A denitration control method and device based on ammonia injection response time lag modeling, equipment and storage medium

Abstract

The application discloses a denitration control method and device based on ammonia injection response time lag modeling, equipment and storage medium, relates to the flue gas denitration control technical field of coal-fired power plants, and comprises the following steps: time correlation processing is carried out on historical operation data, and the outlet nitrogen oxide response sequence is obtained; the change relationship between ammonia injection action and response sequence is determined, the first response time from action to ammonia water flow change and the second response time from ammonia water flow to outlet nitrogen oxide change are obtained, and the ammonia injection response time lag relationship is established; based on the time lag relationship, the ammonia injection influence characteristics are extracted, the denitration state prediction model is constructed by combining process monitoring and operation data; the model is used to train the ammonia injection optimization intelligent agent, candidate actions are generated according to online data, and control actions are output after verification, so that the denitration control precision is improved.

Description

Technical Field

[0001] This invention relates to the field of flue gas denitrification control technology in coal-fired power plants, and in particular to a denitrification control method, device, equipment and storage medium based on ammonia injection response time delay modeling. Background Technology

[0002] Currently, nitrogen oxides (NOx) in flue gas from coal-fired power plants are one of the main sources of air pollution. Selective catalytic reduction (SCR) technology is currently the mainstream technology for controlling NOx emissions from coal-fired power plant flue gas. SCR denitrification systems reduce NOx to nitrogen and water by injecting ammonia water or liquid ammonia into the flue gas duct, under the action of a catalyst. However, the control of SCR denitrification systems faces many challenges.

[0003] First, SCR denitrification systems exhibit significant time delays. There is a response lag between the adjustment of the ammonia injection valve position and the actual change in ammonia flow rate, and a time delay between the change in ammonia flow rate and the actual change in outlet nitrogen oxide concentration, due to chemical reactions and flue gas transport. These time delays can range from several minutes to over ten minutes. Under such conditions, traditional proportional-integral-derivative (PID) control or model predictive control struggles to achieve ideal control results, easily leading to either excessive or insufficient ammonia injection. The former results in ammonia escape and increased operating costs, while the latter increases the risk of exceeding emission standards.

[0004] Secondly, SCR denitrification systems involve complex coupling relationships between multiple variables. Upstream disturbance variables such as inlet nitrogen oxide concentration, flue gas flow rate, flue gas temperature, and production load have a significant impact on outlet nitrogen oxide concentration. Control variables such as ammonia injection valve position, ammonia flow rate setpoint, and pump frequency affect outlet nitrogen oxide concentration through chemical reaction processes. The causal relationships between these variables have clear process directions and time delays. Traditional data-driven modeling methods, which rely solely on statistical correlation for prediction, are prone to learning spurious correlations that contradict process common sense. For example, they might mistakenly attribute a decrease in outlet nitrogen oxide concentration to a reduction in current ammonia injection, rather than a lagged response to a previous increase in ammonia injection.

[0005] Third, intelligent optimization methods such as reinforcement learning have received widespread attention in the field of industrial control in recent years. However, in industrial scenarios with high safety requirements, such as SCR denitrification, it is unacceptable to directly deploy reinforcement learning agents into actual systems for online trial-and-error learning, because unreasonable actions may lead to excessive emissions or equipment safety issues. Therefore, how to train a safe and reliable ammonia injection optimization strategy in an offline environment is an urgent technical problem that needs to be solved in this field.

[0006] Fourth, SCR denitrification systems frequently experience abnormal operation periods during actual operation, such as shutdowns, maintenance, CEMS (Continuous Emission Monitoring System) calibration, instrument purging, and sensor malfunctions. The operational data from these abnormal periods cannot be used to train normal ammonia injection control strategies; otherwise, it will contaminate the training samples and cause the model to learn incorrect control patterns. Traditional methods for identifying and removing abnormal periods rely on manual experience, which is inefficient and lacks standardized criteria.

[0007] As can be seen from the above, how to improve the efficiency of controlling the SCR denitrification system is an urgent problem to be solved. Summary of the Invention

[0008] In view of this, the purpose of this invention is to provide a denitrification control method, apparatus, equipment, and storage medium based on ammonia injection response time delay modeling, which can prevent excessive emissions and ammonia injection accidents during the denitrification control process based on ammonia injection response time delay modeling, thereby improving the economy, safety, and stability of denitrification control. The specific solution is as follows: In a first aspect, this application provides a denitrification control method based on ammonia injection response time delay modeling, including: Based on the historical ammonia injection control range, the historical denitrification system operation data is subjected to time correlation processing to obtain time correlation data, and the outlet nitrogen oxide response sequence is determined based on the time correlation data; the historical denitrification system operation data includes process monitoring data, ammonia injection control data, and operation status data; The relationship between the historical ammonia injection sequence and the outlet nitrogen oxide response sequence is determined, and the first response time from the execution of the historical ammonia injection to the change in ammonia flow rate during the denitrification reaction is determined. Then, the second response time from the change in ammonia flow rate to the change in outlet nitrogen oxides is determined, so as to determine the ammonia injection response time lag relationship based on the relationship, the first response time and the second response time. Based on the time delay relationship of the ammonia injection action response, the ammonia injection control data of the historical ammonia injection control range is subjected to time delay correlation processing to obtain the ammonia injection influence characteristics. Based on the ammonia injection influence characteristics, the process monitoring data and the operating status data, a denitrification status prediction model is constructed. The denitrification state prediction model is used to construct a target ammonia injection optimization agent based on the historical ammonia injection action sequence. Based on the online operation data of the denitrification system, candidate values ​​for ammonia injection control actions are generated. The target ammonia injection control action is then generated based on the candidate values ​​using the denitrification state prediction model and the target ammonia injection optimization agent. The denitrification control is then performed using the target ammonia injection control action.

[0009] Optionally, the step of performing time-correlation processing on historical denitrification system operating data based on historical ammonia injection control range to obtain time-correlation data includes: Historical operating data of the denitrification system are collected from the distributed control system (DCS), the flue gas online monitoring system (CEMS), and the production operation system corresponding to the denitrification system. Based on the historical denitrification system operation data, process monitoring data, ammonia injection control data, and operating status data are acquired. The process monitoring data includes inlet nitrogen oxide concentration, outlet nitrogen oxide concentration, outlet oxygen content, and reactor pressure difference. The ammonia injection control data includes ammonia injection valve position, ammonia water flow rate setpoint, and pump frequency. The operating status data includes flue gas flow rate, flue gas temperature, and production load. Identify the abnormal data segments that occur in the process monitoring data, the ammonia injection control data, and the operating status data, and then remove each of the abnormal data segments from the process monitoring data, the ammonia injection control data, and the operating status data to obtain continuous operating data; the abnormal data segments are data segments that have occurred during shutdown, maintenance, CEMS calibration abnormality, instrument purging abnormality, and sensor abnormality. The continuous operation data is processed by time correlation based on the historical ammonia injection control range to obtain time-correlated data.

[0010] Optionally, determining the outlet nitrogen oxide response sequence based on the time-correlation data includes: The time-related data are arranged based on the historical ammonia injection control range to obtain a historical ammonia injection action sequence; The time range corresponding to the historical ammonia injection action sequence is determined, and an outlet nitrogen oxide response sequence is generated based on the outlet nitrogen oxide concentration data within the time range.

[0011] Optionally, determining the relationship between the historical ammonia injection sequence and the outlet nitrogen oxide response sequence, and determining the first response time from the execution of the historical ammonia injection to the change in ammonia flow rate during the denitrification reaction, and then determining the second response time from the change in ammonia flow rate to the change in outlet nitrogen oxides, to determine the ammonia injection response time lag relationship based on the relationship, the first response time, and the second response time, includes: The action change process is established based on the historical ammonia injection action sequence, and the pollutant response process is established based on the outlet nitrogen oxide response sequence. Then, the relationship between the action change process and the pollutant response process is determined. Based on the changes in ammonia injection valve position, ammonia flow rate, and pump frequency in the historical ammonia injection action sequence, the execution response process corresponding to the ammonia injection control action is determined. The process of change in outlet nitrogen oxide concentration after the change in ammonia injection control action is determined based on the outlet nitrogen oxide response sequence. Based on the execution response process and the outlet nitrogen oxide concentration change process, determine the first response time from the ammonia injection control action to the change in ammonia flow rate and the second response time from the change in ammonia flow rate to the change in outlet nitrogen oxide concentration. Based on the aforementioned change relationship, the first response time, and the second response time, a time delay relationship for ammonia injection action response is generated.

[0012] Optionally, the step of performing time-delay correlation processing on the ammonia injection control data of the historical ammonia injection control range based on the ammonia injection action response time-delay relationship to obtain ammonia injection influence characteristics includes: Determine the ammonia injection control data within the historical ammonia injection control range, and extract the ammonia injection valve position change, ammonia water flow rate change, and pump frequency change from the ammonia injection control data; Based on the time lag relationship of the ammonia injection action response, and based on the changes in the ammonia injection valve position, the changes in the ammonia flow rate, and the changes in the pump frequency, the ammonia injection influence characteristics are constructed.

