is determined based on the preset action weight matrix and the candidate actions. The proportional-integral-derivative (PID) action set generation unit is used to obtain proportional-integral-derivative actions from a preset proportional-integral-derivative controller array, including temperature PID, air-fuel ratio PID, outlet nitrogen oxide PID, ammonia-nitrogen molar ratio PID, ammonia injection flow rate PID, and branch balance PID, and merge all the PID actions to obtain a proportional-integral-derivative action set. The equipment risk score extraction unit is used to extract unprocessed data, including state distribution distance, prediction uncertainty, data quality score and proportional-integral-derivative takeover times, from the input state vector, and to extract equipment risk scores from equipment alarms. The product determination unit is used to determine each product based on each preset weight coefficient, the state distribution distance, the prediction uncertainty, the equipment risk score, the data quality score, and the number of proportional-integral-derivative takeovers. The policy confidence generation unit is used to determine the difference result based on the preset benchmark value and each of the products, perform a nonlinear transformation on the difference result to obtain the policy confidence, and determine the weighted reinforcement learning action based on the policy confidence and the action to be issued. The target action generation unit is used to determine a weighted safety action based on the proportional-integral-differential action set and the reinforcement learning control weights, and to determine the target action based on the weighted reinforcement learning action and the weighted safety action. The control command generation unit is used to convert the target-issued action into control commands corresponding to each denitrification actuator, so as to perform industrial flue gas denitrification based on the control commands.

[0136] 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. The electronic device 20 may specifically 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 industrial flue gas denitrification control method disclosed in any of the foregoing embodiments. Furthermore, the electronic device 20 in this embodiment may specifically be an electronic computer.

[0137] 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.

[0138] 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.

[0139] 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 industrial flue gas denitrification control method 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.

[0140] 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 denitrification control method for industrial flue gas. Specific steps of this method can be found in the corresponding content disclosed in the foregoing embodiments, and will not be repeated here.

[0141] 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.

[0142] 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.

[0143] 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.

[0144] 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.

