An intelligent ammonia injection optimization control method and system for SCR denitration based on working condition prediction
Patent Information
- Application Number
- CN202611177127.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-05
- Publication Date
- 2026-10-23
AI Technical Summary
[0010]随着双碳目标推进与火电机组深度调峰常态化，脱硝系统运行工况大幅拓宽，传统控制算法的局限性愈发突出：过量喷氨导致氨逃逸超标，引发空预器堵塞、催化剂失活等次生问题，大幅增加运维成本与停机风险；人工调节的不确定性也给稳定达标排放带来挑战
[0035]控制时效性显著提升：通过提前2分钟的NOx预测，从根源抵消系统传输时滞，变负荷工况下出口NOx波动幅度大幅收窄，超标风险显著降低。
Smart Images
Figure CN122955347A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent control technology for flue gas denitrification in thermal power plants, and particularly to an intelligent ammonia injection optimization control method and system for SCR denitrification based on operating condition prediction. Through innovative software architecture combining machine learning algorithms and multi-level control strategies, this invention upgrades traditional passive ammonia injection control based on outlet feedback to an active control mode based on combustion condition prediction, solving the core technical challenges of large time delays, strong coupling, and poor adaptability to operating conditions in denitrification systems from an algorithmic perspective. Background Technology
[0002] 1. Technical defects of existing control methods
[0003] Currently, the ammonia injection control of SCR denitrification systems in thermal power plants generally adopts traditional PID feedback control or simplified feedforward control schemes, which face the following technical bottlenecks at the algorithm and control logic levels that are difficult to overcome:
[0004] The large time lag leads to control lag: there is a physical delay of 2-5 minutes in the transmission of flue gas from the furnace combustion zone to the SCR reactor, and the feedback control based on the measured NOx value at the outlet always lags behind the actual operating conditions. During load changes, the unit experiences significant regulation oscillations, resulting in large fluctuations in NOx emissions and a high risk of exceeding emission standards.
[0005] The feedforward calculation is too simplified: existing feedforward schemes are mostly based on simplified stoichiometric formulas, which cannot accurately describe the nonlinear effects of multiple variables such as coal quality fluctuations, air-coal ratio distribution, and air distribution adjustment on NOx generation. The feedforward accuracy is insufficient and it is difficult to effectively offset the time lag.
[0006] Poor adaptability of fixed parameters: Traditional controller parameters are tuned based on specific operating conditions. When the load drifts or the coal type is switched, the control performance drops significantly, requiring manual periodic retuning of parameters. It cannot adapt to all operating conditions.
[0007] The lack of decoupling mechanism in zone control: existing zone ammonia injection mostly uses independent PID loops, which do not consider the coupling effect of lateral diffusion of flue gas in adjacent zones. Frequent adjustments are prone to causing system oscillations, and total quantity control and zone balance are difficult to coordinate.
[0008] Operation relies on manual intervention: Many units still rely on manual adjustments by operators. The operation is intensive and inconsistent when the load fluctuates. Adjustment deviations are prone to occur at night and during shift changes, and excessive ammonia injection is common.
[0009] 2. Industry Pain Points
[0010] With the advancement of dual-carbon targets and the normalization of deep peak shaving in thermal power units, the operating conditions of denitrification systems have been greatly expanded, and the limitations of traditional control algorithms have become increasingly prominent: excessive ammonia injection leads to ammonia escape exceeding standards, causing secondary problems such as air preheater blockage and catalyst deactivation, significantly increasing operation and maintenance costs and downtime risks; the uncertainty of manual adjustment also poses challenges to achieving stable emission standards. The industry urgently needs to fundamentally upgrade its control algorithms and decision-making logic to achieve accurate, stable, and economical operation under all operating conditions. Summary of the Invention
[0011] 1. Technical issues
[0012] To address the five major algorithmic bottlenecks in existing control methods—difficulty in canceling time delays, difficulty in modeling multivariate coupling, difficulty in adapting fixed parameters, susceptibility to oscillations in zoned regulation, and high dependence on manual intervention—this invention achieves the following technical objectives through software architecture and algorithm innovation:
[0013] Predicting NOx concentration changes in advance can offset system transport delays at the source, transforming passive feedback into active control.