[0013] Optionally, the step of constructing a denitrification status prediction model based on the ammonia injection impact characteristics, the process monitoring data, and the operating status data includes: Based on the ammonia injection impact characteristics, the process monitoring data, and the operating status data, an initial denitrification state prediction model is trained to obtain a target denitrification state prediction model. The target denitrification state prediction model is then used to predict the first state prediction results corresponding to the future control cycle, including the outlet nitrogen oxide concentration, ammonia water flow rate, and equipment operating status. Based on the first state prediction result, the prediction error, action response direction consistency and multi-cycle prediction stability of the target denitrification state prediction model are verified, and the verified target denitrification state prediction model is encapsulated as a virtual denitrification operation environment.

[0014] Optionally, the step of constructing a target ammonia injection optimization agent using the denitrification state prediction model and based on the historical ammonia injection action sequence, and generating candidate values ​​for ammonia injection control actions based on the online operation data of the denitrification system, so as to generate a target ammonia injection control action based on the candidate values ​​of ammonia injection control action using the denitrification state prediction model and the target ammonia injection optimization agent, and to use the target ammonia injection control action for denitrification control, includes: In the virtual denitrification operation environment, based on each historical ammonia injection control action in the historical ammonia injection action sequence, data to be processed is determined, including the results of changes in outlet nitrogen oxide concentration, ammonia water consumption, and constraint judgment results. Then, based on the data to be processed, the initial ammonia injection optimization agent is trained to obtain the target ammonia injection optimization agent. Acquire the online operating data of the denitrification system, and generate candidate values ​​for the initial ammonia injection control action corresponding to the current control cycle based on the online operating data; Determine the action evaluation results corresponding to each of the initial ammonia injection control action candidate values, and predict the future state based on the initial ammonia injection control action candidate values ​​in the virtual denitrification operation environment to obtain the second state prediction result. The target ammonia injection optimization agent is used, and the initial ammonia injection control action candidate value is adjusted based on the second state prediction result and the action evaluation result to obtain the target ammonia injection control action candidate value; the target ammonia injection control action candidate value satisfies the preset outlet nitrogen oxide emission constraint condition and the preset ammonia water consumption condition. The candidate values ​​of the target ammonia injection control action are subjected to safety verification, and the candidate values ​​of the target ammonia injection control action that pass the safety verification are set as the target ammonia injection control action, so as to use the target ammonia injection control action for denitrification control.

[0015] Secondly, this application provides a denitrification control device based on ammonia injection response time delay modeling, comprising: The time-correlation data generation module is used to perform time-correlation processing on historical denitrification system operation data based on historical ammonia injection control range to obtain time-correlation data, and to determine the outlet nitrogen oxide response sequence based on the time-correlation data; the historical denitrification system operation data includes process monitoring data, ammonia injection control data, and operation status data; The ammonia injection action response time delay relationship determination module is used to determine the change relationship between the historical ammonia injection action sequence and the outlet nitrogen oxide response sequence, and to determine the first response time from the execution of the historical ammonia injection action to the change of ammonia water flow rate during the denitrification reaction, and then to determine the second response time from the change of ammonia water flow rate to the change of outlet nitrogen oxides, so as to determine the ammonia injection action response time delay relationship based on the change relationship, the first response time and the second response time; The denitrification status prediction model construction module is used to perform time-delay correlation processing on the ammonia injection control data of the historical ammonia injection control range based on the time-delay relationship of the ammonia injection action response, to obtain the ammonia injection influence characteristics, and to construct a denitrification status prediction model based on the ammonia injection influence characteristics, the process monitoring data and the operating status data. The ammonia injection control action generation module is used to construct a target ammonia injection optimization agent based on the denitrification state prediction model and the historical ammonia injection action sequence, and to generate candidate values ​​for ammonia injection control actions based on the online operation data of the denitrification system. The module then uses the denitrification state prediction model and the target ammonia injection optimization agent to generate a target ammonia injection control action based on the candidate values, and uses the target ammonia injection control action to perform denitrification control.

[0016] 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 denitrification control method based on ammonia injection response time delay modeling.

[0017] 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 denitrification control method based on ammonia injection response time delay modeling.

[0018] As can be seen from the above, before performing denitrification control based on ammonia injection response time delay modeling, this application needs to perform time correlation processing on historical denitrification system operation data based on historical ammonia injection control range to obtain time correlation data, and determine the outlet nitrogen oxide response sequence based on the time correlation data. Then, the change relationship between the historical ammonia injection action sequence and the outlet nitrogen oxide response sequence is determined, and the first response time from the execution of historical ammonia injection action to the change of ammonia water flow rate and the second response time from the change of ammonia water flow rate to the change of outlet nitrogen oxide are determined during the denitrification reaction. Based on the above change relationship and the above response time, the ammonia injection action response time delay relationship is determined, and the ammonia injection influence characteristics and denitrification state prediction model are constructed based on the above time delay relationship. Finally, the above denitrification state prediction model and the target ammonia injection optimization agent are constructed based on the historical ammonia injection action sequence, and the target ammonia injection optimization agent is used to generate ammonia injection control actions to achieve denitrification control.

[0019] Therefore, this application first needs to extract the relationship between ammonia injection and the response of nitrogen oxides at the outlet from historical operating data. This allows for the determination of the first response time from the execution of the action to the change in ammonia flow rate and the second response time from the change in ammonia flow rate to the change in nitrogen oxides at the outlet. Then, based on the relationship and response time, the time delay relationship of the ammonia injection action is determined. Finally, based on this time delay relationship, an ammonia injection impact characteristic reflecting the delayed effect of the ammonia injection action on the outlet nitrogen oxides is constructed, thereby building an accurate denitrification state prediction model. This improves the accuracy and safety of ammonia injection control in the denitrification control process based on ammonia injection response time delay modeling. It significantly reduces ammonia consumption and ammonia escape while ensuring that nitrogen oxide emissions do not exceed standards. Simultaneously, the safety filtration module ensures system operational safety, thereby enhancing the user experience. Attached Figure Description

[0020] 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.

[0021] Figure 1 This is a flowchart of a denitrification control method based on ammonia injection response time delay modeling disclosed in this application; Figure 2 This is a schematic diagram of the hierarchical structure of a specific denitrification control system based on ammonia injection response time delay modeling disclosed in this application; Figure 3 This is a schematic diagram of a specific denitrification control process based on ammonia injection response time delay modeling disclosed in this application; Figure 4 This is a schematic diagram of a denitrification control device based on ammonia injection response time delay modeling disclosed in this application; Figure 5 This is a structural diagram of an electronic device disclosed in this application. Detailed Implementation

[0022] 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.

[0023] SCR denitrification systems in coal-fired power plants are characterized by large time delays, multivariate coupling, and strong nonlinearity. Improving the accuracy and economy of ammonia injection control is a pressing issue in this field. To address this, this application provides a denitrification control method based on ammonia injection response time delay modeling. This method fully considers the causal time delay relationship between ammonia injection and outlet nitrogen oxides, constructs a high-precision denitrification state prediction model, and trains an ammonia injection optimization agent offline in a frozen virtual environment. A safety filtering module ensures the safety of online deployment, thereby minimizing ammonia consumption while guaranteeing emission standards.

[0024] See Figure 1 As shown in the figure, an embodiment of the present invention discloses a denitrification control method based on ammonia injection response time delay modeling, comprising: Step S11: Perform time correlation processing on the historical denitrification system operation data based on the historical ammonia injection control range to obtain time correlation data, and determine the outlet nitrogen oxide response sequence based on the time correlation data; the historical denitrification system operation data includes process monitoring data, ammonia injection control data, and operation status data.

[0025] In this embodiment, the historical denitrification system operation data comes from the corresponding DCS (Distributed Control System), CEMS (Continuous Emission Monitoring System), and production operation system. These systems collect and record various parameters during the denitrification system's operation at different sampling periods.

[0026] It is worth mentioning that the DCS system is mainly responsible for collecting process control parameters such as ammonia injection valve position, ammonia flow rate setpoint, pump frequency, flue gas temperature, and reactor differential pressure; the CEMS system is mainly responsible for collecting flue gas emission parameters such as inlet nitrogen oxide concentration, outlet nitrogen oxide concentration, and outlet oxygen content; and the production operation system is mainly responsible for collecting operation management parameters such as production load and start / stop status.

[0027] Furthermore, the sampling periods of the three systems may differ. For example, the sampling period of the CEMS system is usually 60 to 300 seconds, the sampling period of the DCS system is usually 1 to 10 seconds, and the data update period of the production operation system may be even longer.