[0145] 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 denitrification control of industrial flue gas, characterized in that, include: Industrial data streams are acquired from the denitrification control system of industrial flue gas, and an offline training sample set is determined based on the industrial data streams; wherein, the offline training sample set can be represented as: ； in, This is a real-world industrial offline dataset; ( )for Time-state vector; ( ) refers to historical actions performed; ( ) represents the reward value; ( +1) represents the state at the next moment; ( This is an identifier for data quality or device status. The total number of samples; Based on the state vector corresponding to the offline training sample set, an input state vector is determined, including flue gas load and emission variables, temperature and combustion variables, reducing agent and injection variables, reaction and equipment risk variables and auxiliary variables. Based on the input state vector, each preset evaluation function, and each preset limiting function, a regional reaction effectiveness evaluation matrix is ​​determined. The current nitrogen oxide load is determined based on the flue gas flow rate and inlet nitrogen oxide concentration in the input state vector. A reducing agent effective concentration correction term is determined based on the ammonia flow rate, urea solution flow rate, and total flow rate in the input state vector, thereby determining the basic reducing agent action amount. The regional reaction effectiveness evaluation matrix is ​​used to determine the reducing agent dosage corresponding to the current region. Specifically, it determines the temperature term based on the input state vector, preset temperature weighting coefficient, and preset temperature evaluation function, and determines the nitrogen oxide load based on the preset nitrogen oxide concentration... The nitrogen oxide term is determined based on a weighting coefficient and a preset local nitrogen oxide evaluation function; the flow rate term is determined based on a preset flow rate weighting coefficient and a preset flow rate evaluation function; the residence time term is determined based on a preset residence time weighting coefficient and a preset residence time evaluation function; and an accumulated result is determined based on the temperature term, the nitrogen oxide term, the flow rate term, and the residence time term. Based on the accumulated result, a preset ammonia slip risk weighting coefficient, a preset ammonia slip risk, a preset differential pressure risk weighting coefficient, a preset branch differential pressure risk, a historical response score weighting coefficient, and a historical response score, an intermediate qualification evaluation quantity is determined. This intermediate qualification evaluation quantity is then further... A nonlinear transformation is performed to obtain a first transformation result. Based on the first transformation result and a preset limiting function, a temperature window qualification value is determined. A regional response effectiveness evaluation matrix is ​​determined according to each preset region number and each temperature window qualification value. The flue gas flow rate and the inlet nitrogen oxide concentration are extracted from the input state vector to determine the current nitrogen oxide load. The target nitrogen oxide removal amount is determined based on the target removal rate and the current nitrogen oxide load. A preset safety correction coefficient is determined based on the ammonia-nitrogen oxide molar ratio in the input state vector. A reducing agent effective concentration correction term is determined based on the ammonia water flow rate, urea solution flow rate, and total flow rate in the input state vector. A basic reducing agent action amount is determined based on the preset reducing agent stoichiometry coefficient, the safety correction coefficient, the target nitrogen oxide removal amount, and the reducing agent effective concentration correction term. The regional response effectiveness evaluation matrix is ​​used to determine the reduction dose corresponding to the current region, including: determining the partition injection ratio based on each temperature window qualification value in the regional response effectiveness evaluation matrix and the partition preference score output by the reinforcement learning strategy, and using the partition injection ratio to determine the reduction dose corresponding to the current region. The basic heating amount is determined based on the lower limit of the preset effective reaction temperature window and the current temperature. A process baseline action set, including basic reducing agent action amount, basic fuel action amount, and basic combustion air action amount, is determined based on the input state vector and the basic heating amount. An original action vector is then determined using a reinforcement learning policy network based on the input state vector, the regional reaction effectiveness evaluation matrix, and the process baseline action set. Candidate actions are then determined based on the original action vector. Specifically, the process involves using a reinforcement learning policy network to determine the original action vector based on the input state vector, the regional reaction effectiveness evaluation matrix, and the process baseline action set, and then determining candidate actions based on the original action vector. The process includes: utilizing a reinforcement learning policy network and determining an original action vector based on the input state vector, the regional response effectiveness evaluation matrix, and the process baseline action group; extracting the dimension positions corresponding to the denitrification actuators in manual mode or interlocking action state from manual / automatic state and equipment interlocking state; generating an authorization mask vector based on the dimension positions; determining a filtered action vector based on the indication dimension corresponding to the authorization mask vector in the original action vector; constraining the filtered action vector based on a preset action lower limit vector, a preset action upper limit vector, and a preset change rate upper limit vector to obtain a rate-constrained action vector; and determining candidate actions based on the rate-constrained action vector. Based on the candidate actions and the input state vector, the action to be issued is determined, and a proportional-integral-derivative (PID) action set is obtained from a preset PID controller array. The strategy confidence level is determined based on the risk score of each device. The target action to be issued is determined based on the PID action set, the strategy confidence level, and the action to be issued, so as to perform industrial flue gas denitrification based on the target action. Specifically: the candidate actions and the input state vector are concatenated to obtain a concatenated vector. A preset environmental prediction model is used, and based on the concatenated vector, predicted values ​​for several future time steps, including predicted values ​​for outlet nitrogen oxides, ammonia escape, key temperature, carbon monoxide, oxygen content, and pressure difference, are determined. It is determined whether each predicted value meets its corresponding emission limit constraint. If all are met, the action to be issued is determined based on a preset action weight matrix and the candidate actions. The PID action, including temperature PID action, air-fuel ratio PID action, outlet nitrogen oxide PID action, ammonia-nitrogen molar ratio PID action, and ammonia injection PID action, is obtained from the preset PID controller array. The proportional-integral-derivative (PID) actions of flow rate and branch balance are combined to obtain a set of PID actions. Data to be processed, including state distribution distance, prediction uncertainty, data quality score, and PID takeover times, is extracted from the input state vector. Equipment risk scores are extracted from equipment alarms. Products are determined based on preset weight coefficients, the state distribution distance, the prediction uncertainty, the equipment risk score, the data quality score, and the PID takeover times. Differences are determined based on preset benchmark values ​​and the products. A nonlinear transformation is performed on the differences to obtain policy confidence. Weighted reinforcement learning actions are determined based on the policy confidence and the actions to be issued. Weighted safety actions are determined based on the set of PID actions and reinforcement learning control weights. Target actions are determined based on the weighted reinforcement learning actions and the weighted safety actions. Target actions are converted into control commands corresponding to each denitrification actuator for industrial flue gas denitrification.

