An SCR denitration system control method, device, equipment and storage medium

Abstract

This application relates to a control method, apparatus, equipment, and storage medium for an SCR denitrification system, belonging to the field of flue gas denitrification technology in thermal power plants. The method includes obtaining a priori predicted value of the outlet nitrogen oxide concentration based on a pre-trained time-series prediction model and historical operating data; obtaining an observed estimated value of the outlet nitrogen oxide concentration based on a pre-trained operating condition regression model and current operating condition characteristics; updating noise statistical parameters online using an adaptive algorithm, and using the noise statistical parameters to weightedly fuse the priori predicted value and the observed estimated value to obtain a posterior estimated value of the outlet nitrogen oxide concentration; setting an initial control target based on historical outlet nitrogen oxide concentration data, and dynamically correcting the initial control target according to real-time emission status to obtain a dynamic target value; calculating the deviation between the posterior estimated value and the dynamic target value, and adjusting the opening of the ammonia injection regulating valve according to the deviation. This solves the problem of difficulty in achieving precise closed-loop control of denitrification systems in existing technologies.

Description

Technical Field

[0001] This invention relates to the field of flue gas denitrification technology in thermal power plants, specifically to a control method, device, equipment, and storage medium for an SCR denitrification system. Background Technology

[0002] Selective catalytic reduction (SCR) is currently the most widely used technology for denitrification of nitrogen oxides (NOx) in flue gas from coal-fired power plants and industrial boilers. x SCR (Self-Reduction) is a mainstream emission reduction technology. The core control objective of the SCR system is to reduce NO2 at the outlet by adjusting the ammonia injection rate. x The concentration should be controlled below environmental limits, while minimizing ammonia slip rate to reduce operating costs and secondary pollution. However, the SCR denitrification process is a typical industrial process with large time lag, strong nonlinearity, and time-varying complexity. Its operating status is affected by a combination of factors such as boiler load, coal quality, flue gas temperature, and catalyst activity. Achieving accurate prediction and control has always been a technical challenge.

[0003] In existing technologies, data-driven models based on single algorithms such as Long Short-Term Memory (LSTM), Support Vector Machine (SVR), or Extreme Gradient Boosting Tree (XGBoost) are often constructed to predict export NO. x For concentration, some schemes further introduce the Kalman filter algorithm to process sensor data. By constructing a linear state-space model and pre-setting a fixed process noise covariance matrix Q and measurement noise covariance matrix R, recursive state estimation is performed. As for the setting of control targets, fixed target values ​​are usually generated by multiplying the environmental emission standards by a safety factor or manually set by the operators. The control system performs closed-loop adjustment based on the deviation between the target value and the feedback value.

[0004] However, a single data-driven model struggles to balance the dynamic nature of time series with the coupling of multiple features. SCR denitrification systems exhibit strong nonlinearity and large time lag. Using only time series models such as LSTM can easily overlook the combined influence of multiple features under the current operating conditions, while using only regression models such as XGBoost can hardly capture the long-term dependence and periodic fluctuations of the time series, leading to a decrease in prediction accuracy under complex and variable operating conditions, which in turn affects the timeliness of the control system. Furthermore, the traditional Kalman filter algorithm makes overly idealistic assumptions about the statistical characteristics of noise. The operating conditions of SCR systems often fluctuate, and the noise characteristics are time-varying. Fixed parameters can easily lead to slow convergence or even divergence of the filter, making it difficult to provide reliable state estimates for the control system and resulting in the problem of difficulty in achieving accurate closed-loop control of the denitrification system. Summary of the Invention

[0005] This application provides a control method, device, equipment, and storage medium for an SCR denitrification system. It can solve the problem that existing single data-driven models cannot take into account both time-series dynamics and multi-feature coupling. SCR denitrification systems have strong nonlinearity and large time lag characteristics. Using only time-series models such as LSTM can easily ignore the comprehensive influence of multiple features of the current operating condition. Using only regression models such as XGBoost can hardly capture the long-term dependence and periodic fluctuations of the time series, resulting in a decrease in prediction accuracy under complex variable operating conditions, which in turn affects the timeliness of the control system. Furthermore, the traditional Kalman filter algorithm makes overly idealistic assumptions about the statistical characteristics of noise. The operating conditions of the SCR system often fluctuate, and the noise characteristics are time-varying. Fixed parameters can easily lead to slow convergence or even divergence of the filter, making it difficult to provide reliable state estimates for the control system. This presents the problem of difficulty in achieving accurate closed-loop control of the denitrification system.

[0006] In a first aspect, embodiments of this application provide a control method for an SCR denitrification system, comprising: Based on a pre-trained time-series prediction model and historical operating data, a priori predicted value of the outlet nitrogen oxide concentration is obtained. Based on the pre-trained working condition regression model and the current working condition characteristics, the observed estimate of the outlet nitrogen oxide concentration is obtained; The noise statistics parameters are updated online based on an adaptive algorithm. The prior predicted values ​​and observed estimates are then weighted and fused using the noise statistics parameters to obtain the posterior estimate of the outlet nitrogen oxide concentration. An initial control target is set based on historical outlet nitrogen oxide concentration data, and the initial control target is dynamically corrected according to the real-time emission status to obtain a dynamic target value. Calculate the deviation between the posterior estimate and the dynamic target value, and adjust the opening of the ammonia injection regulating valve according to the deviation.