[0014] Establish a multivariate nonlinear mapping model to accurately characterize the coupling relationship between combustion parameters and NOx generation;
[0015] A three-level collaborative control strategy is constructed, taking into account dynamic response speed, steady-state control accuracy, and uniformity of zonal distribution.
[0016] It enables automatic iterative updates of models and parameters, adaptive operation under all working conditions, and reduces the need for manual intervention;
[0017] It has built-in multiple security constraint mechanisms to ensure that the control process is safe, reliable, traceable, and auditable.
[0018] 2. Technical Solution
[0019] This invention constructs a complete technical solution from the perspectives of software algorithms and control logic, adopting a four-layer software architecture of "perception-prediction-decision-execution". The core innovation lies in the prediction engine and the three-level control strategy, with hardware serving only as the implementation carrier.
[0020] (1) Data Acquisition and Feature Processing Layer
[0021] By connecting to the power plant's DCS system through standard industrial protocols, real-time, full-dimensional operational data from the combustion side is collected. Preprocessing such as outlier detection, filtering and noise reduction, and time-series alignment are performed. Through multi-scale feature engineering, a standardized feature set covering 13 subsystems and more than 80 dimensions of physical and derived features is constructed to provide high-quality input for the prediction model.
[0022] (2) Multi-model fusion prediction layer
[0023] A dual-model integrated architecture of "structured feature learning + temporal deep learning" is adopted to predict the NOx concentration at the SCR inlet in advance by utilizing the physical time window of flue gas transport.
[0024] Structured feature learning model: Based on gradient boosting ensemble algorithm, it establishes a nonlinear mapping between combustion process features and NOx concentration, which is highly interpretable and supports feature contribution quantification and confidence interval output;
[0025] Temporal deep learning model: Based on gated recurrent networks and attention mechanisms, it automatically captures the coupling patterns of multiple parameters and adapts to complex dynamic working conditions;
[0026] Dynamic weighted fusion: The weights of the two models are dynamically allocated according to the operating condition range and the prediction confidence level, and the output is a stable and reliable early prediction result with a prediction lead of no less than 2 minutes.
[0027] (3) Three-level collaborative optimization control layer
[0028] Based on the prediction results, a three-level control logic is constructed: "feedforward-driven, feedback-corrected, and zone-balanced".
[0029] Feedforward predictive control: Using the predicted NOx concentration as input, the baseline ammonia injection rate is calculated in advance by combining the operating condition correction coefficient, and the system time lag is actively offset;
[0030] Feedback correction control: The measured value of NOx at the outlet is used as feedback, and the feedforward quantity is finely adjusted online through adaptive PID to ensure steady-state accuracy;
[0031] Zonal equilibrium optimization: The ammonia nitrogen ratio deviation is calculated based on the zonal monitoring data. The adjustment amount of each zonal zone is corrected by the decoupling compensation matrix. A slow-cycle fine-tuning strategy is adopted to avoid coupling oscillations and achieve uniform distribution of outlet concentration.
[0032] (4) Security constraints and closed-loop iteration layer
[0033] It incorporates four safety constraints: hard emission limits, ammonia slip limits, confidence thresholds, and mode switching. All control commands are verified before issuance. It supports seamless switching to traditional control modes. Simultaneously, a closed-loop iterative mechanism is established, periodically updating the model and control parameters based on newly added operating data. When switching coal types, it automatically loads the corresponding pre-trained model and performs rapid fine-tuning, achieving full-condition adaptive control.
[0034] 3. Beneficial effects
[0035] Significantly improved control timeliness: By predicting NOx 2 minutes in advance, the system transmission delay is offset at the source, the fluctuation range of NOx at the outlet is greatly narrowed under variable load conditions, and the risk of exceeding the standard is significantly reduced.
[0036] Dual improvement in control accuracy and economy: Multi-model fusion prediction ensures accuracy, and three-level collaborative control enables precise ammonia injection, effectively reducing excessive ammonia injection, reducing reducing agent consumption and ammonia escape, and alleviating blockage and corrosion of downstream equipment.
[0037] Strong adaptability under all working conditions: The automatic model iteration and working condition sub-model switching mechanism can adapt to complex scenarios such as deep peak shaving and coal type switching, without the need for frequent manual parameter tuning.
[0038] Operational safety and traceability: Multiple safety constraints protect operational boundaries, feature attribution and full log support decision traceability, balancing automation and controllability.