2. The industrial flue gas denitrification control method according to claim 1, characterized in that, The process of acquiring industrial data streams from the industrial flue gas denitrification control system and determining an offline training sample set based on the industrial data streams includes: The first subset of collected data, including action values, feedback values, manual / automatic status and equipment interlock status, of each denitrification actuator is obtained from the industrial flue gas denitrification control system. The second subset of collected data, including inlet nitrogen oxide concentration, outlet nitrogen oxide concentration, oxygen concentration, carbon monoxide concentration, flue gas flow rate, humidity, pressure, sulfur dioxide estimate and sulfur trioxide estimate, is obtained from the continuous emission monitoring system. A third subset of data is acquired from the ammonia escape meter, including ammonia escape concentration and corresponding quality label. A fourth subset of data is acquired from the temperature measuring point and the equipment health system, including selective catalytic reduction inlet temperature, catalyst layer temperature, selective non-catalytic reduction spray layer temperature, spray gun layer temperature, branch pressure difference, catalyst pressure difference, air preheater pressure difference, valve position feedback deviation, fan current and equipment alarm information. The fifth subset of data collected from the operation log includes boiler or kiln load, load change rate, fuel type, coal quality or waste calorific value, catalyst operating time, soot blowing status, spray gun or ammonia grid maintenance records, manual intervention records, and historical proportional-integral-derivative outputs. A target data subset is constructed based on the first, second, third, fourth, and fifth data subsets, and a standardized industrial data stream is determined based on the target data subset. Data points that are abnormal in the industrial data stream are removed to obtain the first intermediate data, and each segment in the first intermediate data that exceeds the standard is marked as a safe negative sample to obtain the second intermediate data; The missing data in the second intermediate data is filled by linear interpolation, and the filled data is set to meet the preset low quality label to obtain the cleaned data. The cleaned data is divided into several working condition segments, and the response time delays between the occurrence time of each action and the start time of the response are determined in each working condition segment. Based on the response time delays, each response variable and the corresponding action variable in the industrial data stream are aligned on the time axis to obtain a quintuple including a state vector, action, reward value, data quality identifier, and the state vector of the next moment after being shifted according to their respective response time delays. An offline training sample set is generated based on each quintuple.

3. The industrial flue gas denitrification control method according to claim 2, characterized in that, The determination of the input state vector based on the state vector corresponding to the offline training sample set includes: flue gas load and emission variables, temperature and combustion variables, reducing agent and injection variables, reaction and equipment risk variables, and auxiliary variables. First separation data, including inlet nitrogen oxide concentration, outlet nitrogen oxide concentration and nitrogen oxide emission limit, is separated from the state vector of the offline training sample set, and flue gas load and emission variables are determined based on the first separation data; Separate second separation data from the state vector, including selective catalytic reduction inlet temperature, catalyst layer temperature, selective non-catalytic reduction injection layer temperature, current temperature and target temperature, and determine temperature and combustion variables based on the second separation data; A third set of separation data is extracted from the state vector, including ammonia flow rate, urea solution flow rate, liquid ammonia regulating valve opening feedback value, reducing agent main pipe pressure, ammonia-nitrogen oxide molar ratio, basic reducing agent action amount, ammonia injection grid branch valve opening feedback value, spray gun layer valve start / stop status, branch flow rate, branch pressure difference, and historical injection volume accumulation value. The reducing agent and injection variables are determined based on the third set of separation data. The ammonia-nitrogen oxide molar ratio is the ratio between the number of ammonia moles and the number of nitrogen oxide moles. A fourth set of separation data is extracted from the state vector, including catalyst differential pressure, air preheater differential pressure, branch differential pressure, catalyst operating time, estimated catalyst activity, ammonium bisulfate deposition risk level, spray gun blockage alarm status, valve position feedback deviation, pump / fan current, and equipment interlock status. This fourth set of separation data is used to determine reaction and equipment risk variables. The estimated catalyst activity is a value determined based on the catalyst operating time and the historical decay curve of the inlet nitrogen oxide concentration. The ammonium bisulfate deposition risk level is a level determined based on ammonia slip concentration, catalyst bed temperature, estimated sulfur dioxide, and estimated sulfur trioxide. The input state vector is determined based on the flue gas load and emission variables, the temperature and combustion variables, the reducing agent and injection variables, and the reaction and equipment risk variables.