[0028] To achieve the fusion of multi-source data, this embodiment uses the control action update cycle as the main time axis, aggregating high-frequency variables into the main cycle and aligning low-frequency variables by forward padding with the most recent valid value no later than the current time. Specifically, this embodiment strictly prohibits using future detection values ​​to fill historical states to avoid future information leakage, which could lead to unreasonable prediction accuracy during model training, but which would be impossible to reproduce in actual deployment.

[0029] In one specific implementation, this application requires time-correlation processing of historical denitrification system operation data based on historical ammonia injection control range to obtain time-correlation data, specifically including the following steps: First, historical operational data of the denitrification system was collected from the corresponding DCS system, CEMS system, and production operation system. This data collection process covered at least one full operating year to fully encompass operating scenarios under different load conditions, ambient temperatures, and coal qualities. The collected data was correlated using timestamps as the primary key across multiple sources.

[0030] Secondly, based on the aforementioned historical denitrification system operation data, process monitoring data, ammonia injection control data, and operational status data were acquired. The process monitoring data included inlet nitrogen oxide concentration (mg / Nm³), outlet nitrogen oxide concentration (mg / Nm³), outlet oxygen content (volume percentage), and reactor differential pressure (Pa). The ammonia injection control data included ammonia injection valve position (percentage opening), ammonia flow rate setpoint (L / h), and pump frequency (Hz). The operational status data included flue gas flow rate (Nm³ / h), flue gas temperature (°C), and production load (MW or percentage).

[0031] Next, identify the abnormal data segments in the process monitoring data, ammonia injection control data, and operating status data, and then remove each abnormal data segment from the process monitoring data, ammonia injection control data, and operating status data to obtain continuous operating data.

[0032] The abnormal data segments are those resulting from shutdowns, maintenance, CEMS calibration anomalies, instrument purging anomalies, and obvious sensor malfunctions. Taking CEMS calibration as an example, the CEMS analyzer in the denitrification system requires periodic zero-point and range calibration using standard gases. During calibration, the nitrogen oxide and oxygen content values ​​output by the CEMS are standard gas concentration values, not the actual emission concentrations within the flue gas duct. Therefore, the data from calibration segments cannot be used to train control models under normal operating conditions.

[0033] Similarly, during the instrument purging phase, compressed air is introduced into the sampling pipeline for purging, and the CEMS reading does not represent the actual emission concentration. Significant sensor anomalies include values ​​exceeding the reasonable measurement range, values ​​remaining unchanged for extended periods (stuck), and value fluctuations exceeding the equipment's permissible limits. Identification of these abnormal data segments can be achieved through status flags in the equipment's operating log or through automatic detection using data analysis algorithms.

[0034] Finally, this embodiment of the application requires time correlation processing of the aforementioned continuous operation data based on the historical ammonia injection control range to obtain time-correlated data. Specifically, for each sampling time t, the system forms a field observation vector O. t (Including process monitoring data and operating status data), ammonia injection action vector A t (Including ammonia injection valve position, ammonia flow rate setpoint, and pump frequency) and uncontrollable disturbance vector D t(Including inlet nitrogen oxide concentration, flue gas flow rate, flue gas temperature, and outlet oxygen content, etc.). For each time t, the system concatenates the observation vectors from several past sampling periods into a historical state window, and concatenates the action vectors from several past sampling periods into an action queue. The length of the state window and the length of the action queue can be the same, or they can be set separately according to the ammonia injection action response time delay.

[0035] In one specific implementation, determining the outlet nitrogen oxide response sequence based on the aforementioned time-related data includes: arranging the time-related data based on historical ammonia injection control ranges to obtain historical ammonia injection action sequences; determining the time range corresponding to the historical ammonia injection action sequences, and generating an outlet nitrogen oxide response sequence based on the outlet nitrogen oxide concentration data within that time range. Specifically, when the ammonia injection valve position is adjusted from its current opening to a new opening at a certain moment, the change in outlet nitrogen oxide concentration at that moment and for a period thereafter is recorded, forming a response curve with time as the horizontal axis and outlet nitrogen oxide concentration as the vertical axis. This response curve reflects the time-dependent influence pattern of ammonia injection actions on the outlet nitrogen oxide concentration.

[0036] In this embodiment, Figure 2 This is a schematic diagram of the hierarchical structure of a denitrification control system based on ammonia injection response time delay modeling. The system corresponding to the embodiment of this application includes at least a data access module, a sample construction module, a causal time delay relationship construction module, a world model training module, a virtual environment encapsulation module, an agent training module, an online inference module, and a security filtering module.

[0037] The system comprises the following modules: a data access module for acquiring multi-source time-series data from DCS, CEMS, and production operation systems; a sample construction module for cleaning, time alignment, window construction, and training set segmentation of the acquired data, outputting a standardized training sample set; a causal time-delay relationship construction module for defining the influence direction and time-delay range between variables in the denitrification system, providing process prior constraints for the world model; a world model training module for learning the state transition dynamics model of the denitrification system conditioned on ammonia injection actions; a virtual environment encapsulation module for encapsulating the successfully trained world model into a virtual denitrification environment with reset and step interfaces; an agent training module for generating imagined trajectories and updating the ammonia injection strategy within the frozen world model; an online inference module for generating candidate values ​​for ammonia injection actions based on the current state during actual operation; and a safety filtering module for converting the strategy output into suggested actions that conform to on-site procedures, preventing unreasonable actions from being issued to the execution mechanism.

[0038] Step S12: Determine the relationship between the historical ammonia injection action sequence and the outlet nitrogen oxide response sequence, and determine the first response time from the execution of the historical ammonia injection action to the change of ammonia water flow rate during the denitrification reaction process. Then, determine the second response time from the change of ammonia water flow rate to the change of outlet nitrogen oxides, so as to determine the ammonia injection action response time delay relationship based on the relationship, the first response time and the second response time.

[0039] In this embodiment, the ammonia injection response time lag reflects the causal relationship between control actions and emission responses in the SCR denitrification system. Specifically, when the ammonia injection valve position in the control system is adjusted, the change in valve opening requires mechanical action of the actuator to be converted into an actual change in ammonia flow rate; this process involves a first response time. Subsequently, the ammonia gas with the changed flow rate undergoes a selective catalytic reduction reaction with nitrogen oxides in the flue gas in the catalyst layer. The reaction products continue to flow with the flue gas and eventually reach the outlet CEMS measuring point for detection; this process involves a second response time. The aforementioned first and second response times together constitute the ammonia injection response time lag.

[0040] In one specific embodiment, this application requires determining the relationship between the historical ammonia injection sequence and the outlet nitrogen oxide response sequence, and determining the first response time from the execution of the historical ammonia injection to the change in ammonia flow rate and the second response time from the change in ammonia flow rate to the change in outlet nitrogen oxides during the denitrification reaction. Specifically, this includes the following steps: First, an action change process was established based on historical ammonia injection sequence, and a pollutant response process was established based on the outlet nitrogen oxide response sequence. Then, the relationship between the aforementioned action change process and the pollutant response process was determined. Specifically, this relationship can be expressed as a decrease in outlet nitrogen oxide concentration following an increase in ammonia injection and an increase in outlet nitrogen oxide concentration following a decrease in ammonia injection. By analyzing the significance and consistency of these relationships, a preliminary verification can be made as to whether a process-reasonable causal relationship exists between ammonia injection and outlet nitrogen oxides.

[0041] Secondly, based on the changes in ammonia injection valve position, ammonia flow rate, and pump frequency in the historical ammonia injection action sequence, the execution response process corresponding to the ammonia injection control action is determined. Specifically, the time difference between the issuance of the ammonia injection valve position command and the actual change in ammonia flow rate is recorded, and the execution response times of multiple actions are statistically analyzed to obtain the statistical distribution characteristics of the first response time. This first response time reflects the process delay from the output of the control command to the actual change of the fluid state by the actuator (valve or pump), and is mainly affected by factors such as the mechanical inertia of the actuator, the pressure wave transmission speed in the pipeline, and the compressibility of the fluid.

[0042] Secondly, the change process of outlet nitrogen oxide concentration after the change of ammonia injection control action was determined based on the outlet nitrogen oxide response sequence. Specifically, the time difference between the moment the ammonia flow rate changed and the moment the outlet nitrogen oxide concentration actually began to change was recorded, and multiple response times were statistically analyzed to obtain the statistical distribution characteristics of the second response time. This second response time reflects the process delay from the injection of the reducing agent (ammonia) into the flue to the arrival of the reaction products at the outlet measuring point and their detection. It is mainly affected by factors such as the residence time of the flue gas in the flue and reactor, the reaction rate on the catalyst surface, and the transmission delay of the CEMS sampling pipeline. It should be particularly noted that in SCR denitrification systems, this second response time is usually much longer than the first response time, and the two should not be confused.