[0039] Deployment-friendly and low-risk: Supports progressive verification and deployment, and connects to existing DCS systems through standard protocols without requiring modifications to existing control hardware, resulting in low transformation risk and cost. Attached Figure Description
[0040] Figure 1 The overall software architecture diagram of the intelligent ammonia injection control system (showing the data flow and logical relationships of the four software modules: data acquisition, prediction engine, three-level control, and safety iteration).
[0041] Figure 2 Flowchart of the multi-model fusion NOx prediction algorithm (showing the complete algorithm flow of feature input, dual-model operation, dynamic weighted fusion, and confidence output)
[0042] Figure 3 : Three-level coordinated ammonia injection control logic block diagram (showing the control link and command output logic of feedforward calculation, feedback correction, and zone balancing)
[0043] Figure 4 : Flowchart of security constraint verification and closed-loop iteration (showing the closed-loop logic of the four-fold security verification mechanism and the periodic iteration of model parameters)
[0044] Figure 5 NOx prediction performance comparison curve (showing the fitting effect between predicted values and actual values 2 minutes in advance) Detailed Implementation
[0045] 1. Data Acquisition and Feature Engineering Implementation
[0046] The system connects to the power plant's DCS real-time database through standard industrial protocols such as OPC-UA / Modbus, collecting operational measurement points covering 13 subsystems including coal feeding system, air distribution system, SOFA system, burner tilt angle, flue gas temperature, and unit load, with no fewer than 200 original measurement points.
[0047] In the data preprocessing stage, an improved DBSCAN algorithm is used for outlier detection, combined with Kalman filtering to denoise high-frequency data, and unified clock alignment is performed. Through feature engineering, process features such as load change rate, total air volume / coal feed ratio, air distribution ratio of each layer, and flue gas temperature gradient are derived, ultimately forming a standardized feature set of no less than 80 dimensions, which serves as the input to the prediction model.
[0048] 2. Implementation of Multi-Model Fusion Prediction Engine
[0049] The prediction engine adopts a dual-model integrated architecture, developed based on the TensorFlow framework, and deployed in a lightweight manner on an industrial control all-in-one machine.
[0050] Structured Feature Learning Model: Employing a gradient boosting tree ensemble algorithm, this model performs nonlinear fitting on process-derived features, outputting NOx predictions while simultaneously calculating the contribution of each input feature. It supports quantile regression to output a 95% confidence interval. This model offers strong interpretability and high prediction accuracy under steady-state conditions.
[0051] Temporal deep learning model: It adopts a GRU gated recurrent network combined with a multi-head attention mechanism to learn the original temporal features end-to-end, automatically capture the time delay relationship and dynamic coupling law between multiple parameters, and performs better under dynamic conditions such as variable load.
[0052] Dynamic fusion mechanism: The system determines the current operating condition range based on indicators such as real-time load and load change rate, and dynamically allocates output weights by combining the historical prediction errors and confidence levels of the two models. Offline verification shows that the average absolute error of the fused prediction is ≤8mg / m³, the relative error is about 3.1%, and the prediction lead is stable at 2 minutes, which can completely cover the response time of the ammonia injection system.
[0053] 3. Implementation of the three-level collaborative control strategy
[0054] The control layer is implemented using software modularization, and outputs standard control commands to interface with the DCS system.
[0055] Feedforward predictive control module: Receives NOx prediction values, combines the baseline ammonia-nitrogen molar ratio with the operating condition correction coefficients (load, coal quality, flue gas temperature) to calculate the baseline ammonia injection rate, and completes the pre-adjustment of the ammonia injection rate before the NOx concentration change reaches the reactor, eliminating the time lag effect from the root cause.
[0056] Feedback correction module: Taking the deviation between the measured and setpoint NOx values at the outlet as input, it uses an adaptive PID algorithm to calculate the correction amount and fine-tune the feedforward reference value. The PID parameters are automatically adjusted according to operating conditions to ensure steady-state control accuracy under different loads.
[0057] The zone-level balancing optimization module receives NOx and ammonia slip monitoring data from each zone and calculates the ammonia-nitrogen molar ratio deviation for each zone. It incorporates a built-in adjacent zone coupling compensation matrix, corrects the adjustment amount for each zone based on the lateral diffusion influence coefficient, and employs a slow-cycle fine-tuning strategy to avoid coupling oscillations caused by frequent adjustments in adjacent zones, thereby achieving uniform outlet concentration distribution.