[0007] In one implementation, obtaining the prior prediction value of the outlet nitrogen oxide concentration based on the pre-trained time-series prediction model and historical operating data includes: Based on historical operational data, historical sequences of outlet nitrogen oxide concentrations and historical windows of auxiliary variables are extracted, a sliding time window is set, and an input feature matrix is ​​constructed. The input feature matrix is ​​transformed in the frequency domain using a time series prediction model to extract the dominant period, and the one-dimensional time series data is reconstructed into a two-dimensional matrix form based on the dominant period. By using multiple convolutional branches of different scales in the time series prediction model to extract features from the two-dimensional matrix, a multi-scale feature map is obtained. Multi-scale feature maps are fused, and the fused features are weighted using the temporal attention mechanism in the temporal prediction model to obtain prior prediction values.

[0008] In one implementation, obtaining the observed estimate of the outlet nitrogen oxide concentration based on the pre-trained operating condition regression model and current operating condition characteristics includes: Obtain various operating parameters at the current moment and construct a feature vector of the current operating condition; The current operating condition feature vector is input into a pre-trained operating condition regression model. The operating condition regression model is used to explore the nonlinear mapping relationship between the current operating condition feature vector and the outlet nitrogen oxide concentration, and the observed estimated value is output.

[0009] In one implementation, the method of updating noise statistical parameters online based on an adaptive algorithm, and using the noise statistical parameters to weightedly fuse prior predicted values ​​and observed estimated values ​​to obtain a posterior estimated value of the outlet nitrogen oxide concentration, includes: A Kalman filter framework is constructed, in which the prior predictions are used as the output of the state equation and the observation estimates are used as the output of the observation equation. Based on an adaptive algorithm, the process noise covariance matrix and measurement noise covariance matrix are updated online using historical residual sequences; The Kalman gain is calculated based on the updated process noise covariance matrix and measurement noise covariance matrix. The posterior estimate is obtained by weighting the deviation between the prior prediction and the observed estimate based on Kalman gain and correcting the prior prediction.

[0010] In one implementation, the online updating of the process noise covariance matrix and the measurement noise covariance matrix using historical residual sequences includes: Calculate the residual between the prior prediction and the observed estimate at the current time. By introducing a forgetting factor, the statistical properties of the noise covariance matrix and residual at the previous time step are weighted and averaged to obtain the process noise covariance matrix and measurement noise covariance matrix at the current time step.

[0011] In one implementation, the step of setting an initial control target based on historical outlet nitrogen oxide concentration data and dynamically correcting the initial control target according to real-time emission status to obtain a dynamic target value includes: Historical outlet nitrogen oxide concentration data during the initial stage of system operation were selected as the sample set, and the statistical quantile values ​​of the sample set were calculated as the initial control target. Use the initial control target or the dynamic target value updated in the previous time step as the current target value; The dynamic target value is updated based on the relationship between the export nitrogen oxide concentration and the environmental protection limit and the current target value.

[0012] In one implementation, updating the dynamic target value based on the relationship between the export nitrogen oxide concentration and environmental limits and the current target value includes: If the concentration of nitrogen oxides in the export exceeds the environmental protection limit for a continuous period of time, the current target value will be reduced by the first preset ratio to obtain the updated dynamic target value. If the concentration of nitrogen oxides in the export is continuously lower than the current target value and does not exceed the environmental protection limit for a second preset time, the current target value will be adjusted upward according to the second preset ratio to obtain the updated dynamic target value. If the concentration of nitrogen oxides in exports is within the preset fluctuation range of the current target value and does not exceed the environmental protection limit, or does not meet the conditions for downward or upward adjustment, then the current target value will be maintained.

[0013] Secondly, embodiments of this application also provide an SCR denitrification system control device, which includes: The prior prediction value acquisition module is used to obtain the prior prediction value of the outlet nitrogen oxide concentration based on the pre-trained time series prediction model and historical operating data. The observation estimate acquisition module is used to obtain the observation estimate of the outlet nitrogen oxide concentration based on the pre-trained working condition regression model and the current working condition characteristics. The posterior estimate acquisition module is used to update noise statistical parameters online based on an adaptive algorithm, and to use the noise statistical parameters to perform weighted fusion of prior predictions and observation estimates to obtain the posterior estimate of the outlet nitrogen oxide concentration. The dynamic target value acquisition module is used to set an initial control target based on historical outlet nitrogen oxide concentration data, and dynamically correct the initial control target according to the real-time emission status to obtain the dynamic target value. The calculation and execution module is used to calculate the deviation between the posterior estimate and the dynamic target value, and adjust the opening of the ammonia injection regulating valve according to the deviation.

[0014] Thirdly, this application also provides an SCR denitrification system control device, which includes a processor, a memory, and an SCR denitrification system control program stored in the memory and executable by the processor. When the SCR denitrification system control program is executed by the processor, it implements the steps of the above-described SCR denitrification system control method.

[0015] Fourthly, embodiments of this application also provide a storage medium storing an SCR denitrification system control program, wherein when the SCR denitrification system control program is executed by a processor, it implements the steps of the above-described SCR denitrification system control method.

