NOx emission concentration prediction system of circulating fluidized bed boiler

Abstract

The invention relates to a NOx emission concentration prediction system for a circulating fluidized bed boiler. The system comprises a data acquisition and dynamic feature construction module of a distributed sensor network, which realizes multi-dimensional characterization of a combustion state through space-time coupling analysis; according to the hybrid prediction model, a double-layer architecture is adopted to decouple a static combustion rule and a dynamic delay effect, and the prediction precision is improved by utilizing stable output of a base layer and self-adaptive compensation of a correction layer; the closed-loop control module is used for forming a synergistic effect of feedforward and feedback based on an intelligent decision-making mechanism of a working condition mapping model so as to avoid control overshoot; and the health guarantee module constructs a triple protection system, so that the data reliability is ensured, and redundancy protection can be realized under a fault working condition. According to the method, high-precision prediction and real-time control of NOx emission of the coal-fired unit in the full working condition range can be achieved, and the stability and the environmental protection performance of boiler operation are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of clean combustion and pollution control, and in particular to a NOx emission concentration prediction system for a circulating fluidized bed boiler. Background Art

[0002] Circulating fluidized bed boilers are a typical application of efficient and clean combustion technology, and their nitrogen oxide emission control is an important part of the environmental protection management of coal-fired power plants. In the existing technology, the mainstream control method mostly adopts feedforward control based on the prediction of operating parameters. However, due to the dynamic characteristics of the combustion process of circulating fluidized bed boilers, such as strong nonlinearity and large hysteresis, the existing predictive control system has the problem of insufficient coupling between combustion state perception and emission response. Specifically, traditional systems usually rely on static operating condition modeling. When the coal quality fluctuates or the unit load is adjusted rapidly, they fail to reflect the spatiotemporal delay characteristics of the air-coal ratio coupling process in real time, resulting in a systematic deviation between the actual emission concentration and the predicted value. This deviation further causes a phase difference between the control command and the process response, causing the actuators such as the secondary air damper to operate frequently, which in turn causes the combustion process to be unstable, resulting in a decrease in emission control effect and the risk of excessive emissions during the variable load transition process. Summary of the Invention

[0003] In view of this, the present invention provides a NOx emission concentration prediction system for a circulating fluidized bed boiler to solve the technical defects in the prior art.

[0004] Specifically, the present invention provides a circulating fluidized bed boiler NOx emission concentration prediction system, comprising: A distributed sensor network synchronously collects the target circulating fluidized bed boiler's operating parameters, fuel characteristics, and environmental variables. Operating parameters include dynamic data such as exhaust pressure and generator power acquired in real time through the DCS system. Fuel characteristics include coal particle size distribution obtained through coal sampling and screening tests or online particle size monitoring devices, as well as volatile matter content measured through coal sampling and testing. Environmental variables include flue gas humidity and inlet air temperature continuously recorded by temperature and humidity sensors. The dynamic feature construction module uses a sliding time window mechanism to calculate the dynamic change rate of operating parameters, and associates the instantaneous ratio of coal feed rate to primary air volume to generate air-coal coupling characteristics. Under variable load conditions, it generates time-lag compensation characteristics through phase difference analysis. A hybrid prediction model consists of a base layer and a correction layer. The base layer outputs NOx concentration predictions using the LC mean model. The correction layer activates the LSTM timing correction network to compensate for combustion delay effects when load fluctuations are detected. The closed-loop control module uses a preset operating condition mapping model to adjust the secondary air damper opening and bed temperature setpoints based on the deviation between the predicted NOx concentration and the emission limit; The health assurance module implements a triple protection mechanism of sensor abnormality diagnosis, prediction credibility assessment, and safety degradation control.

[0005] In some embodiments, the distributed sensing network further includes an embedded infrared spectrometer array for capturing the combustion flame morphology and temperature field distribution in the furnace in real time through multi-band optical sensors.

[0006] In some embodiments, the dynamic feature construction module uses the Pearson correlation coefficient to screen static strong correlation features under steady-state conditions.

[0007] In some embodiments, the hybrid prediction model has an online self-update mechanism. When it is detected that the continuous prediction error exceeds a preset threshold, an incremental learning process is triggered to update the model parameters, thereby improving the adaptability of the model under dynamic working conditions.