[0058] 4. Safety constraints and closed-loop iterative implementation
[0059] The system has a built-in four-fold safety constraint verification logic, and all control commands must pass the verification sequentially before being issued:
[0060] Emission hard limit verification: Predict that the hourly average NOx value at the export outlet will not exceed the standard;
[0061] Ammonia slip limit verification: The amount of ammonia injected does not exceed the upper limit value corresponding to the ammonia slip threshold;
[0062] Confidence verification: When the prediction confidence is lower than the threshold, the feedforward weight is automatically reduced and the system reverts to a conservative mode.
[0063] Instruction amplitude verification: The change in instruction value in a single cycle is limited to a safe range to prevent large disturbances.
[0064] The system supports three levels of seamless switching: manual, original PID, and intelligent control. During the switching process, the commands are processed with gradual amplitude limiting, resulting in no control jumps. It also supports three levels of progressive online modes: bypass monitoring, suggested output, and automatic control, ensuring a zero-risk verification process.
[0065] Closed-loop iteration mechanism: The system automatically performs incremental training on the prediction model based on newly added operating data according to shift, and fine-tunes the model parameters; when a coal type switch or long-term operating condition drift is detected, the system automatically matches the pre-trained sub-model of the corresponding operating condition, and the model fine-tuning time is ≤2 hours, always maintaining the optimal control performance.
[0066] 5. Explainability and Implementation of Operational and Maintenance Functions
[0067] The system has a built-in feature attribution analysis module that can quantify and display the main driving factors and their contribution percentages for each prediction and control decision, helping operators understand system behavior and assisting in combustion optimization and adjustments. All control commands, model outputs, and operational data are fully logged, supporting historical backtracking and audit traceability.
[0068] This invention significantly improves the control quality and operational economy of the denitrification system through pure software-level algorithm innovation and control logic reconstruction without changing the original hardware architecture, providing a low-risk, high-return software solution for the intelligent transformation of denitrification in thermal power plants.
Claims
1. A method for intelligent ammonia injection optimization control of SCR denitrification based on operating condition prediction, characterized in that, The proactive adaptive regulation of the ammonia injection process is achieved through software control logic of "sensing-prediction-decision-execution," including the following steps: (1) Multi-source data acquisition and feature engineering steps: Real-time acquisition of DCS operation data on the combustion side, data cleaning, time alignment and multi-scale feature derivation processing are performed to construct a combustion feature set covering multiple subsystems such as coal feeding, air distribution, flue gas temperature and load; (2) NOx concentration prediction steps of multi-model fusion: The integrated architecture of structured feature learning model and time-series deep learning model is adopted to calculate the SCR inlet NOx concentration after the preset advance time based on the current combustion feature sequence, and output the predicted value and confidence index simultaneously. (3) Three-level coordinated ammonia injection optimization control steps: The baseline ammonia injection amount is calculated with the predicted NOx concentration as the feedforward input, and closed-loop feedback correction is performed in combination with the measured value of NOx at the outlet. At the same time, the ammonia-nitrogen ratio balance adjustment of the zone is performed based on the zone monitoring data. After multi-objective solution, the final ammonia injection valve control command is generated. (4) Safety constraint verification and closed-loop iteration steps: Perform multiple safety boundary verifications on the control commands, and send them to the execution unit after the verification is passed; collect control effect data in real time, and automatically iterate and update the model parameters and control strategies according to the cycle to form a closed-loop optimization of the whole process.
2. The intelligent ammonia injection optimization control method for SCR denitrification based on operating condition prediction as described in claim 1, characterized in that, The multi-model fusion prediction in step (2) specifically includes: Structured feature learning model: Combustion process-derived indicators are constructed through time-series feature engineering, and a gradient boosting ensemble learning algorithm is used to establish a nonlinear mapping from combustion parameters to NOx concentration, outputting interpretable feature contribution and quantile confidence intervals; Temporal deep learning model: It adopts a gated recurrent network combined with a multi-head attention mechanism to learn the dynamic evolution law of working conditions from multivariate time series data and automatically capture the time delay coupling relationship between multiple parameters; Model fusion strategy: The output weights of the two models are dynamically allocated according to the working condition range and confidence level. After fusion, the final prediction result is obtained, and the prediction lead time is not less than 2 minutes.