[0043] Finally, based on the above-mentioned changes, the first response time, and the second response time, the ammonia injection action response time lag relationship is generated. In a preferred embodiment, the ammonia injection action response time lag relationship is stored and represented in the form of a causal time lag relationship graph. This causal time lag relationship graph includes multiple nodes and directed edges. Nodes represent key variables in the denitrification system, and directed edges represent the causal influence direction between variables. Each edge is labeled with the minimum response time, typical response time, and maximum response time. The above-mentioned response time parameters can be estimated by lag correlation analysis of historical operating data. For example, by calculating the cross-correlation function (CCF) between the ammonia injection action and the outlet nitrogen oxides at different time lags, the lag time when the correlation coefficient reaches its peak is taken as the estimated value of the response time lag. However, it should be emphasized that the cross-correlation analysis is only used for time lag estimation. The results of data analysis cannot change the influence direction determined by the process prior, that is, the ammonia injection action is the cause, and the change in outlet nitrogen oxides is the result.

[0044] Among them, the above-mentioned ammonia injection action response time delay relationship includes at least the following types of process-determined relationship paths: (1) Causal relationship between production load or flue gas flow rate and inlet nitrogen oxide concentration. Changes in production load alter the intensity of flue gas emissions from combustion, and this causal relationship is effective from the present moment to several minutes in the future.

[0045] (2) Causal relationship between inlet nitrogen oxide concentration and outlet nitrogen oxide concentration. Nitrogen oxides in flue gas are transported with flue gas and reach the outlet measuring point after passing through the reactor. This causal relationship has a time lag of several minutes to more than ten minutes.

[0046] (3) Causal relationship between ammonia injection valve position or pump frequency and ammonia flow rate. After the actuator changes the valve opening or pump operating frequency, the flow rate in the ammonia pipeline changes, and this causal relationship takes effect from the current moment to several minutes later.

[0047] (4) Causal relationship between ammonia flow rate and outlet nitrogen oxide concentration. Changes in ammonia injection rate affect the degree of reduction reaction in the SCR reactor, and this causal relationship has a time lag of several minutes to more than ten minutes.

[0048] (5) Causal relationship between flue gas temperature and denitrification efficiency. Flue gas temperature affects the reactivity and catalytic efficiency of the catalyst surface, and this causal relationship is effective from the present moment to several minutes in the future.

[0049] (6) The causal relationship between the measured nitrogen oxide concentration at the export and the oxygen content at the export points to the nitrogen oxide concentration at the export oxygen content. The oxygen content is calculated according to the environmental protection conversion rules, and this causal relationship takes effect at the same moment or within a very short delay.

[0050] In this embodiment, at least three types of masks are established to constrain the learnable information paths of the model: (1) Time causal mask: When predicting the state at time t, only data at time t and before can be accessed. It is strictly forbidden to use data from future times to assist in the prediction of the current time.

[0051] (2) Variable role mask: Uncontrollable disturbances (such as inlet nitrogen oxide concentration and flue gas flow rate) cannot be controlled by the strategy output, and outlet nitrogen oxide concentration cannot be used as the reason for the current ammonia injection action.

[0052] (3) Action intervention mask: The ammonia injection valve position, ammonia water flow rate setpoint or pump frequency are entered into the dynamic model as external intervention variables. Active intervention modeling should be carried out rather than passive correlation analysis as ordinary correlation variables.

[0053] By constructing the time delay relationship of the ammonia injection action response and applying the three types of masks, the model can retain its data-driven learning ability during subsequent training, while effectively avoiding learning unreasonable patterns such as "future information leakage" or "erroneous control paths".

[0054] Step S13: Based on the time delay relationship of the ammonia injection action response, perform time delay correlation processing on the ammonia injection control data of the historical ammonia injection control range to obtain the ammonia injection influence characteristics, and construct a denitrification status prediction model based on the ammonia injection influence characteristics, the process monitoring data and the operating status data.

[0055] In this embodiment, the ammonia injection impact feature refers to a quantitative representation that reflects the degree of influence of historical ammonia injection actions on the current and future denitrification status. Unlike the traditional approach of directly using the current ammonia injection action as the model input, this embodiment, based on the time delay relationship of the ammonia injection action response, associates and combines ammonia injection actions from multiple past moments according to their time delay weights to form a feature vector that can characterize the historical cumulative impact of ammonia injection actions.

[0056] In one specific implementation, ammonia injection control data within the historical ammonia injection control range is processed by time-delay correlation based on the time-delay relationship of ammonia injection action response to obtain ammonia injection impact characteristics. Specifically, this includes: determining ammonia injection control data within the historical ammonia injection control range, and extracting ammonia injection valve position change, ammonia flow rate change, and pump frequency change from the ammonia injection control data; and constructing ammonia injection impact characteristics based on the time-delay relationship of ammonia injection action response and based on the aforementioned ammonia injection valve position change, ammonia flow rate change, and pump frequency change.

[0057] More specifically, for each sampling time t, an action queue is constructed: ; Among them, L a The length of the action queue is typically set based on the upper limit of the ammonia injection action response time delay. Then, the actions at each time step in the action queue are weighted and summed with their corresponding time delay weight coefficients to obtain the ammonia injection influence characteristic F. t Methods for determining the time delay weight coefficients include, but are not limited to: determining the peak position and decay rate of the weights based on the statistical distribution of the first and second response times; or using a learnable attention mechanism to automatically learn the importance weights for different lag times.

[0058] After constructing the characteristics of ammonia injection impact, a denitrification state prediction model is built based on these characteristics, process monitoring data, and operational status data. This denitrification state prediction model, also known as the World Model, describes the dynamic process of how the denitrification system evolves in the future, given its current state and ammonia injection actions.

[0059] In one specific implementation, constructing a denitrification status prediction model includes the following steps: First, an initial denitrification state prediction model is trained based on the characteristics of ammonia injection impact, process monitoring data, and operational status data to obtain a target denitrification state prediction model. This target denitrification state prediction model is then used to predict the first state prediction results for future control cycles. The first state prediction results include predicted values ​​of key variables such as outlet nitrogen oxide concentration, ammonia flow rate, and equipment operating status (e.g., low-temperature injection ban state, CEMS operating state).

[0060] Specifically, the denitrification state prediction model includes a state encoder, an action lag encoder, an action-conditional dynamics model, a state decoder, and a reward and constraint calculator.

[0061] The status encoder receives the historical status window: ; in, The historical state window at time t is a window that contains past states. A sequence of observation vectors at each time step; The field observation vector at time t represents the specific process monitoring and operating status variables, including inlet nitrogen oxide concentration, outlet nitrogen oxide concentration, flue gas flow rate, flue gas temperature, outlet oxygen content, reactor pressure difference, and production load. Indicates the state window length, in control cycles, used to determine the historical observation steps the model can backtrack during prediction; subscript to t Indicates from the past number The continuous time range from the previous moment to the current moment.

[0062] It is worth noting that the state encoder can be implemented using a Temporal Convolutional Network (TCN), a Gated Recurrent Unit (GRU), or a Causal Attention Network with a causal mask. Regardless of the network structure used, the encoding of the state at time t must strictly avoid using real data after time t to ensure causal consistency.

[0063] The motion delay encoder receives the motion queue: ; in, express The action queue matrix at each moment; express The ammonia injection action vector at any given moment (including ammonia injection valve position, ammonia flow rate setpoint, and pump frequency). The action queue length (in units of control cycles) is usually set based on the upper limit of the ammonia injection action response time delay to ensure that the action queue covers all historical action information that affects the current outlet nitrogen oxide concentration.

[0064] It is worth mentioning that the action delay of the ammonia injection action response is expressed as follows: The action lag is used to express the cumulative impact of ammonia injection actions over a number of minutes on current and future export indicators. Preferably, the action lag representation includes, but is not limited to: current action value, action difference, cumulative action amount, action change rate, and action value corresponding to the main lag order.

[0065] Action condition dynamics model based on potential state Action delay representation and the current perturbation vector Predict the next potential state : ; in, For the predicted next moment (i.e. t The latent state vector at time +1 is the latent space representation of the denitrification system at future times; For parameters A neural network dynamics model is used to fit the state transition laws of the denitrification system; For the current moment The latent state vector is obtained by the state encoder from the historical state window. Extracted from; The action delay representation vector is obtained from the action queue by the action delay encoder. Extracted from [the data], it is used to characterize the cumulative impact of historical ammonia injection actions on the present and future. For the current moment The uncontrollable disturbance vector specifically includes external disturbance variables such as inlet nitrogen oxide concentration, flue gas flow rate, flue gas temperature, and outlet oxygen content.