[0016] The beneficial effects of the technical solutions provided in this application include: By constructing a dual-model-driven prediction framework, a time-series prediction model is used to capture historical evolution patterns and obtain prior predictions. A working condition regression model is then combined to mine multi-feature nonlinear mappings and obtain observation estimates. This effectively overcomes the shortcomings of a single model in simultaneously addressing the dynamic nature of time series and the coupling of multiple features, significantly improving prediction accuracy under complex and varying working conditions. Furthermore, an adaptive algorithm is introduced to update noise statistical parameters online, weighting and fusing prior predictions and observation estimates. This solves the problem of slow convergence or divergence caused by fixed noise parameters in traditional Kalman filtering, enhancing the robustness and adaptability of state estimation. Simultaneously, an initial control target is set based on historical data statistical quantiles and dynamically corrected using real-time emission status. This avoids excessive ammonia injection while ensuring environmental compliance, achieving a dynamic balance between economy and compliance. Finally, closed-loop control is performed based on high-precision posterior estimates, effectively overcoming measurement lag and improving system response speed and control stability. This enables precise closed-loop control of the denitrification system and low-cost emission compliance. This paper addresses the challenge that existing technologies often rely on single data-driven models to simultaneously account for temporal dynamics and multi-feature coupling. SCR denitrification systems exhibit strong nonlinearity and large time lag. Using only time-series models like LSTM can easily overlook the combined impact of multiple features under current operating conditions, while using only regression models like XGBoost struggles to capture the long-term dependencies and periodic fluctuations of time series data. This leads to decreased prediction accuracy under complex and variable operating conditions, consequently affecting the timeliness of the control system. Furthermore, traditional Kalman filtering algorithms make overly idealistic assumptions about noise statistics. SCR systems often experience fluctuating operating conditions, and noise characteristics are time-varying. Fixed parameters can cause slow filter convergence or even divergence, making it difficult to provide reliable state estimates for the control system and hindering the achievement of precise closed-loop control of the denitrification system. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a schematic flowchart of the first embodiment of the SCR denitrification system control method of the present invention.

[0019] Figure 2 This is a schematic diagram of the hardware structure of the SCR denitrification system control device involved in the embodiments of this application.

[0020] Figure 3 Design a table for the parameters of the TimesNet model.

[0021] Figure 4The composition and symbol explanation of the working condition feature vector uk. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0023] Firstly, this application provides a control method for an SCR denitrification system.

[0024] In one embodiment, reference is made to Figure 1 , Figure 1 This is a schematic flowchart of the first embodiment of the SCR denitrification system control method of this application. Figure 1 As shown, the control methods for the SCR denitrification system include: S1: Based on a pre-trained time-series prediction model and historical operating data, obtain the prior prediction value of the outlet nitrogen oxide concentration.

[0025] S2: Based on the pre-trained operating condition regression model and the current operating condition characteristics, obtain the observed estimate of the outlet nitrogen oxide concentration.

[0026] S3: Based on an adaptive algorithm, noise statistical parameters are updated online. The noise statistical parameters are used to weight and fuse the prior predicted values ​​and the observed estimated values ​​to obtain the posterior estimated value of the outlet nitrogen oxide concentration.

[0027] S4: Set an initial control target based on historical outlet nitrogen oxide concentration data, and dynamically correct the initial control target according to the real-time emission status to obtain a dynamic target value.

[0028] S5: Calculate the deviation between the posterior estimate and the dynamic target value, and adjust the opening of the ammonia injection regulating valve according to the deviation.

[0029] By constructing a dual-model-driven prediction framework, a time-series prediction model is used to capture historical evolution patterns and obtain prior predictions. A working condition regression model is then combined to mine multi-feature nonlinear mappings and obtain observation estimates. This effectively overcomes the shortcomings of a single model in simultaneously addressing the dynamic nature of time series and the coupling of multiple features, significantly improving prediction accuracy under complex and varying working conditions. Furthermore, an adaptive algorithm is introduced to update noise statistical parameters online, weighting and fusing prior predictions and observation estimates. This solves the problem of slow convergence or divergence caused by fixed noise parameters in traditional Kalman filtering, enhancing the robustness and adaptability of state estimation. Simultaneously, an initial control target is set based on historical data statistical quantiles and dynamically corrected using real-time emission status. This avoids excessive ammonia injection while ensuring environmental compliance, achieving a dynamic balance between economy and compliance. Finally, closed-loop control is performed based on high-precision posterior estimates, effectively overcoming measurement lag and improving system response speed and control stability. This enables precise closed-loop control of the denitrification system and low-cost emission compliance.

[0030] In this example, when adjusting the opening of the ammonia injection regulating valve: a three-stage PID ammonia injection control is adopted. The first-stage PID main regulating loop takes the deviation between the dynamic target value and the posterior estimated value as input and outputs the theoretical ammonia flow rate setpoint. The second-stage PID secondary regulating loop takes the deviation between the theoretical ammonia flow rate setpoint and the actual ammonia injection flow rate as input and outputs the ammonia injection regulating valve opening command. The third-stage feedforward compensation loop takes the inlet nitrogen oxide concentration, flue gas flow rate, and unit load as feedforward inputs to correct the ammonia injection regulating valve opening command.

[0031] In this example, before obtaining the prior predicted value of the outlet nitrogen oxide concentration based on the pre-trained time series prediction model and historical operating data, the following steps are also included: collecting key parameters such as ammonia injection rate, ammonia injection regulating valve opening, ammonia-air ratio, inlet nitrogen oxide concentration, flue gas temperature, flue gas flow rate, boiler load, and fuel nitrogen content; performing outlier removal, missing value filling, and standardization on the collected data; and performing time delay compensation on the outlet nitrogen oxide concentration data based on the transmission delay time to align it with the current operating data on the time axis.

[0032] In this process, the dynamic evolution of the nitrogen oxide concentration at the outlet of the SCR denitrification system is regarded as the system state, and the state variable is defined as: in, express The real value of the net flue gas nitrogen oxide concentration at any given time (one-dimensional state vector, focusing on outlet nitrogen oxide concentration prediction and control). Combined with a sliding time window design, the length of the sliding time window is set to... The historical sequence of nitrogen oxide concentration at the outlet is defined as follows:

[0033] The set of auxiliary variables is ,in, These represent key operating parameters such as ammonia injection rate, ammonia injection regulating valve opening, ammonia-air ratio, inlet nitrogen oxide concentration, flue gas temperature, flue gas flow rate, boiler load, and fuel nitrogen content.