3. The intelligent ammonia injection optimization control method for SCR denitrification based on operating condition prediction according to claim 1, characterized in that, The three-level synergistic ammonia injection optimization control in step (3) specifically includes: Feedforward predictive control: Based on the predicted trend of NOx concentration change at the SCR inlet, combined with stoichiometry and operating condition correction coefficient, the optimal amount of ammonia to be injected is calculated in advance to offset the regulation lag caused by system transmission time delay. Feedback correction control: The measured value of the outlet NOx concentration is used as the feedback value, and the feedforward reference ammonia injection quantity is finely adjusted online through an adaptive PID algorithm to ensure steady-state control accuracy; Zonal equalization optimization: Based on the NOx and ammonia slip monitoring data of each zone, the deviation of the ammonia nitrogen molar ratio of each zone is calculated, and the ammonia injection adjustment amount of each zone is corrected by the decoupling compensation matrix to achieve uniformity of outlet concentration distribution.
4. The intelligent ammonia injection optimization control method for SCR denitrification based on operating condition prediction according to claim 1, characterized in that, The feature engineering in step (1) specifically includes: extracting no less than 80 physical features from the DCS system, covering 13 combustion-related subsystems; performing outlier removal, Kalman filtering for noise reduction and time-series alignment on the original data to derive process features such as load change rate, air-coal ratio, and air distribution ratio, which constitute the model input feature set.
5. The intelligent ammonia injection optimization control method for SCR denitrification based on operating condition prediction according to claim 1, characterized in that, The multiple security constraint verification in step (4) includes: Hard emission limits: Control the hourly average NOx concentration at the export level to not exceed the set emission limit; Ammonia slip upper limit constraint: The upper limit of ammonia injection is limited by the ammonia slip concentration threshold; Confidence threshold constraint: When the prediction confidence is lower than the set threshold, it automatically reverts to conservative control mode; Mode switching constraints: Supports one-click switching to traditional control mode, and the switching process is performed with non-disruptive gradual change processing.
6. The intelligent ammonia injection optimization control method for SCR denitrification based on operating condition prediction according to claim 1, characterized in that, It also includes feature attribution and interpretability steps: a built-in feature attribution analysis module that quantifies the main driving factors and contribution ratios for each prediction and control decision, stores full logs of all decision-making processes, and supports post-event auditing, traceability, and operational optimization assistance.
7. The intelligent ammonia injection optimization control method for SCR denitrification based on operating condition prediction according to claim 1, characterized in that, The closed-loop iteration steps specifically include: incrementally updating the prediction model and control parameters based on newly added operating data according to a set period; when a coal type switch or operating condition drift is detected, automatically loading the pre-trained sub-model of the corresponding operating condition and performing rapid fine-tuning to achieve full-condition adaptive adaptation.
8. The intelligent ammonia injection optimization control method for SCR denitrification based on operating condition prediction according to claim 3, characterized in that, In the partition balance optimization, a slow-cycle fine-tuning strategy is adopted to perform partition adjustment. The lateral diffusion interference is offset by the coupling compensation matrix of adjacent partitions, thus avoiding system oscillation caused by frequent adjustments.
9. The intelligent ammonia injection optimization control method for SCR denitrification based on operating condition prediction according to claim 1, characterized in that, It supports a progressive verification operation mode, which is verified and put into operation step by step through bypass monitoring mode, suggested output mode and automatic control mode, without any forced intervention to the original control system.
10. A smart ammonia injection optimization control system for SCR denitrification based on operating condition prediction, characterized in that, include: Data acquisition and feature processing module: used to acquire DCS operation data in real time, perform data cleaning and time-series feature engineering, and output a standardized combustion feature set; Multi-model fusion prediction module: It has built-in structured feature learning unit and time-series deep learning unit, which are used to predict the NOx concentration and confidence level at the SCR inlet in advance based on combustion feature set; The three-level collaborative control module is used to perform feedforward prediction calculations, feedback correction and adjustment, and zonal balance optimization to solve and generate the optimal ammonia injection control command. Safety constraint and closed-loop iteration module: used to perform multiple safety checks on control commands, issue and execute them, and periodically iterate and update the model and control parameters; The system is integrated into an industrial control all-in-one machine in the form of software functional modules, and communicates bidirectionally with the power plant's DCS system through standard industrial protocols.