[0008] In some embodiments, the closed-loop control module dynamically adjusts the control target weight during the load change process to avoid system oscillation while ensuring that the emission limit is not exceeded.

[0009] In some embodiments, when the health assurance module detects that a sudden change in infrared spectrometer data exceeds a preset standard deviation threshold, it automatically switches to a backup data source to ensure the continuity and stability of the monitoring data.

[0010] In some embodiments, the LSTM timing correction network of the correction layer generates a time lag compensation factor by memorizing historical operating condition change patterns. The calculation formula for the time lag compensation factor includes:

[0011] in, is the time lag compensation factor, which is used to correct the phase delay between the predicted value and the actual value of NOx; is the length of the historical time window, which is determined by the sliding time window mechanism of the dynamic feature construction module; is the generator power derivative at the ti-th moment, which is the dynamic rate of change calculated by the generator power in the operating parameters; is the phase difference between the load change rate and the NOx generation rate; is the fuel-air mixing efficiency of the jth combustion zone, which is calculated based on the coal feed particle size distribution and volatile matter content in the fuel characteristic data; is the primary air volume deviation of the j-th combustion zone, which is calculated based on the difference between the instantaneous value of the primary air volume collected in real time by the distributed sensor network and the reference value; is the spatial variation rate of the fuel volatile content in the jth combustion area, which is calculated based on the volatile content monitored online by the near-infrared spectrometer; is the exponential decay coefficient, In order to control the nonlinear intensity of the influence of air volume change, is the suppression coefficient of the fuel gradient disturbance, where 、 and The initial setting is a preset value and is updated through LSTM timing correction network training.

[0012] In some embodiments, the formula for calculating the rate phase difference includes:

[0013] in: is the instantaneous value of the primary air volume at the kth sampling time, which is collected in real time by the distributed sensor network; is the phase angle of the wind-coal coupling characteristic, which is generated by the ratio correlation of the dynamic characteristic construction module; is the standard deviation of the furnace temperature field distribution, which is calculated from the multi-band data of the infrared spectrometer array; is the mean value of the temperature field, The sequence is obtained by taking the sliding average; is the flame morphology fluctuation rate, obtained by analyzing the time series data of the infrared spectrometer array; is the ambient humidity correction coefficient, which is obtained by normalizing the flue gas humidity recorded by the temperature and humidity sensor; M is the number of sampling points in the infrared spectrometer space.

[0014] In some embodiments, the operating condition mapping model calibrates the nonlinear mapping relationship between NOx concentration and operating parameters through experiments, and preferentially adjusts the secondary air valve opening.

[0015] In some embodiments, the safety degradation control module switches to a preset operating curve when the bed temperature or wind pressure exceeds a safety range, and the curve is generated based on similarity matching of historical safety operating conditions.

[0016] The beneficial effects of at least one embodiment of the present invention include: a data acquisition and dynamic feature construction module of a distributed sensor network, which realizes multi-dimensional characterization of the combustion state through spatiotemporal coupling analysis; a hybrid prediction model, which uses a two-layer architecture to decouple the static combustion law and the dynamic delay effect, and uses the stable output of the base layer and the adaptive compensation of the correction layer to improve the prediction accuracy; a closed-loop control module, which forms a synergistic effect of feedforward and feedback based on the intelligent decision-making mechanism of the operating condition mapping model to avoid control overshoot; a health protection module, which constructs a triple protection system to ensure data reliability and realize redundant protection under fault conditions. Through the above design, the present invention can achieve high-precision prediction and real-time control of NOx emissions of coal-fired units within the full operating range, significantly improving the stability and environmental protection performance of boiler operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 This is a structural block diagram of a circulating fluidized bed boiler NOx emission concentration prediction system provided by the present invention. DETAILED DESCRIPTION

[0018] The following description sets forth many specific details to facilitate a thorough understanding of this specification. However, this specification can be implemented in many other ways than those described herein, and those skilled in the art can make similar generalizations without violating the scope of this specification. Therefore, this specification is not limited to the specific implementations disclosed below.