[0034] Auxiliary variable history window:

[0035] The input feature matrix of the TimesNet model is historical NO. x splicing of sequence and auxiliary variable history windows ,in, The total number of input features consists of 1 NO x Historical sequence and It consists of 1 auxiliary variable, which is consistent with the dimension (T,n) of the model input layer.

[0036] Furthermore, in one embodiment, obtaining the prior predicted value of the outlet nitrogen oxide concentration based on the pre-trained time-series prediction model and historical operating data includes: Based on historical operational data, historical sequences of outlet nitrogen oxide concentrations and historical windows of auxiliary variables are extracted, a sliding time window is set, and an input feature matrix is ​​constructed. The input feature matrix is ​​transformed in the frequency domain using a time series prediction model to extract the dominant period, and the one-dimensional time series data is reconstructed into a two-dimensional matrix form based on the dominant period. By using multiple convolutional branches of different scales in the time series prediction model to extract features from the two-dimensional matrix, a multi-scale feature map is obtained. Multi-scale feature maps are fused, and the fused features are weighted using the temporal attention mechanism in the temporal prediction model to obtain prior prediction values.

[0037] In this embodiment, prior predictions obtained through a pre-trained time-series prediction model effectively capture the long-term dependence and periodic fluctuation characteristics of historical nitrogen oxide concentrations at outlets. Specifically, the dominant period is extracted through frequency domain transformation, and the one-dimensional time-series data is reconstructed into a two-dimensional matrix, transforming temporal variations into spatial variations and enhancing the model's ability to uncover evolution patterns within and between periods. Multiple convolutional branches of different scales are used to extract short-term, medium-term, and long-term features, achieving comprehensive capture of multi-granularity temporal features. A temporal attention mechanism is used to weight the fused features, highlighting the influence weight of key operating conditions. This approach overcomes the shortcomings of traditional time-series models in simultaneously considering multiple periodic components and local fluctuations, significantly improving the fitting accuracy of prior predictions to the dynamic evolution of concentration time-series under complex varying operating conditions, and providing highly reliable state equation prior information for subsequent Kalman filter fusion.

[0038] In this example, the time series prediction model is the TimesNet model.

[0039] Using the TimesNet model to analyze state variables Perform prior estimation and construct the state equation:

[0040] In the formula ,represent Exit NO x Prior prediction of concentration; It represents the mapping relationship of a time-series deep learning model, realizing a non-linear transformation from the input feature matrix to prior predictions; This represents process noise, characterizing interference factors not captured in time-series modeling (such as minor fluctuations in fuel characteristics, slight changes in catalyst activity, etc.), and follows a zero-mean Gaussian distribution. , The process noise covariance is initially obtained through statistical analysis of historical operating data.

[0041] Subsequent updates will use the Sage-Husa adaptive algorithm, whose rules are as follows:

[0042] In the formula, Forgetting factor (taken in this paper) ), For Kalman gain, For prior covariance, For posterior covariance, The noise covariance was measured at the previous time step.

[0043] The specific structure of the TimesNet model is set as follows: the input layer dimension is (60, n), where 60 is the sliding window length and n is the number of input features (including one historical NOx concentration sequence from the outlet and n-1 historical sequences of auxiliary variables). T=60 is set as the sliding window length, and n is the feature dimension (including one-dimensional NOx concentration sequence from the outlet and n-1 historical sequences of auxiliary variables). (1-dimensional auxiliary variable). The dominant period is extracted using Fourier transform, and the amplitude of each frequency component is calculated:

[0044] In the formula, For Fast Fourier Transform, For amplitude calculation, This is the average amplitude. (Using...) The top k significant periods were selected:

[0045] Two-dimensional reconstruction, based on the dominant period pt, reconstructs a one-dimensional sequence into a two-dimensional matrix, capturing features within or during the period:

[0046] Multi-scale feature extraction involves extracting features from each two-dimensional matrix using three convolutional branches at three different scales (short-term / medium-term / long-term).

[0047] in, The kernel size is used to fuse multi-scale features.

[0048] Temporal attention weighting applies time-step weighting to the fused features, highlighting key operational moments:

[0049] Furthermore, in one embodiment, obtaining the observed estimate of the outlet nitrogen oxide concentration based on the pre-trained operating condition regression model and the current operating condition characteristics includes: Obtain various operating parameters at the current moment and construct a feature vector of the current operating condition; The current operating condition feature vector is input into a pre-trained operating condition regression model. The operating condition regression model is used to explore the nonlinear mapping relationship between the current operating condition feature vector and the outlet nitrogen oxide concentration, and the observed estimated value is output.

[0050] In this embodiment, by obtaining observation estimates based on a pre-trained operating condition regression model, the complex nonlinear mapping relationship between current multi-dimensional operating condition features and outlet nitrogen oxide concentration can be effectively explored. Specifically, by constructing a current operating condition feature vector containing multiple operating parameters, the real-time coupled influence of key factors such as boiler load, flue gas temperature, and ammonia injection rate on denitrification effect is comprehensively reflected, overcoming the defect of traditional time series models that easily ignore the comprehensive influence of multiple features of the current operating condition. The observation estimates provided by this method serve as inputs to the observation equation of the Kalman filter, complementing the prior predictions and enhancing the sensitivity of state estimation to sudden changes in operating conditions. This provides a high-confidence measurement basis for subsequent weighted fusion, further improving the robustness of the overall prediction system.

[0051] In this example, the working condition regression model is the XGBoost model.