4. The industrial flue gas denitrification control method according to claim 3, characterized in that, The process of determining the input state vector based on the flue gas load and emission variables, the temperature and combustion variables, the reducing agent and injection variables, and the reaction and equipment risk variables includes: The state vector is separated into a fifth set of data, which includes the value sequence of each key variable over several past frames as a historical window, frame change rate, each response time delay feature, historical proportional-integral-derivative output value, key measurement point quality identifier, state distribution distance, prediction uncertainty, and strategy confidence. Auxiliary variables are determined based on each of the fifth set of data. Wherein, the frame change rate is the difference between the current frame and the previous frame for each key variable; the state distribution distance is the Mahalanobis distance between the current state vector and the state distribution of the training set; and the prediction uncertainty is the variance of the prediction result generated using the environmental prediction model. By concatenating the flue gas load and emission variables, the temperature and combustion variables, the reducing agent and injection variables, the reaction and equipment risk variables, and the auxiliary variables, the input state vector of the reinforcement learning strategy is obtained.

5. The industrial flue gas denitrification control method according to claim 3, characterized in that, The process baseline action group, which determines the basic heating amount based on the lower limit of the preset effective reaction temperature window and the current temperature, and determines the basic reducing agent action amount, basic fuel action amount, and basic combustion air action amount based on the input state vector and the basic heating amount, includes: The current critical temperature is determined based on the selective catalytic reduction inlet temperature, catalyst layer temperature, and selective non-catalytic reduction spray layer temperature in the input state vector. When the current critical temperature is lower than the lower limit of the preset effective reaction temperature window, the basic heating amount is determined based on the flue gas flow rate, the preset flue gas isobaric specific heat capacity, the lower limit of the preset effective reaction temperature window, and the current critical temperature in the input state vector. The basic fuel action quantity is determined based on the basic heating quantity, preset heating efficiency and preset fuel lower heating value, and the basic combustion air action quantity is determined based on the air-fuel ratio and excess air coefficient in the input state vector and the basic fuel action quantity. The process reference action group is determined based on the basic reducing agent action quantity, the basic fuel action quantity and the basic combustion air action quantity.