[0019] The terms used in one or more embodiments of this specification are for the purpose of describing specific embodiments only and are not intended to limit one or more embodiments of this specification. The singular forms of "a" and "the" used in one or more embodiments of this specification and the appended claims are also intended to include plural forms, unless the context clearly indicates otherwise. It should also be understood that the term "and / or" used in one or more embodiments of this specification refers to and includes any or all possible combinations of one or more associated listed items. The modifications of "one" and "a plurality" mentioned in this disclosure are illustrative and not restrictive, and those skilled in the art should understand that unless the context clearly indicates otherwise, it should be understood as "one or more".

[0020] It should be understood that although the terms first, second, etc. may be used to describe various information in one or more embodiments of this specification, such information should not be limited to these terms. These terms are only used to distinguish the same type of information from each other. For example, without departing from the scope of one or more embodiments of this specification, the first may also be referred to as the second, and similarly, the second may also be referred to as the first. Depending on the context, the word "if" as used herein may be interpreted as "at the time of" or "when" or "in response to determining".

[0021] First, the terms involved in one or more embodiments of this specification are explained.

[0022] NOx: Nitrogen Oxides.

[0023] LC: LightGBM-CatBoost Ensemble Model, LightGBM-CatBoost integrated mean model.

[0024] LSTM timing correction network: Long Short-Term Memory Timing Correction Network, long short-term memory timing correction network.

[0025] Pearson correlation coefficient: Pearson Correlation Coefficient, Pearson correlation coefficient.

[0026] See also Figure 1 , Figure 1 A structural block diagram of a NOx emission concentration prediction system for a circulating fluidized bed boiler provided according to some embodiments of the present specification is shown, and the NOx emission concentration prediction system for a circulating fluidized bed boiler includes: a distributed sensor network, which synchronously collects operating parameters, fuel characteristic data and environmental variables of a target circulating fluidized bed boiler, wherein the operating parameters include dynamic data including exhaust pressure and generator power obtained in real time through a DCS system, the fuel characteristic data include coal particle size distribution obtained through a coal sampling and screening test or an online particle size monitoring device, and volatile matter content measured through coal sampling and testing, and environmental variables include flue gas humidity and intake air temperature continuously recorded by temperature and humidity sensors; a dynamic feature construction module, which uses a sliding time window mechanism to calculate The dynamic change rate of the operating parameters is calculated, and the instantaneous ratio of the coal feed rate to the primary air volume is associated to generate the air-coal coupling feature, and the time lag compensation feature is generated through phase difference analysis under variable load conditions; the hybrid prediction model includes a basic layer and a correction layer. The basic layer outputs the NOx concentration prediction value through the LC mean model, and the correction layer activates the LSTM timing correction network to compensate for the combustion delay effect when load fluctuations are detected; the closed-loop control module calls the preset operating condition mapping model to adjust the secondary air door opening and bed temperature set value based on the deviation between the predicted NOx concentration and the emission limit, so as to maintain stable system operation while ensuring that emissions meet the limit; the health protection module implements a triple protection mechanism of sensor abnormality diagnosis, prediction credibility assessment and safety degradation control.

[0027] A distributed sensor network can refer to a data acquisition system with multiple nodes working in collaboration. For example, an embedded sensor array monitors boiler operating status in real time, enabling the simultaneous collection of circulating fluidized bed boiler operating parameters, fuel characteristics, and environmental variables, improving the comprehensiveness and real-time nature of data collection. Operating parameters can refer to key dynamic indicators during boiler operation, such as exhaust pressure and generator power, acquired in real time through a DCS system. These data reflect the boiler's current operating status and can optimize combustion control strategies. Fuel characteristic data can refer to the physical and chemical properties of the fuel, such as coal particle size distribution and volatile matter content obtained through coal sampling, screening, and testing. These can be used to assess combustion efficiency and optimize NOx generation predictions. Environmental variables can refer to external conditions that affect boiler operation, such as flue gas humidity and inlet temperature recorded through temperature and humidity sensors. These can be used to modify the combustion model's environmental adaptability and improve prediction accuracy.