[0052] Define the observed value as the outlet NO under the current operating condition. x The observed estimate of the concentration is denoted as: ,in The output NO reflects the characteristics of the current measurable operating conditions. xThe actual concentration level is used to complement the TimesNet prior predictions and improve the accuracy of state estimation.

[0053] By combining multi-dimensional operating condition characteristics, a Kalman filter observation equation is constructed, and the multivariate regression advantage of the XGBoost model is utilized to achieve measurement estimation.

[0054] in, ,represent The feature vector of the current operating condition at any given time. The total feature dimension includes original operating condition features and time-series derived features. Its specific composition is defined using a matrix and symbolic representation; see details below. Figure 3 . , representing a multivariate nonlinear mapping function based on gradient boosting trees, realizes the mapping from the working condition feature vector to NO. x Mapping of concentration observation estimates; , representing measurement noise, characterizes unmodeled disturbances such as sensor detection errors and random disturbances in operating conditions, and follows a zero-mean Gaussian distribution. ,in, To measure the noise covariance, the initial value is statistically calibrated based on the detection accuracy of each sensor in the denitrification system. Subsequent values ​​are synchronized with the process noise covariance Q and updated online using the Sage-Husa adaptive algorithm. The update formula is:

[0055] In the formula, , This represents the deviation between prior predictions and measurement estimates.

[0056] Further, in one embodiment, the online updating of noise statistical parameters based on an adaptive algorithm, and the weighted fusion of prior predicted values ​​and observed estimated values ​​using the noise statistical parameters to obtain a posterior estimated value of the outlet nitrogen oxide concentration, includes: A Kalman filter framework is constructed, in which the prior predictions are used as the output of the state equation and the observation estimates are used as the output of the observation equation. Based on an adaptive algorithm, the process noise covariance matrix and measurement noise covariance matrix are updated online using historical residual sequences; The Kalman gain is calculated based on the updated process noise covariance matrix and measurement noise covariance matrix. The posterior estimate is obtained by weighting the deviation between the prior prediction and the observed estimate based on Kalman gain and correcting the prior prediction.

[0057] In this embodiment, a Kalman filter framework is constructed, with prior predictions and observation estimates serving as the outputs of the state equation and observation equation, respectively, achieving effective fusion of time-series dynamic information and multi-feature coupled information. In particular, by using an adaptive algorithm to update the process noise covariance matrix and measurement noise covariance matrix online using historical residual sequences, the shortcomings of traditional Kalman filters, which struggle to adapt to the time-varying characteristics of SCR denitrification systems due to preset fixed noise parameters, are overcome. This avoids the slow convergence or even divergence problems caused by overly idealized assumptions about noise statistical characteristics. By calculating the Kalman gain based on the updated noise parameters and weighting the deviation, the confidence levels of time-series prior information and operational observation information can be adaptively balanced, thereby obtaining high-precision and robust posterior estimates. This effectively smooths measurement noise interference and provides a reliable state feedback basis for subsequent control strategies.

[0058] Furthermore, in one embodiment, the online update of the process noise covariance matrix and the measurement noise covariance matrix using historical residual sequences includes: Calculate the residual between the prior prediction and the observed estimate at the current time. By introducing a forgetting factor, the statistical properties of the noise covariance matrix and residual at the previous time step are weighted and averaged to obtain the process noise covariance matrix and measurement noise covariance matrix at the current time step.

[0059] In this embodiment, by calculating the residual between the prior predicted value and the observed estimated value at the current moment, and introducing a forgetting factor to weight the statistical characteristics of the noise covariance matrix and the residual at the previous moment, online real-time updates of the process noise covariance matrix and the measurement noise covariance matrix are achieved. The introduction of the forgetting factor can effectively adjust the weight ratio between historical noise information and current residual information, ensuring smooth and stable parameter updates while ensuring rapid response to sudden changes in system operating conditions and time-varying noise characteristics. This adaptive update mechanism overcomes the shortcomings of traditional fixed noise parameters in adapting to the complex operating environment of SCR denitrification systems, avoids the risk of filter divergence due to model mismatch, provides a reliable basis for the accurate calculation of Kalman gain, and further improves the accuracy and robustness of the posterior estimate.

[0060] like Figure 4 As shown, in this example, by combining the advantages of TimesNet time series modeling and XGBoost multi-feature regression, a Kalman filter-like fusion framework is constructed, in which the state equation and observation equation are provided by the two models respectively.

[0061] Equations of state:

[0062] Observation equation:

[0063] The TimesNet prior output and XGBoost observation estimation are incorporated into a Kalman-like filtering iterative process. By calculating the residual bias and solving the Kalman gain, the confidence levels of time-series prior information and operating condition observation information are adaptively balanced. The prior predictions are then corrected to obtain the fused posterior estimation result. Simultaneously, a Sage-Husa adaptive update mechanism is introduced to correct the process noise covariance Q and measurement noise covariance R online in real time, continuously updating the prior and posterior error covariances to achieve NO under time-varying operating conditions. x High-precision and robust concentration prediction.

[0064] Furthermore, in one embodiment, the step of setting an initial control target based on historical outlet nitrogen oxide concentration data and dynamically correcting the initial control target according to real-time emission status to obtain a dynamic target value includes: Historical outlet nitrogen oxide concentration data during the initial stage of system operation were selected as the sample set, and the statistical quantile values ​​of the sample set were calculated as the initial control target. Use the initial control target or the dynamic target value updated in the previous time step as the current target value; The dynamic target value is updated based on the relationship between the export nitrogen oxide concentration and the environmental protection limit and the current target value.