[0066] The state decoder will determine the next potential state. This involves reducing the variables to interpretable field variables, specifically predicting key variables such as outlet nitrogen oxide concentration, ammonia flow rate, and reactor pressure differential at the next time step. ; in, The predicted field observation vector for the next time step (i.e., time t+1) includes interpretable key process variables (such as outlet oxygen and nitrogen oxide concentrations, actual ammonia flow rate, valve position feedback, reactor differential pressure, etc.). For parameters The state decoder neural network is responsible for mapping the latent states in the latent space back to the original variable space; This is the potential state vector for the next moment, output by the action condition dynamics model.

[0067] The model training objectives include: prediction errors of key variables (mean squared error or mean absolute error), consistency error of action response (the outlet nitrogen oxide concentration should decrease after the ammonia injection action increases, and vice versa), multi-step rolling error (the cumulative error when the model's own prediction results are used as input for the next step of continuous multi-step prediction), and constraint judgment error (the classification and judgment error of constraints such as low temperature spray restriction state and action overrun). During training, the prediction weights of key variables such as outlet nitrogen oxide concentration, ammonia flow rate, valve position feedback, low temperature spray restriction state, and reactor pressure difference should be increased to avoid the model only pursuing the global average error while ignoring the prediction accuracy of key control quantities.

[0068] Secondly, based on the first-state prediction results, the target denitrification state prediction model was validated for prediction error, consistency of action response direction, and stability of multi-cycle prediction. Specifically, the model's single-step prediction error and multi-step rolling prediction error were evaluated on an independent test set; it was checked whether the predicted outlet nitrogen oxide concentration showed a decreasing trend when the input ammonia injection action increased, and whether the response direction was consistent with common process knowledge; it was checked whether the multi-step rolling prediction curve diverged. If the state curve showed unreasonable oscillations or divergence after multiple consecutive sampling cycles of prediction, the model failed the validation. The validated target denitrification state prediction model was then encapsulated as a virtual denitrification operating environment.

[0069] In this embodiment, Figure 3 This is a schematic diagram of the denitrification control process based on ammonia injection response time delay modeling: the process is divided into three stages: The system first collects data from multiple sources, then removes segments showing shutdown, maintenance, CEMS calibration, instrument purging, and obvious sensor anomalies. It then aligns the state, action, and disturbance variables according to the control cycle. After alignment, the system constructs a historical state window using data from several past sampling cycles, and constructs an action queue using ammonia injection valve positions, ammonia flow rates, or pump frequencies from several past sampling cycles. A single world model training sample can be represented as: ; in, The true state window at time t+1 (i.e., the observation value at the next time step) is used as the target label for model prediction.

[0070] It is worth mentioning that the following constraints need to be implemented when constructing the sample: (1) Time sequence constraint: The training set, validation set and test set must be divided in time sequence and random window shuffling is prohibited to ensure that the model evaluation can reflect the performance in the real online prediction scenario; (2) Operation segment constraint: The shutdown, maintenance, CEMS calibration, instrument purging and long-term missing segments must not be spliced ​​with the normal operation segments into the same training window; (3) Action legality constraint: Data with valve position below zero or above the upper limit, pump frequency out of range, and action changes exceeding the allowable limit of the equipment must be marked and cannot be directly used as normal control samples; (4) Oxygen conversion constraint: The outlet oxygen conversion nitrogen oxides can be directly provided by the field system, or calculated by the outlet measured nitrogen oxides and oxygen content according to the environmental protection conversion rules. The same conversion caliber must be used for training and deployment to ensure consistency; (5) Low temperature no-spray constraint: When the flue gas temperature is lower than the field allowable ammonia injection temperature or the equipment is in no-spray mode, this segment can be used to identify abnormal modes, but cannot be used to train normal ammonia injection strategies.

[0071] Subsequently, the system establishes a causal time-delay relationship graph by combining process priors and data response analysis, and then trains a world model that can predict the next state and multi-step states. After stability checks such as prediction error on the test set, action response direction, rolling stability, and constraint judgment, the world model is frozen and encapsulated as a virtual denitrification environment with reset and step interfaces. The reset interface initializes the virtual environment according to the state window at a certain historical moment, resetting the virtual environment's state to the specified historical operating point; the step interface receives the ammonia injection action output by the agent and returns the next virtual observation, immediate reward, end flag, and constraint information.

[0072] The reward and constraint calculator does not rely on neural networks for free learning, but rather calculates explicitly based on the denitrification control objectives. In a feasible specific scheme, the reward function takes the following form: ; Where R is the immediate reward value, used to evaluate the quality of the current ammonia spraying action, and the training objective of the agent is to maximize the cumulative reward. This is the penalty weighting coefficient for exceeding emission limits. It is a positive real number used to control the intensity of the penalty for exceeding the export nitrogen oxide standard in the reward. Its value is usually higher than other weights to prioritize ensuring compliance with emission standards. The current or next time point is the export concentration of oxygen and nitrogen oxides (in milligrams per standard cubic meter, mg / Nm³) predicted by the world model. For emission limits or internal control limits (unit: milligrams per standard cubic meter, mg / Nm³), when Exceeding this limit triggers a quadratic penalty; The ammonia consumption penalty weighting coefficient is a positive real number used to control the economy of ammonia injection and suppress excessive ammonia consumption. The instantaneous flow rate of ammonia water at the current moment (in liters per hour, L / h) reflects the direct material cost of ammonia injection; The smoothness penalty weight coefficient is a positive real number used to suppress drastic fluctuations in ammonia injection and extend the service life of the actuator. The change in ammonia injection action at the current moment relative to the ammonia injection action at the previous moment (e.g., percentage change in valve position or Hertz change in frequency), and its squared term is used to penalize large abrupt changes in action. The safety penalty weighting coefficient is a large positive real number. In one specific implementation, The value range is [0.2, 0.6], and it is generally taken as 0.5 to ensure that the safety constraint has the highest priority; To incorporate safety penalties, this item will take a large positive value when the system triggers safety risk events such as low-temperature spray ban, excessive ammonia escape prediction, abnormal reactor pressure difference, or excessive operation. Otherwise, it will take zero.

[0073] Furthermore, the constraint information should include at least: whether the export exceeds the limit, whether the action goes beyond the boundary, whether the action changes too quickly, whether the temperature is below the prohibited spray threshold, whether the CEMS is in a calibrated or faulty state, whether there is a risk of excessive ammonia spraying, and whether there is a risk of abnormal pressure difference.

[0074] Then, the control agent outputs actions and receives reward feedback within the frozen world model, learning the ammonia injection strategy through a large number of offline imagined trajectories. The agent takes current virtual observations, target limits, constraint margins, and historical action information as input, and outputs the ammonia injection valve position, ammonia flow rate setpoint, pump frequency, or a sequence of actions for the next few steps. Depending on the on-site control method, the actions can be a single continuous value or a combined pump and valve action.

[0075] In this embodiment, the system first trains an initial policy for behavior cloning using historical operation data, making the policy output close to the executable range in the field. Then, it further optimizes the policy in a frozen world model, thereby reducing the occurrence of a large number of unreasonable actions in the early stages of offline training. After offline training is completed, the agent is only responsible for generating candidate actions during online operation. The final action must be checked by a safety filtering module before it can be output to the operator or automatic control system. The system also retains records of input state, candidate actions, filtered actions, predicted trajectories, and constraint judgments.

[0076] Step S14: Utilize the denitrification state prediction model and construct a target ammonia injection optimization agent based on the historical ammonia injection action sequence, and generate candidate values ​​for ammonia injection control actions based on the online operation data of the denitrification system. Then, use the denitrification state prediction model and the target ammonia injection optimization agent to generate a target ammonia injection control action based on the candidate values ​​for ammonia injection control action, and use the target ammonia injection control action for denitrification control.

[0077] In one specific implementation, a target ammonia injection optimization agent is constructed using a denitrification status prediction model and based on historical ammonia injection action sequences, specifically including the following steps: First, in the virtual denitrification operation environment, based on each historical ammonia injection control action in the historical ammonia injection action sequence, the data to be processed, including the results of the change in outlet nitrogen oxide concentration, the results of ammonia water consumption, and the results of constraint judgment, are determined. Then, based on the above data to be processed, the initial ammonia injection optimization agent is trained to obtain the target ammonia injection optimization agent.

[0078] Specifically, in this embodiment, a historical state window needs to be extracted from the training set, and the virtual environment's reset interface is called to initialize the virtual denitrification environment to that historical operating point. The agent outputs candidate ammonia injection actions based on the current observations. The virtual environment calls the step interface to predict the next state, reward, and constraint information. The above process is repeated to form an imagined trajectory, which records the state, action, reward, constraint, and safety events. The agent parameters are updated based on the cumulative reward and constraint violations to make it more inclined to select actions that meet the standards, have low ammonia consumption, are stable, and safe.

[0079] It is worth mentioning that the above training process can be implemented using reinforcement learning algorithms, such as Proximal Policy Optimization (PPO) and Soft Actor-Critic (SAC). In a preferred embodiment, the system first trains an initial policy for a behavior clone using historical operation data, making the policy output close to the field's executable range, and then further optimizes it in a frozen world model.