[0028] The "dynamic feature construction module" can refer to a data processing unit. For example, it uses a sliding time window mechanism to calculate the dynamic rate of change and associates it with the air-to-coal ratio to generate a coupling feature, which is used to enhance the dynamic response capability of the prediction model and reduce the impact of load fluctuations. The "sliding time window mechanism" can refer to a dynamic data processing method, such as using a fixed time window (e.g., 60 seconds) to slide and calculate the rate of change of operating parameters. This can capture short-term fluctuations in boiler operation and enhance model adaptability. The "air-to-coal coupling feature" can refer to an indicator that correlates coal feed rate and primary air volume. For example, it can be generated through real-time ratio calculation and can be used to reflect the fuel-air mixing state during the combustion process to improve NOx prediction accuracy. The "time lag compensation feature" can refer to the phase difference analysis results for variable load conditions. For example, by using historical data to calculate the delayed relationship between load changes and NOx generation, it can correct the time lag error of the prediction model and improve prediction stability under dynamic conditions.

[0029] A hybrid prediction model can refer to an intelligent algorithm framework that combines basic prediction with dynamic correction. For example, the base layer uses the LC mean model to initially predict NOx concentration, while the correction layer activates LSTM timing correction when load fluctuates to improve prediction accuracy and adapt to complex operating conditions. An LSTM timing correction network can refer to a deep learning model. For example, by memorizing historical operating condition change patterns to generate a time lag compensation factor, it can be used to correct for combustion delay effects and improve real-time prediction. A bed temperature setpoint can refer to the target control temperature of the circulating fluidized bed combustion zone. For example, a safety range (850-900°C) is set based on the ash melting point to balance combustion efficiency and coking risk to avoid sudden NOx increases caused by high temperatures. Phase difference analysis can refer to the calculation of the delay relationship between time series signals. For example, a cross-correlation algorithm can be used to determine the time lag between load changes and NOx generation. This can be used to correct the dynamic errors of the prediction model to improve accuracy under variable operating conditions. The combustion delay effect can refer to the time lag between load changes and NOx concentration responses. For example, historical data analysis found that NOx changes are delayed by 8-12 seconds after generator power fluctuations. This can be compensated through LSTM network modeling to improve dynamic prediction accuracy.

[0030] The closed-loop control module can refer to an automatic adjustment system based on prediction results. For example, it can optimize the secondary air damper opening and bed temperature setting values ​​through a preset operating condition mapping model, and dynamically adjust combustion parameters to achieve NOx emission standards. The emission limit can refer to the specified NOx concentration threshold, such as the ultra-low emission standard for thermal power plants (50mg / m 3), which is used as a benchmark target for closed-loop control to achieve compliant emissions. The operating condition mapping model can refer to a multivariable optimization model. For example, a polynomial regression model is used to establish a mapping relationship between the secondary air damper opening, bed temperature and NOx concentration, which is used to dynamically solve the optimal control parameter combination based on the prediction deviation to achieve precise emission control. The secondary air damper opening can refer to the valve position parameter that adjusts the combustion air flow. For example, the air damper opening and closing angle is controlled by a servo motor (range 30-85%), which can optimize the oxygen distribution in the furnace and thus suppress NOx generation.

[0031] ‌Health assurance module can refer to the fault protection mechanism of system operation, such as the implementation of sensor anomaly diagnosis and prediction credibility assessment, which can be used to prevent the risk of miscontrol and improve the reliability of the system. ‌Sensor anomaly diagnosis can refer to the fault detection algorithm, such as identifying abnormal data through the 3σ principle or machine learning model, which can timely isolate the failed sensor signal to avoid the risk of miscontrol. ‌Prediction credibility assessment can refer to the reliability analysis of model output, such as calculating the fluctuation range of the prediction result through KL divergence or confidence interval, which can be used to trigger the model self-update mechanism to improve long-term prediction stability. ‌Safety degradation control can refer to safety protection strategies, such as switching to preset safety parameters (such as a fixed air-coal ratio) when the system is abnormal, which can prevent the boiler from running out of control and ensure the safety of equipment and personnel.

[0032] The present invention is further described below by a detailed embodiment: This example uses a 350t / h ultra-high-pressure circulating fluidized bed boiler at a power plant as an example, detailing the system's operating mechanism. The boiler burns a mixture of local gangue and washed medium coal (volatile matter content 18-25%), achieving precise NOx emissions control within a BMCR load range of 50%-100%.