[0065] In this embodiment, the initial control target is set based on the statistical quantile values ​​of historical outlet nitrogen oxide concentration data. This objectively reflects the baseline level under stable system operation, avoiding the subjectivity and blindness of manually setting or fixing limit ratios. Furthermore, the initial control target or the dynamically updated target value from the previous moment is used as the current target value, and dynamic adjustments are made based on the relationship between real-time emission status, environmental limits, and the current target value, achieving adaptive iterative adjustment of the control target. When there is a risk of exceeding standards, the target value is automatically lowered to ensure environmental compliance; when operation is stable and the concentration is low, the target value is automatically raised to reduce ammonia injection. This dynamic target setting mechanism, while ensuring stable compliance of outlet nitrogen oxide concentrations, effectively avoids ammonia waste and secondary pollution caused by excessive ammonia injection, and solves the problem of insufficient economy or safety due to the inability of traditional fixed target values ​​to adapt to changes in operating conditions. It achieves the optimal balance between the operating cost and environmental risk of the denitrification system.

[0066] Furthermore, in one embodiment, updating the dynamic target value based on the relationship between the export nitrogen oxide concentration and the environmental limit and the current target value includes: If the concentration of nitrogen oxides in the export exceeds the environmental protection limit for a continuous period of time, the current target value will be reduced by the first preset ratio to obtain the updated dynamic target value. If the concentration of nitrogen oxides in the export is continuously lower than the current target value and does not exceed the environmental protection limit for a second preset time, the current target value will be adjusted upward according to the second preset ratio to obtain the updated dynamic target value. If the concentration of nitrogen oxides in exports is within the preset fluctuation range of the current target value and does not exceed the environmental protection limit, or does not meet the conditions for downward or upward adjustment, then the current target value will be maintained.

[0067] In this embodiment, refined dynamic correction of the control target is achieved through specific rules for downward adjustment, upward adjustment, and maintenance. Using a first preset duration of continuous exceedance of environmental limits as the downward adjustment trigger condition ensures timely reduction of the control target under persistent risk of exceeding standards, prioritizing environmental compliance. Using a second preset duration of continuous below-the-current target value without exceeding the limit as the upward adjustment trigger condition effectively identifies and utilizes the system's safety margin, reducing unnecessary ammonia injection and lowering ammonia consumption costs. Simultaneously, by maintaining the current target value within a preset fluctuation range or when adjustment conditions are not met, and combining this with the judgment logic of continuous reach of preset durations, interference from instantaneous measurement noise or brief operational disturbances is effectively filtered out, avoiding frequent jumps and oscillations in the control target value, ensuring the stability of the control system, and ultimately achieving the optimal balance between the denitrification system's operational economy and safety while ensuring compliance with emission standards.

[0068] In this example, the initial internal control target (first tertiary) is defined as one consecutive month during the initial operation of the SCR denitrification system (denoted as the time interval). ) exports NO x The historical concentration data is used as a sample set, denoted as:

[0069] in, NO for exports within one month x Total number of concentration samples (calculated at a 1-minute sampling interval, N0=43200). For the i-th sample, the export NO x True concentration (unit: mg / m³). For the sample set The first tertiary (Q1, 25th percentile) is defined as the initial internal control target value, arranged in ascending order, and denoted as . Its mathematical definition is: Select the nth element from the ordered sample sequence. The value corresponding to the position, that is:

[0070] In the formula, To ensure that the selected target value is the actual concentration value in actual operation, the rounding operation is performed downwards.

[0071] The initial target value It can objectively reflect the output NO under the stable operating state of the system. x The concentration baseline level avoids setting the target value too high (higher than) Excessive (leading to NO export) x This avoids the problem of concentrations easily exceeding standards and causing environmental risks, and also prevents the target value from being too low (below the limit). Excessive ammonia injection leads to problems such as excessive ammonia volume and increased ammonia consumption, providing a reasonable initial baseline for subsequent control strategies.

[0072] However, to adapt to the dynamic changes in the operating conditions of the SCR denitrification system (such as boiler load fluctuations, fuel characteristic changes, flue gas parameter disturbances, etc.) and environmental emission requirements, the design is based on the outlet NO x An adaptive correction rule for concentration operation status adjusts the internal control target value in real time, and the corrected target value at time k is recorded as... The specific correction logic, combined with textual explanations and mathematical expressions, is as follows: Rules for Lowering Standards for Exceeding Limits: Defining environmental limits as follows (Unit: mg / m³), if the exported NO x The concentration was maintained for 3 consecutive hours (denoted as the time interval [k]). 3,k]) all satisfy (i.e., exceeding the standard) (This refers to any time within this interval), indicating that the current control target is too high and poses an environmental risk. In this case, the target value will be lowered by 5%. The 10% reduction formula is as follows:

[0073] in, To lower the coefficient, The greater the deviation ( The larger ( The closer the value is to 0.10, the more this adjustment will continue until the export NO... x Concentration restored to the following.

[0074] Low concentration upward adjustment rule: If the exported NO x The concentration was maintained for 24 consecutive hours (referred to as the time interval). All of them satisfy And there was no time exceeding the standard within that interval (i.e.) This indicates that the current control target is too low, the ammonia injection rate is excessive, and the ammonia consumption is too high. In this case, the target value should be increased by 3%. 5%, the upward adjustment formula is:

[0075] in, To adjust the coefficient upwards, This ensures that, while complying with environmental regulations, unnecessary ammonia injection is reduced, thus lowering ammonia consumption.

[0076] Stability maintenance rule: If exports NO x The concentration remains within a stable range, meaning that the condition is met at any time t. Since there are no instances of exceeding the standard, it indicates that the current target value is suitable for the current working conditions and no adjustment is needed; therefore, the target value should remain unchanged. .