6. A denitrification control device for industrial flue gas, characterized in that, include: The offline training sample set determination module is used to acquire industrial data streams from the industrial flue gas denitrification control system and determine the offline training sample set based on the industrial data streams; wherein, the offline training sample set can be represented as: ； in, This is a real-world industrial offline dataset; ( )for Time-state vector; ( ) refers to historical actions performed; ( ) represents the reward value; ( +1) represents the state at the next moment; ( This is an identifier for data quality or device status. The total number of samples; The input state vector generation module is used to determine the input state vector, including flue gas load and emission variables, temperature and combustion variables, reducing agent and injection variables, reaction and equipment risk variables and auxiliary variables, based on the state vector corresponding to the offline training sample set. The basic reducing agent action quantity determination module is used to determine the regional reaction effectiveness evaluation matrix based on the input state vector, each preset evaluation function, and each preset limiting function; to determine the current nitrogen oxide load based on the flue gas flow rate and inlet nitrogen oxide concentration in the input state vector; and to determine the reducing agent effective concentration correction term based on the ammonia water flow rate, urea solution flow rate, and total flow rate in the input state vector, thereby determining the basic reducing agent action quantity. The regional reaction effectiveness evaluation matrix is ​​used to determine the reducing agent dosage corresponding to the current region; specifically, it is used to determine the reducing agent dosage based on the input state vector, preset temperature weighting coefficient, and preset temperature evaluation function. A temperature term is defined, and a nitrogen oxide term is determined based on a preset nitrogen oxide weighting coefficient and a preset local nitrogen oxide evaluation function. A flow rate term is determined based on a preset flow rate weighting coefficient and a preset flow rate evaluation function. A residence time term is determined based on a preset residence time weighting coefficient and a preset residence time evaluation function. An accumulated result is determined based on the temperature term, the nitrogen oxide term, the flow rate term, and the residence time term. An intermediate qualification evaluation quantity is determined based on the accumulated result, a preset ammonia slip risk weighting coefficient, a preset ammonia slip risk, a preset differential pressure risk weighting coefficient, a preset branch differential pressure risk, a historical response score weighting coefficient, and a historical response score. The qualification evaluation intermediate quantity is nonlinearly transformed to obtain a first transformation result. Based on the first transformation result and a preset limiting function, a temperature window qualification value is determined. A regional response effectiveness evaluation matrix is ​​determined according to each preset region number and each temperature window qualification value. The flue gas flow rate and the inlet nitrogen oxide concentration are extracted from the input state vector to determine the current nitrogen oxide load. The target nitrogen oxide removal amount is determined based on the target removal rate and the current nitrogen oxide load. A preset safety correction coefficient is determined based on the ammonia-nitrogen oxide molar ratio in the input state vector. A reducing agent effective concentration correction term is determined based on the ammonia water flow rate, urea solution flow rate, and total flow rate in the input state vector. A basic reducing agent action amount is determined based on the preset reducing agent stoichiometry coefficient, the safety correction coefficient, the target nitrogen oxide removal amount, and the reducing agent effective concentration correction term. The regional response effectiveness evaluation matrix is ​​used to determine the reduction dose corresponding to the current region, including: determining the partition injection ratio based on each temperature window qualification value in the regional response effectiveness evaluation matrix and the partition preference score output by the reinforcement learning strategy, and using the partition injection ratio to determine the reduction dose corresponding to the current region. The candidate action generation module is used to determine the basic heating amount based on the lower limit of the preset effective reaction temperature window and the current temperature, and to determine a process baseline action group including the basic reducing agent action amount, the basic fuel action amount, and the basic combustion air action amount based on the input state vector and the basic heating amount. Then, it uses a reinforcement learning policy network and the input state vector, the regional reaction effectiveness evaluation matrix, and the process baseline action group to determine the original action vector, and determines candidate actions based on the original action vector. Specifically, the process involves using a reinforcement learning policy network and the input state vector, the regional reaction effectiveness evaluation matrix, and the process baseline action group to determine the original action vector, and then using the original action vector to determine candidate actions. The process for determining candidate actions includes: using a reinforcement learning policy network and based on the input state vector, the regional response effectiveness evaluation matrix, and the process baseline action group to determine an original action vector; extracting the dimension positions corresponding to the denitrification actuators in manual or interlocked action states from manual / automatic and equipment interlock states; generating an access control mask vector based on the dimension positions; and determining a filtered action vector based on the indicator dimension corresponding to the access control mask vector in the original action vector; constraining the filtered action vector based on a preset action lower limit vector, a preset action upper limit vector, and a preset rate of change upper limit vector to obtain a rate-constrained action vector; and determining candidate actions based on the rate-constrained action vector. The action generation module is used to determine the action to be issued based on the candidate actions and the input state vector, and to obtain a proportional-integral-derivative (PID) action set from a preset proportional-integral-derivative (PID) controller array. It determines the strategy confidence level based on the risk score of each device, and determines the target action based on the PID action set, the strategy confidence level, and the action to be issued, so as to perform industrial flue gas denitrification based on the target action. Specifically, it concatenates the candidate actions and the input state vector to obtain a concatenated vector, uses a preset environmental prediction model, and determines the predicted values ​​for several future time steps, including predicted values ​​for outlet nitrogen oxides, ammonia slip, key temperature, carbon monoxide, oxygen content, and pressure difference, based on the concatenated vector. It determines whether each predicted value meets the corresponding emission limit constraints. If all are met, it determines the action to be issued based on a preset action weight matrix and the candidate actions. It obtains temperature PID actions, air-fuel ratio PID actions, outlet nitrogen oxide PID actions, and ammonia-nitrogen molar ratio PID actions from the preset PID controller array. The proportional-integral-derivative (PID) actions of the sub-actions, ammonia injection flow rate PID actions, and branch balance PID actions are merged to obtain a set of PLD actions. Data to be processed, including state distribution distance, prediction uncertainty, data quality score, and PLD intervention times, is extracted from the input state vector. Equipment risk scores are extracted from equipment alarms. Products are determined based on preset weight coefficients, the state distribution distance, the prediction uncertainty, the equipment risk score, the data quality score, and the PLD intervention times. Differences are determined based on preset benchmark values ​​and the products. A nonlinear transformation is performed on the differences to obtain policy confidence. Weighted reinforcement learning actions are determined based on the policy confidence and the actions to be issued. Weighted safety actions are determined based on the set of PLD actions and reinforcement learning control weights. Target actions are determined based on the weighted reinforcement learning actions and the weighted safety actions. Target actions are converted into control commands corresponding to each denitrification actuator to perform industrial flue gas denitrification based on the control commands.

7. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor for executing the computer program to implement the denitrification control method for industrial flue gas as described in any one of claims 1 to 5.

8. 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 denitrification control method for industrial flue gas as described in any one of claims 1 to 5.

Images