[0080] Secondly, the trained strategy is evaluated on the initial state of the validation set, and evaluation metrics such as the success rate, average ammonia consumption, valve position fluctuation, number of out-of-limit occurrences, maximum out-of-limit magnitude, and number of safety filter triggers are calculated. Once the above metrics meet the preset requirements, the strategy model is deployed to the online inference module.

[0081] During the online inference phase, the system reads the latest runtime data each control cycle and constructs a state window using the same variable order and standardized parameters as in the training phase. Actions output by the agent must not be issued directly and unconditionally; they must pass through a security filtering module.

[0082] The security filtering includes the following steps: (1) Filtering of upper and lower limits of operation: The ammonia injection valve position, pump frequency and ammonia flow rate shall not exceed the allowable range of the equipment. For example, the ammonia injection valve position should be between 0% and 100%, and the pump frequency should be between the lowest and highest frequencies specified on the equipment nameplate.

[0083] (2) Action change rate filtering: The action change in adjacent control cycles shall not exceed the allowable change range of the actuator to prevent overload of the actuator or drastic fluctuations in pipeline pressure.

[0084] (3) Low temperature do not spray filtration: When the flue gas temperature is lower than the allowable ammonia injection temperature or the system is in do not spray mode, the automatic increase of ammonia injection is not allowed to prevent the catalyst from being deactivated at low temperature or generating ammonium bisulfate (ABS) to block the catalyst pores.

[0085] (4) CEMS status filtering: When CEMS is calibrated, purged, faulty or data quality is abnormal, the system outputs a manual review mark and does not directly use the model results for closed-loop control to avoid making inappropriate control decisions based on erroneous data.

[0086] (5) Emission risk filtering: If the world model predicts that the current candidate action has a risk of exceeding the limit in the future, the conservative ammonia injection action will be increased within the allowable range or manual handling will be prompted.

[0087] (6) Excessive ammonia injection filtration: If the predicted compliance margin is sufficient and the ammonia consumption is too high, the action should be restricted from continuing to increase or the ammonia injection action should be gradually reduced to reduce the risk of ammonia escape and the cost of ammonia water consumption.

[0088] After passing the safety filter, the system outputs the final suggested action, the action before filtering, the reason for filtering, the future predicted trajectory, and audit logs. These results can be displayed as operator suggestions on the human-machine interface, or, when authorized, written into the automatic control system to achieve closed-loop control.

[0089] Regarding validation metrics, the world model validation metrics should include at least: one-step prediction error for key variables, multi-step rolling prediction error, accuracy of judging oxygen and nitrogen oxide exceedances at the outlet, consistency of ammonia injection action response in the up and down directions, whether the rolling trajectory diverges, and stability of judgments regarding low-temperature injection bans and action overruns. Control strategy validation metrics should include at least: compliance rate, average ammonia consumption, ammonia consumption per unit volume of flue gas, valve position fluctuation amplitude, number of times the action change rate exceeds limits, number of times low-temperature injection bans are triggered, number of times excessive ammonia injection risks occur, number of times manual verification is triggered, and the maximum exceedance amplitude.

[0090] The system outputs include at least: training sample configuration, variable list, causal time-delay graph, world model parameters, standardized parameters, virtual environment configuration, strategy model parameters, security filtering configuration, test set evaluation report, action suggestion record, and audit log. These outputs constitute a complete configuration package for the denitrification intelligent control system, supporting system migration and deployment across different units.

[0091] In one specific implementation, in the SCR denitrification system of a 300MW coal-fired power plant, the original control method, which used manual setting of the ammonia injection valve position combined with PID feedback regulation, suffered from problems such as lag in ammonia injection regulation, large emission fluctuations, and high ammonia consumption. The main parameter ranges of the plant's denitrification system are as follows: inlet nitrogen oxide concentration 200–600 mg / Nm³, outlet nitrogen oxide concentration target limit 50 mg / Nm³, flue gas flow rate 800,000–1,200,000 Nm³ / h, flue gas temperature 280–420℃, ammonia water flow rate setpoint range 0–1500 L / h, and ammonia injection valve position range 0–100%.

[0092] In implementing the technical solution of this application, a 12-month historical operating data collection period is first conducted. Data such as ammonia injection valve position, ammonia flow setpoint, pump frequency, flue gas temperature, and reactor differential pressure are collected from the DCS system, with a sampling period of 5 seconds; data such as inlet nitrogen oxide concentration, outlet nitrogen oxide concentration, and outlet oxygen content are collected from the CEMS system, with a sampling period of 60 seconds; and data such as production load and start / stop status are collected from the production operation system, with an update period of 1 minute. Using the control action update period (60 seconds) as the main time axis, the DCS data sampled every 5 seconds is aggregated into a 60-second average value, and the production operation data updated every minute is aligned to the 60-second time axis according to the most recent valid value.

[0093] Then, abnormal data segments were identified and removed. Based on the equipment operation logs, 7 CEMS periodic calibration segments (approximately 30 minutes each), 3 shutdown maintenance segments, multiple instrument purging segments, and segments with obvious sensor anomalies (such as values ​​exceeding reasonable ranges or values ​​remaining unchanged for extended periods) were marked and removed. The cleaned continuous operating data is approximately 10 months of valid operating data.

[0094] Based on this, sample construction is performed. The state window length Ls is set to 10 control cycles (i.e., a 10-minute historical state window), and the action queue length L... a It is also set to 10 control periods. A single training sample is represented as (S t A t , d t ,S {t+1} The samples were divided into training set (70%), validation set (15%), and test set (15%) in chronological order.

[0095] In the causal time-delay relationship construction stage, the main response time delays were estimated through cross-correlation analysis. The analysis results show that the response time delay from the change in ammonia injection valve position to the change in ammonia flow rate is approximately 30–90 seconds (first response time), the response time delay from the change in ammonia flow rate to the change in outlet nitrogen oxides is approximately 4–8 minutes (second response time), and the total ammonia injection response time delay is approximately 5–9 minutes. These time-delay estimates are basically consistent with the process design parameters and field experience. Based on these analysis results, the action queue length L was determined. a Cover the upper limit of the total response time delay (10 minutes) to ensure that the action queue contains all historical action information that has an impact on the current outlet nitrogen oxide concentration.

[0096] During the world model training phase, a gated recurrent unit (GRU) was used as the core network structure for the state encoder and action delay encoder, and a multilayer perceptron (MLP) was used as the action conditional dynamics model and state decoder. During training, the prediction weight for the outlet oxygen and nitrogen oxide concentration was specifically increased (set to three times the weight of other variables). Evaluation results on the test set showed that the mean absolute error of the one-step prediction of the outlet oxygen and nitrogen oxide concentration was 2.3 mg / Nm³, the mean absolute error of the multi-step rolling prediction (10 steps, i.e., 10 minutes) was 4.1 mg / Nm³, the accuracy of exceeding the limit judgment reached 96.7%, the consistency of the ammonia injection action's lifting and lowering direction response was 98.2%, and no divergence was observed in the multi-step rolling trajectory. All of the above indicators met the stability verification requirements, and the world model was encapsulated as a virtual denitrification environment.

[0097] During the agent training phase, an initial policy for behavioral cloning was first trained using historical operational data to ensure that the policy output values, such as the ammonia injection valve position and ammonia flow rate setpoint, remained within the historical ammonia injection control range. Then, offline policy optimization was performed using the PPO algorithm in a frozen virtual environment. After approximately 100,000 steps of imagined trajectory training, the policy achieved a 98.5% compliance rate on the validation set, with average ammonia consumption reduced by approximately 8.7% and valve position fluctuation amplitude reduced by approximately 22% compared to historical operational data.

[0098] During online deployment, the safety filtering module checks each candidate action output by the agent. For example, in one online operation, the agent outputs a candidate ammonia injection valve position of 78%, but the current flue gas temperature is 295°C, lower than the minimum allowable ammonia injection temperature for the catalyst (300°C). The safety filtering module triggers low-temperature injection restriction filtration, limiting the valve position to the current 65% opening to prevent the risk of catalyst deactivation at low temperatures. As another example, in another operation, the CEMS system is in periodic calibration mode, and the CEMS outputs the outlet nitrogen oxide concentration as the standard gas concentration rather than the actual emission concentration. After detecting the CEMS status flag, the safety filtering module rejects the agent's output and prompts for manual verification.

[0099] After three consecutive months of online trial operation, the compliance rate of the denitrification intelligent control system has stabilized at over 99%, the average ammonia consumption has decreased by about 7.2% compared to before the trial operation, the number of manual verification triggers has decreased from several times a day in the early stage to 1-2 times a week, and the operator's adoption rate of system output suggestions has increased from less than 60% in the early stage to over 95%, which fully verifies the practicality and effectiveness of this application.