[0077] Secondly, this application also provides a control device for an SCR denitrification system, comprising: The prior prediction value acquisition module is used to obtain the prior prediction value of the outlet nitrogen oxide concentration based on the pre-trained time series prediction model and historical operating data. The observation estimate acquisition module is used to obtain the observation estimate of the outlet nitrogen oxide concentration based on the pre-trained working condition regression model and the current working condition characteristics. The posterior estimate acquisition module is used to update noise statistical parameters online based on an adaptive algorithm, and to use the noise statistical parameters to perform weighted fusion of prior predictions and observation estimates to obtain the posterior estimate of the outlet nitrogen oxide concentration. The dynamic target value acquisition module is used to set an initial control target based on historical outlet nitrogen oxide concentration data, and dynamically correct the initial control target according to the real-time emission status to obtain the dynamic target value. The calculation and execution module is used to calculate the deviation between the posterior estimate and the dynamic target value, and adjust the opening of the ammonia injection regulating valve according to the deviation.

[0078] The functions of each module in the SCR denitrification system control device correspond to the steps in the above-mentioned SCR denitrification system control method embodiment, and their functions and implementation processes will not be described in detail here.

[0079] The prior prediction acquisition module, the observation estimate acquisition module, the posterior estimate acquisition module, the dynamic target value acquisition module, and the calculation execution module can be devices with data processing capabilities, such as personal computers (PCs), laptops, and servers.

[0080] like Figure 2 As shown, in a third aspect, this application also provides an SCR denitrification system control device, which includes a processor, a memory, and an SCR denitrification system control program stored in the memory and executable by the processor, wherein when the SCR denitrification system control program is executed by the processor, it implements the steps of the above-mentioned SCR denitrification system control method.

[0081] In this embodiment of the application, an SCR denitrification system control device may further include a communication interface and a communication bus.

[0082] The communication bus can be of any type and is used to interconnect the processor, memory, and communication interface.

[0083] Communication interfaces include input / output (I / O) interfaces, physical interfaces, and logical interfaces used for interconnecting devices within the SCR denitrification system control unit, as well as interfaces used for interconnecting the SCR denitrification system control unit with other devices (such as other computing devices or user equipment). Physical interfaces can be Ethernet interfaces, fiber optic interfaces, ATM interfaces, etc.; user equipment can be displays, keyboards, etc.

[0084] Memory can be various types of storage media, such as random access memory (RAM), read-only memory (ROM), non-volatile RAM (NVRAM), flash memory, optical storage, hard disk, programmable ROM (PROM), erasable PROM (EPROM), electrically erasable PROM (EEPROM), etc.

[0085] The processor can be a general-purpose processor, which can call the SCR denitrification system control program stored in the memory and execute the SCR denitrification system control method provided in the embodiments of this application. For example, the general-purpose processor can be a central processing unit (CPU). The method executed when the SCR denitrification system control program is called can be referred to the various embodiments of the SCR denitrification system control method of this application, and will not be repeated here.

[0086] Fourthly, embodiments of this application also provide a storage medium.

[0087] The storage medium stores an SCR denitrification system control program, wherein when the SCR denitrification system control program is executed by the processor, the steps of the above-described SCR denitrification system control method are implemented.

[0088] The method implemented when the SCR denitrification system control program is executed can be referred to in the various embodiments of the SCR denitrification system control method of this application, and will not be repeated here.

[0089] It should be noted that the sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0090] The terms "comprising" and "having," and any variations thereof, in the specification, claims, and accompanying drawings of this application are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to such process, method, product, or apparatus. The terms "first," "second," and "third," etc., are used to distinguish different objects, etc., and do not indicate a sequence, nor do they limit "first," "second," and "third" to different types.

[0091] In the description of the embodiments of this application, terms such as "exemplary," "for example," or "for instance" are used to indicate examples, illustrations, or explanations. Any embodiment or design described as "exemplary," "for example," or "for instance" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of terms such as "exemplary," "for example," or "for instance" is intended to present the relevant concepts in a concrete manner.

[0092] In the description of the embodiments of this application, unless otherwise stated, " / " means "or". For example, A / B can mean A or B. The "and / or" in the text is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, A and / or B can mean: A exists alone, A and B exist simultaneously, and B exists alone. In addition, in the description of the embodiments of this application, "multiple" means two or more.

[0093] In some processes described in the embodiments of this application, multiple operations or steps are included in a specific order. However, it should be understood that these operations or steps may not be executed in the order they appear in the embodiments of this application, or they may be executed in parallel. The sequence number of the operation is only used to distinguish different operations, and the sequence number itself does not represent any execution order. In addition, these processes may include more or fewer operations, and these operations or steps may be executed sequentially or in parallel, and these operations or steps may be combined.

[0094] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes several instructions to cause a terminal device to execute the methods described in the various embodiments of this application.

[0095] The above are merely preferred embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A control method for an SCR denitrification system, characterized in that, include: Based on a pre-trained time-series prediction model and historical operating data, a priori predicted value of the outlet nitrogen oxide concentration is obtained. Based on the pre-trained working condition regression model and the current working condition characteristics, the observed estimate of the outlet nitrogen oxide concentration is obtained; The noise statistics parameters are updated online based on an adaptive algorithm. The prior predicted values ​​and observed estimates are then weighted and fused using the noise statistics parameters to obtain the posterior estimate of the outlet nitrogen oxide concentration. An initial control target is set based on historical outlet nitrogen oxide concentration data, and the initial control target is dynamically corrected according to the real-time emission status to obtain a dynamic target value. Calculate the deviation between the posterior estimate and the dynamic target value, and adjust the opening of the ammonia injection regulating valve according to the deviation.