[0033] The distributed sensing network achieves comprehensive parameter acquisition through three sensing layers: the DCS system collects 27 operating parameters at a 1Hz frequency, including exhaust pressure (range 0-10kPa, accuracy ±0.5%FS) and generator power (range 0-350MW); a laser particle size analyzer uses Mie scattering to monitor coal feed particle size distribution online (measurement range 0.1-15mm), and a near-infrared spectrometer analyzes characteristic absorption peaks to determine volatile matter content (detection limit 0.5%). An array of temperature and humidity sensors, positioned at six key locations throughout the boiler room, continuously records flue gas humidity (10-95%RH) and inlet air temperature (200-350°C). Specifically, embedded infrared spectrometer arrays installed at the four corners of the furnace capture the combustion flame morphology and temperature distribution in real time using dual-band imaging at 3.9μm and 4.3μm wavelengths (spatial resolution 5cm, temperature range 600-1200°C).

[0034] The dynamic feature construction module performs multi-scale feature extraction: Under normal working conditions, the Pearson correlation coefficient filter is used to extract 8 static strong correlation features from the original parameters ( >0.82), including the nonlinear coupling term of bed temperature and secondary air volume; under variable load conditions, the sliding time window (default width 60 seconds) calculates the dynamic change rate of generator power (ΔP / Δt), and at the same time associates the instantaneous ratio of coal feed rate to primary air volume to generate the air-coal coupling feature. When the load change rate exceeds 2% / min, the phase difference analysis algorithm is started: combined with the standard deviation R of the infrared temperature field, the phase difference analysis algorithm is used to calculate the dynamic change rate of generator power (ΔP / Δt). m (typical value 15-45℃), generating a time lag compensation characteristic (value range 0.7-1.3).

[0035] The hybrid prediction model operates using a two-layer cascade architecture: the base layer's LC mean model receives 12 steady-state feature inputs and calculates and outputs a baseline prediction value; the correction layer's LSTM network contains 128 memory cells, which activate when the system detects a load fluctuation standard deviation exceeding 5MW / min. It receives a 20-dimensional time series feature vector and outputs a compensation value. Field tests have shown that the model's prediction error under load mutation conditions has been reduced from ±38mg / m / to ±38mg / m / within the traditional method. 3 Reduced to ±9 mg / m 3 .

[0036] The closed-loop control module implements hierarchical regulation based on the working condition mapping model: when the predicted NOx concentration exceeds the limit (40mg / m 3 ), the secondary damper opening is prioritized (in steps of ±3% / s). If the desired effect is not achieved within 30 seconds, the bed temperature setpoint is adjusted (at a rate of change ≤ 2°C / min). Specifically, during rapid load changes, the system dynamically adjusts control target weights to prevent control oscillation. This strategy has reduced fan adjustment frequency by 62%.

[0037] The health assurance module implements triple protection: It detects sensor anomalies through Mahalanobis distance calculation (threshold λ=3.5). For example, if the infrared spectrometer data suddenly changes by more than three times the standard deviation, it automatically switches to virtual sensor data based on the historical temperature field model. It uses the prediction interval coverage ratio (PICP) (target value ≥92%) to assess model credibility. When both the bed temperature (>920°C) and the predicted NOx value exceed the limit, safety degradation control is triggered, and operation is carried out according to the preset "coal feed rate-air volume" optimal ratio curve.

[0038] The beneficial effects of one of the embodiments in the specification include at least: a data acquisition and dynamic feature construction module of the distributed sensing network, which realizes multi-dimensional representation of the combustion state through spatio-temporal coupling analysis; a hybrid prediction model, which adopts a double-layer architecture to decouple the static combustion law and dynamic delay effect, and uses the stable output of the base layer and the adaptive compensation of the correction layer to improve the prediction accuracy; a closed-loop control module, which forms a synergistic effect of feedforward and feedback based on the intelligent decision mechanism of the working condition mapping model, to avoid control overshoot; and a health guarantee module, which constructs a triple protection system to ensure data reliability and achieve redundancy protection under fault conditions. Through the above design, the present application can realize high-precision prediction and real-time control of NOx emission of the coal-fired unit in the full working condition range, and significantly improve the stability and environmental protection performance of the boiler operation.