[0100] As can be seen from the above, this application first needs to perform time correlation processing on historical denitrification system operation data based on historical ammonia injection control range to obtain time correlation data, and then determine the outlet nitrogen oxide response sequence based on the time correlation data. Next, it determines the change relationship between the historical ammonia injection action sequence and the outlet nitrogen oxide response sequence, and determines the first response time from the execution of historical ammonia injection actions to the change in ammonia flow rate and the second response time from the change in ammonia flow rate to the change in outlet nitrogen oxides during the denitrification reaction. Based on the above change relationship and the above response time, it determines the ammonia injection action response time delay relationship, and constructs an ammonia injection influence characteristic and a denitrification state prediction model based on the above time delay relationship. Finally, it uses the above denitrification state prediction model and the historical ammonia injection action sequence to construct a target ammonia injection optimization agent, and uses the target ammonia injection optimization agent to generate ammonia injection control actions to achieve denitrification control. In this way, the accuracy and safety of ammonia injection control are improved in the denitrification control process based on ammonia injection response time delay modeling. While ensuring that nitrogen oxide emissions do not exceed the standard, ammonia consumption is significantly reduced, ammonia escape is reduced, and the safety of system operation is ensured through a safety filtration module.

[0101] Accordingly, see Figure 4 As shown, this application also provides a denitrification control device based on ammonia injection response time delay modeling, comprising: The time-correlation data generation module 11 is used to perform time-correlation processing on historical denitrification system operation data based on historical ammonia injection control range to obtain time-correlation data, and to determine the outlet nitrogen oxide response sequence based on the time-correlation data; the historical denitrification system operation data includes process monitoring data, ammonia injection control data, and operation status data; The ammonia injection action response time delay relationship determination module 12 is used to determine the change relationship between the historical ammonia injection action sequence and the outlet nitrogen oxide response sequence, and to determine the first response time from the execution of the historical ammonia injection action to the change of ammonia water flow rate during the denitrification reaction, and then to determine the second response time from the change of ammonia water flow rate to the change of outlet nitrogen oxides, so as to determine the ammonia injection action response time delay relationship based on the change relationship, the first response time and the second response time; The denitrification status prediction model construction module 13 is used to perform time-delay correlation processing on the ammonia injection control data of the historical ammonia injection control range based on the ammonia injection action response time delay relationship, to obtain ammonia injection influence characteristics, and to construct a denitrification status prediction model based on the ammonia injection influence characteristics, the process monitoring data and the operating status data. The ammonia injection control action generation module 14 is used to construct a target ammonia injection optimization agent based on the denitrification state prediction model and the historical ammonia injection action sequence, and generate ammonia injection control action candidate values ​​based on the online operation data of the denitrification system, so as to use the denitrification state prediction model and the target ammonia injection optimization agent, and generate a target ammonia injection control action based on the ammonia injection control action candidate values, so as to use the target ammonia injection control action for denitrification control.

[0102] In one specific embodiment, the time-related data generation module 11 specifically includes: The data acquisition unit is used to collect historical denitrification system operation data from the corresponding DCS system, CEMS system, and production operation system of the denitrification system. The data classification unit is used to acquire process monitoring data, ammonia injection control data, and operating status data based on the historical denitrification system operation data. An anomaly removal unit is used to identify abnormal data segments that are abnormal in the process monitoring data, the ammonia injection control data, and the operating status data, and then remove each of the abnormal data segments from the process monitoring data, the ammonia injection control data, and the operating status data to obtain continuous operating data; The time correlation processing unit is used to perform time correlation processing on the continuous operation data based on the historical ammonia injection control range to obtain time correlation data.

[0103] In some specific embodiments, the time-related data generation module 11 may specifically include: An action sequence generation unit is used to arrange the time-related data based on the historical ammonia injection control range to obtain a historical ammonia injection action sequence. The response sequence generation unit is used to determine the time range corresponding to the historical ammonia injection action sequence, and generate an outlet nitrogen oxide response sequence based on the outlet nitrogen oxide concentration data within the time range.

[0104] In some specific embodiments, the ammonia injection action response time delay relationship determination module 12 may specifically include: The change relationship determination unit is used to establish an action change process based on the historical ammonia injection action sequence, and to establish a pollutant response process based on the outlet nitrogen oxide response sequence, and then determine the change relationship between the action change process and the pollutant response process; The execution response determination unit is used to determine the execution response process corresponding to the ammonia injection control action based on the changes in ammonia injection valve position, ammonia flow rate, and pump frequency in the historical ammonia injection action sequence. The concentration change determination unit is used to determine the process of outlet nitrogen oxide concentration change after the change of ammonia injection control action based on the outlet nitrogen oxide response sequence. The response time determination unit is used to determine, based on the execution response process and the outlet nitrogen oxide concentration change process, the first response time from the ammonia injection control action to the change in ammonia flow rate and the second response time from the change in ammonia flow rate to the change in outlet nitrogen oxide concentration; The time delay relationship generation unit is used to generate ammonia injection action response time delay relationship based on the change relationship, the first response time and the second response time.

[0105] In some specific embodiments, the denitrification state prediction model construction module 13 may specifically include: The change extraction unit is used to determine the ammonia injection control data within the historical ammonia injection control range, and extract the ammonia injection valve position change, ammonia water flow rate change, and pump frequency change from the ammonia injection control data. The influence feature construction unit is used to construct ammonia injection influence features based on the ammonia injection action response time delay relationship and based on the ammonia injection valve position change, the ammonia water flow rate change, and the pump frequency change.

[0106] In some specific embodiments, the denitrification state prediction model construction module 13 may specifically include: The model training unit is used to train an initial denitrification state prediction model based on the ammonia injection influence characteristics, the process monitoring data, and the operating status data, to obtain a target denitrification state prediction model, and to use the target denitrification state prediction model to predict the first state prediction results corresponding to the future control cycle, including the outlet nitrogen oxide concentration, ammonia water flow rate, and equipment operating status. The model verification unit is used to verify the prediction error, action response direction consistency and multi-cycle prediction stability of the target denitrification state prediction model based on the first state prediction result, and to encapsulate the target denitrification state prediction model that has passed the verification into a virtual denitrification operation environment.

[0107] In some specific embodiments, the ammonia injection control action generation module 14 may specifically include: The agent training unit is used to determine the data to be processed, including the results of the change in outlet nitrogen oxide concentration, the results of ammonia water consumption, and the results of constraint judgment, based on each historical ammonia injection control action in the historical ammonia injection action sequence in the virtual denitrification operation environment. Then, the initial ammonia injection optimization agent is trained based on the data to be processed to obtain the target ammonia injection optimization agent. The candidate value generation unit is used to acquire the online operating data of the denitrification system and generate candidate values ​​for the initial ammonia injection control action corresponding to the current control cycle based on the online operating data. The state prediction unit is used to determine the action evaluation result corresponding to each of the initial ammonia injection control action candidate values, and to perform future state prediction based on the initial ammonia injection control action candidate values ​​in the virtual denitrification operation environment to obtain a second state prediction result. The candidate value adjustment unit is used to optimize the agent by utilizing the target ammonia injection and adjust the initial ammonia injection control action candidate value based on the second state prediction result and the action evaluation result to obtain the target ammonia injection control action candidate value; the target ammonia injection control action candidate value satisfies the preset outlet nitrogen oxide emission constraint condition and the preset ammonia water consumption condition. A safety verification unit is used to perform safety verification on the candidate values ​​of the target ammonia injection control action, and set the candidate values ​​of the target ammonia injection control action that pass the safety verification as the target ammonia injection control action, so as to use the target ammonia injection control action for denitrification control.

[0108] Furthermore, embodiments of this application also disclose an electronic device, Figure 5This is a structural diagram of an electronic device 20 according to an exemplary embodiment. The content of the diagram should not be construed as limiting the scope of this application. Specifically, the electronic device 20 may include: at least one processor 21, at least one memory 22, a power supply 23, a communication interface 24, an input / output interface 25, and a communication bus 26. The memory 22 stores a computer program, which is loaded and executed by the processor 21 to implement the relevant steps in the denitrification control method based on ammonia injection response time delay modeling disclosed in any of the foregoing embodiments. Furthermore, the electronic device 20 in this embodiment may specifically be an electronic computer.

[0109] 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.

[0110] 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.

[0111] 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 denitrification control method based on ammonia injection response time delay modeling disclosed in any of the foregoing embodiments, the computer program 222 may further include computer programs capable of performing other specific tasks.

[0112] 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 based on ammonia injection response time delay modeling. Specific steps of this method can be found in the corresponding content disclosed in the foregoing embodiments, and will not be repeated here.

[0113] 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.

[0114] 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.

[0115] 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.

[0116] 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.