2. The SCR denitrification system control method as described in claim 1, characterized in that, The aforementioned method for obtaining prior predictions of outlet nitrogen oxide concentrations based on a pre-trained time-series prediction model and historical operational data includes: Based on historical operational data, historical sequences of outlet nitrogen oxide concentrations and historical windows of auxiliary variables are extracted, a sliding time window is set, and an input feature matrix is ​​constructed. The input feature matrix is ​​transformed in the frequency domain using a time series prediction model to extract the dominant period, and the one-dimensional time series data is reconstructed into a two-dimensional matrix form based on the dominant period. By using multiple convolutional branches of different scales in the time series prediction model to extract features from the two-dimensional matrix, a multi-scale feature map is obtained. Multi-scale feature maps are fused, and the fused features are weighted using the temporal attention mechanism in the temporal prediction model to obtain prior prediction values.

3. The SCR denitrification system control method as described in claim 1, characterized in that, The method of obtaining the observed estimate of the outlet nitrogen oxide concentration based on the pre-trained operating condition regression model and the current operating condition characteristics includes: Obtain various operating parameters at the current moment and construct a feature vector of the current operating condition; The current operating condition feature vector is input into a pre-trained operating condition regression model. The operating condition regression model is used to explore the nonlinear mapping relationship between the current operating condition feature vector and the outlet nitrogen oxide concentration, and the observed estimated value is output.

4. The SCR denitrification system control method as described in claim 1, characterized in that, The method of online updating of noise statistical parameters based on an adaptive algorithm, which uses the noise statistical parameters to weight and fuse prior predicted values ​​and observed estimates to obtain a posterior estimate of the outlet nitrogen oxide concentration, includes: A Kalman filter framework is constructed, in which the prior predictions are used as the output of the state equation and the observation estimates are used as the output of the observation equation. Based on an adaptive algorithm, the process noise covariance matrix and measurement noise covariance matrix are updated online using historical residual sequences; The Kalman gain is calculated based on the updated process noise covariance matrix and measurement noise covariance matrix. The posterior estimate is obtained by weighting the deviation between the prior prediction and the observed estimate based on Kalman gain and correcting the prior prediction.

5. The SCR denitrification system control method as described in claim 4, characterized in that, The method of online updating the process noise covariance matrix and the measurement noise covariance matrix using historical residual sequences includes: Calculate the residual between the prior prediction and the observed estimate at the current time. By introducing a forgetting factor, the statistical properties of the noise covariance matrix and residual at the previous time step are weighted and averaged to obtain the process noise covariance matrix and measurement noise covariance matrix at the current time step.

6. The SCR denitrification system control method as described in claim 1, characterized in that, The method of setting an initial control target based on historical outlet nitrogen oxide concentration data and dynamically correcting the initial control target according to real-time emission status to obtain a dynamic target value includes: Historical outlet nitrogen oxide concentration data during the initial stage of system operation were selected as the sample set, and the statistical quantile values ​​of the sample set were calculated as the initial control target. Use the initial control target or the dynamic target value updated in the previous time step as the current target value; The dynamic target value is updated based on the relationship between the export nitrogen oxide concentration and the environmental protection limit and the current target value.

7. The SCR denitrification system control method as described in claim 6, characterized in that, The aforementioned updating of dynamic target values ​​based on the relationship between export nitrogen oxide concentrations, environmental limits, and current target values ​​includes: If the concentration of nitrogen oxides in the export exceeds the environmental protection limit for a continuous period of time, the current target value will be reduced by the first preset ratio to obtain the updated dynamic target value. If the concentration of nitrogen oxides in the export is continuously lower than the current target value and does not exceed the environmental protection limit for a second preset time, the current target value will be adjusted upward according to the second preset ratio to obtain the updated dynamic target value. If the concentration of nitrogen oxides in exports is within the preset fluctuation range of the current target value and does not exceed the environmental protection limit, or does not meet the conditions for downward or upward adjustment, then the current target value will be maintained.

8. A control device for an SCR denitrification system, characterized in that, include: The prior prediction value acquisition module is used to obtain the prior prediction value of the outlet nitrogen oxide concentration based on the pre-trained time series prediction model and historical operating data. The observation estimate acquisition module is used to obtain the observation estimate of the outlet nitrogen oxide concentration based on the pre-trained working condition regression model and the current working condition characteristics. The posterior estimate acquisition module is used to update noise statistical parameters online based on an adaptive algorithm, and to use the noise statistical parameters to perform weighted fusion of prior predictions and observation estimates to obtain the posterior estimate of the outlet nitrogen oxide concentration. The dynamic target value acquisition module is used to set an initial control target based on historical outlet nitrogen oxide concentration data, and to dynamically correct the initial control target according to the real-time emission status to obtain the dynamic target value. The calculation and execution module is used to calculate the deviation between the posterior estimate and the dynamic target value, and adjust the opening of the ammonia injection regulating valve according to the deviation.

9. A control device for an SCR denitrification system, characterized in that, The SCR denitrification system control device includes a processor, a memory, and an SCR denitrification system control program stored in the memory and executable by the processor, wherein when the SCR denitrification system control program is executed by the processor, it implements the steps of the SCR denitrification system control method as described in any one of claims 1 to 7.

10. A storage medium, characterized in that, The storage medium stores an SCR denitrification system control program, wherein when the SCR denitrification system control program is executed by a processor, it implements the steps of the SCR denitrification system control method as described in any one of claims 1 to 7.

Images