[0039] In some embodiments, the distributed sensing network further includes an embedded infrared spectrometer array for real-time capture of the combustion flame morphology and temperature field distribution in the furnace through a multi-band optical sensor.

[0040] ‌The embedded infrared spectrometer array can refer to a multi-sensor group integrated at key positions of the boiler, for example, using an 8-14 μm band infrared detector matrix (pitch 20 cm x 20 cm) to continuously scan the furnace cross section, for inversing the flame temperature gradient (resolution ± 5℃) through radiation intensity, which can identify the combustion uneven area in real time to improve the spatial accuracy of NOx prediction.‌The multi-band optical sensor can refer to a composite spectral detection device, such as synchronously collecting radiation signals of three bands of visible light (400-700 nm), near-infrared (700-1100 nm), and thermal infrared (8-14 μm), which can be used to construct three-dimensional features of the flame morphology (such as height / area / flashing frequency) to quantify the influence of combustion stability on NOx generation.‌The combustion flame morphology can refer to the geometric and dynamic characteristics of the flame, such as extracting the flame envelope area (2-8 m 2 ) and pulsation frequency (5-15 Hz) through image processing algorithms, which can reflect the mixing degree of fuel and air, and correct the local deviation of the wind-coal coupling characteristics.‌The temperature field distribution can refer to the thermodynamic state matrix of the furnace cross section, such as interpolating the infrared data to generate a 500x500 pixel temperature cloud map (range 800-1200℃), which is used to detect the position deviation of the local high temperature area (> 1100℃) to avoid the concentrated generation of thermal NOx.

[0041] ‌As a specific example: in the embodiment of the system, an embedded infrared spectrometer array (containing 16 detection units) scans the four corners and the center area of the furnace at a sampling rate of 10 Hz, and synchronously acquires flame shape data through multi-band optical sensors (visible light channel resolution 1920x1080 pixels, infrared channel temperature measurement accuracy ±3°C). The dynamic feature construction module fuses the flame envelope area (real-time calculation value 3.5-6.2 m 2 ) and the temperature field standard deviation (45-90°C) to generate a combustion uniformity index. When the front wall temperature is detected to be too high (more than +50°C than the average), the closed-loop control module is triggered to preferentially adjust the secondary air door of the corresponding area (opening degree increased by 8-12%). The mixed prediction model cross- validates the flame pulsation frequency (7.2 Hz-9.8 Hz) with the volatile content, and if it is found that the combustion delay exceeds the threshold value (ΔT>4s), the LSTM network is used to compensate the prediction value (correction amplitude ±15 mg / m 3 ). The health guarantee module performs bad point diagnosis on the infrared array (continuous 3 frames of data missing switch to standby unit), ensuring that the temperature field data reliability is greater than 98%.

[0042] Through the infrared spectrum array, the combustion state is holographically monitored, the recognition sensitivity to local high temperature and uneven fuel mixing is significantly improved, and the NOx prediction model has spatial dimension correction capability. The deep coupling of multi-band optical data and working condition parameters effectively overcomes the response lag problem of traditional methods to flame shape changes, and still maintains a prediction error of less than 5% under variable load conditions. The closed-loop control module combines the temperature field distribution to accurately adjust the air volume, which avoids the efficiency loss caused by excessive regulation and ensures that the emission concentration meets the standard. The three health guarantee mechanisms further strengthen the long-term reliability of the system in harsh industrial environments, and provide an innovative solution for clean combustion control.

[0043] In some embodiments, the dynamic feature construction module uses Pearson correlation coefficient to screen static strong correlation features under steady state conditions.

[0044] ‌The Pearson correlation coefficient can refer to a statistical correlation measure index, such as calculating the linear correlation between the air-coal ratio and the NOx concentration (range -1~1) through the ratio of covariance and standard deviation, which is used to screen key static parameters with |r|>0.7, and can optimize the input feature dimension of the basic prediction model to reduce noise interference.‌The static strong correlation feature can refer to a stable correlation variable under steady state conditions, such as selecting the correlation coefficient (r=0.82) between the bed temperature and the oxygen concentration as a fixed input, which can be used to construct a benchmark parameter set for the L-C mean model, to improve the prediction consistency under conventional load.