[0117] 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 denitrification control method based on ammonia injection response time delay modeling, characterized in that, include: Based on the historical ammonia injection control range, the historical denitrification system operation data is subjected to time correlation processing to obtain time correlation data, and the outlet nitrogen oxide response sequence is determined based on the time correlation data; the historical denitrification system operation data includes process monitoring data, ammonia injection control data, and operation status data; The relationship between the historical ammonia injection sequence and the outlet nitrogen oxide response sequence is determined, and the first response time from the execution of the historical ammonia injection to the change in ammonia flow rate during the denitrification reaction is determined. Then, the second response time from the change in ammonia flow rate to the change in outlet nitrogen oxides is determined, so as to determine the ammonia injection response time lag relationship based on the relationship, the first response time and the second response time. Based on the time delay relationship of the ammonia injection action response, the ammonia injection control data of the historical ammonia injection control range is subjected to time delay correlation processing to obtain the ammonia injection influence characteristics. Based on the ammonia injection influence characteristics, the process monitoring data and the operating status data, a denitrification status prediction model is constructed. The denitrification state prediction model is used to construct a target ammonia injection optimization agent based on the historical ammonia injection action sequence. Based on the online operation data of the denitrification system, candidate values ​​for ammonia injection control actions are generated. The target ammonia injection control action is then generated based on the candidate values ​​using the denitrification state prediction model and the target ammonia injection optimization agent. The denitrification control is then performed using the target ammonia injection control action.

2. The denitrification control method based on ammonia injection response time delay modeling according to claim 1, characterized in that, The time-correlation processing of historical denitrification system operation data based on historical ammonia injection control range yields time-correlation data, including: Historical operating data of the denitrification system are collected from the distributed control system (DCS), the flue gas online monitoring system (CEMS), and the production operation system corresponding to the denitrification system. Based on the historical denitrification system operation data, process monitoring data, ammonia injection control data, and operating status data are acquired. The process monitoring data includes inlet nitrogen oxide concentration, outlet nitrogen oxide concentration, outlet oxygen content, and reactor pressure difference. The ammonia injection control data includes ammonia injection valve position, ammonia water flow rate setpoint, and pump frequency. The operating status data includes flue gas flow rate, flue gas temperature, and production load. Identify the abnormal data segments that occur in the process monitoring data, the ammonia injection control data, and the operating status data, and then remove each of the abnormal data segments from the process monitoring data, the ammonia injection control data, and the operating status data to obtain continuous operating data; the abnormal data segments are data segments that have occurred during shutdown, maintenance, CEMS calibration abnormality, instrument purging abnormality, and sensor abnormality. The continuous operation data is processed by time correlation based on the historical ammonia injection control range to obtain time-correlated data.

3. The denitrification control method based on ammonia injection response time delay modeling according to claim 2, characterized in that, The determination of the export nitrogen oxide response sequence based on the time-correlation data includes: The time-related data are arranged based on the historical ammonia injection control range to obtain a historical ammonia injection action sequence; The time range corresponding to the historical ammonia injection action sequence is determined, and an outlet nitrogen oxide response sequence is generated based on the outlet nitrogen oxide concentration data within the time range.

4. The denitrification control method based on ammonia injection response time delay modeling according to claim 3, characterized in that, The process of determining the relationship between the historical ammonia injection sequence and the outlet nitrogen oxide response sequence, and determining the first response time from the execution of the historical ammonia injection to the change in ammonia flow rate during the denitrification reaction, and then determining the second response time from the change in ammonia flow rate to the change in outlet nitrogen oxides, to determine the ammonia injection response time lag relationship based on the relationship, the first response time, and the second response time, includes: The action change process is established based on the historical ammonia injection action sequence, and the pollutant response process is established based on the outlet nitrogen oxide response sequence. Then, the relationship between the action change process and the pollutant response process is determined. Based on the changes in ammonia injection valve position, ammonia flow rate, and pump frequency in the historical ammonia injection action sequence, the execution response process corresponding to the ammonia injection control action is determined. The process of change in outlet nitrogen oxide concentration after the change in ammonia injection control action is determined based on the outlet nitrogen oxide response sequence. Based on the execution response process and the outlet nitrogen oxide concentration change process, determine the first response time from the ammonia injection control action to the change in ammonia flow rate and the second response time from the change in ammonia flow rate to the change in outlet nitrogen oxide concentration. Based on the aforementioned change relationship, the first response time, and the second response time, a time delay relationship for ammonia injection action response is generated.

5. The denitrification control method based on ammonia injection response time delay modeling according to claim 1, characterized in that, The ammonia injection control data within the historical ammonia injection control range is subjected to time-delay correlation processing based on the ammonia injection action response time-delay relationship to obtain ammonia injection impact characteristics, including: Determine the ammonia injection control data within the historical ammonia injection control range, and extract the ammonia injection valve position change, ammonia water flow rate change, and pump frequency change from the ammonia injection control data; Based on the time lag relationship of the ammonia injection action response, and based on the changes in the ammonia injection valve position, the changes in the ammonia flow rate, and the changes in the pump frequency, the ammonia injection influence characteristics are constructed.

6. The denitrification control method based on ammonia injection response time delay modeling according to claim 1, characterized in that, The denitrification status prediction model constructed based on the ammonia injection impact characteristics, the process monitoring data, and the operating status data includes: Based on the ammonia injection impact characteristics, the process monitoring data, and the operating status data, an initial denitrification state prediction model is trained to obtain a target denitrification state prediction model. The target denitrification state prediction model is then used to predict the first state prediction results corresponding to the future control cycle, including the outlet nitrogen oxide concentration, ammonia water flow rate, and equipment operating status. Based on the first state prediction result, the prediction error, action response direction consistency and multi-cycle prediction stability of the target denitrification state prediction model are verified, and the verified target denitrification state prediction model is encapsulated as a virtual denitrification operation environment.

7. The denitrification control method based on ammonia injection response time delay modeling according to claim 6, characterized in that, The process of constructing a target ammonia injection optimization agent based on the denitrification state prediction model and the historical ammonia injection action sequence, and generating candidate values ​​for ammonia injection control actions based on the online operation data of the denitrification system, thereby generating a target ammonia injection control action based on the candidate values, and using the target ammonia injection control action for denitrification control, includes: In the virtual denitrification operation environment, based on each historical ammonia injection control action in the historical ammonia injection action sequence, data to be processed is determined, including the results of changes in outlet nitrogen oxide concentration, ammonia water consumption, and constraint judgment results. Then, based on the data to be processed, the initial ammonia injection optimization agent is trained to obtain the target ammonia injection optimization agent. Acquire the online operating data of the denitrification system, and generate candidate values ​​for the initial ammonia injection control action corresponding to the current control cycle based on the online operating data; Determine the action evaluation results corresponding to each of the initial ammonia injection control action candidate values, and predict the future state based on the initial ammonia injection control action candidate values ​​in the virtual denitrification operation environment to obtain the second state prediction result. The target ammonia injection optimization agent is used, and the initial ammonia injection control action candidate value is adjusted based on the second state prediction result and the action evaluation result to obtain the target ammonia injection control action candidate value; the target ammonia injection control action candidate value satisfies the preset outlet nitrogen oxide emission constraint condition and the preset ammonia water consumption condition. The candidate values ​​of the target ammonia injection control action are subjected to safety verification, and the candidate values ​​of the target ammonia injection control action that pass the safety verification are set as the target ammonia injection control action, so as to use the target ammonia injection control action for denitrification control.

8. A denitrification control device based on ammonia injection response time delay modeling, characterized in that, include: The time-correlation data generation module is used to perform time-correlation processing on historical denitrification system operation data based on historical ammonia injection control range to obtain time-correlation data, and to determine the outlet nitrogen oxide response sequence based on the time-correlation data; the historical denitrification system operation data includes process monitoring data, ammonia injection control data, and operation status data; The ammonia injection action response time delay relationship determination module is used to determine the change relationship between the historical ammonia injection action sequence and the outlet nitrogen oxide response sequence, and to determine the first response time from the execution of the historical ammonia injection action to the change of ammonia water flow rate during the denitrification reaction, and then to determine the second response time from the change of ammonia water flow rate to the change of outlet nitrogen oxides, so as to determine the ammonia injection action response time delay relationship based on the change relationship, the first response time and the second response time; The denitrification status prediction model construction module is used to perform time-delay correlation processing on the ammonia injection control data of the historical ammonia injection control range based on the time-delay relationship of the ammonia injection action response, to obtain the ammonia injection influence characteristics, and to construct a denitrification status prediction model based on the ammonia injection influence characteristics, the process monitoring data and the operating status data. The ammonia injection control action generation module is used to construct a target ammonia injection optimization agent based on the denitrification state prediction model and the historical ammonia injection action sequence, and to generate candidate values ​​for ammonia injection control actions based on the online operation data of the denitrification system. The module then uses the denitrification state prediction model and the target ammonia injection optimization agent to generate a target ammonia injection control action based on the candidate values, and uses the target ammonia injection control action to perform denitrification control.

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

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

Images