[0045] As a specific example: In the steady-state stage of the boiler load of 50-100 BMCR%, the dynamic feature construction module first calculates the Pearson coefficient of 12 parameters such as bed temperature and secondary air volume and NOx concentration (sampling period is 5 minutes, sliding window 30 groups of data), and selects strong correlation features (r threshold 0.75) including bed temperature (r=0.79) and flue gas oxygen content (r=0.81). The base layer of the hybrid prediction model combines the above features with the LC mean algorithm to output a benchmark prediction value (such as 180mg / m 3 ±5%; when load fluctuations exceed ±5%, the correction layer automatically switches to dynamic feature mode, compensating for time lag through phase difference analysis (delay time 8-12 seconds). The health assurance module simultaneously monitors sudden changes in the correlation coefficient (such as a sudden drop in the bed temperature r value below 0.6), triggering the sensor verification process (an alarm is issued if the comparison deviation between the infrared array and thermocouple data exceeds 10°C), ensuring the reliability of feature screening.

[0046] The Pearson coefficient is used to scientifically screen features, effectively distinguishing the key influencing factors of steady-state and dynamic conditions and avoiding model overfitting problems caused by redundant parameters. The stable input of static strongly correlated features significantly improves the robustness of the basic prediction layer, keeping the prediction error under normal conditions at a low level for a long time. The intelligent switching mechanism between dynamic and static features not only maintains steady-state computing efficiency but also ensures rapid response capabilities under variable loads. Full-process correlation coefficient monitoring further enhances the system's adaptability to sensor drift and changes in fuel characteristics, providing dual guarantees for environmentally friendly boiler operation.

[0047] In some embodiments, the hybrid prediction model has an online self-update mechanism. When it is detected that the continuous prediction error exceeds a preset threshold, an incremental learning process is triggered to update the model parameters, thereby improving the adaptability of the model under dynamic working conditions.

[0048] The online self-update mechanism can refer to a real-time optimization system for model parameters. For example, when the NOx prediction error exceeds ±15 mg / m3 for three consecutive times, 3 When the fuel is calorific value fluctuates, the incremental learning algorithm (learning rate 0.001-0.005) is automatically called to dynamically adjust the weight matrix of the LSTM network to adapt to changes in emission characteristics caused by fluctuations in the fuel calorific value. The incremental learning process can refer to a parameter iterative update method, such as using small batch gradient descent (batch size = 32) to train only the latest 10% of sample data. It can be used to locally optimize the model while retaining historical knowledge to avoid the waste of computing resources caused by global retraining. The preset threshold can refer to the critical condition that triggers the model update, such as setting the allowable error under static conditions to ±10mg / m 3 , dynamic working condition ±20mg / m 3, when the limit is exceeded for 5 consecutive minutes, the self-update is activated to ensure that the model is within the emission limit (100mg / m 3 ) range of reliability.

[0049] As a specific example: When the boiler is running at 90% load, the hybrid prediction model monitors the deviation between the NOx prediction value and the online analyzer data in real time (sampling interval is 1 minute). When the error exceeds the threshold (static condition ±12mg / m 3 , dynamic ±25mg / m 3 ), the online self-update mechanism is activated: the incremental learning process extracts the latest 30 sets of samples (including 12-dimensional features such as flame shape and bed temperature) from the real-time database, updates the LSTM hidden layer weights through the Adam optimizer (learning rate 0.003), and freezes the base layer parameters to prevent overfitting. The health assurance module synchronously verifies the performance of the updated model (the test set MAE must be <8mg / m 3 ), if the verification fails, it will roll back to the previous version and trigger an alarm. The closed-loop control module adjusts the secondary air damper (opening change ±5%) according to the new model output to stabilize the emission concentration at 92-105mg / m 3 interval.

[0050] Dynamic model optimization is achieved through an online self-update mechanism, significantly reducing prediction deviations caused by changes in fuel composition or equipment aging, ensuring that emissions control always closely matches actual operating conditions. The incremental learning process significantly reduces computing resource consumption while ensuring real-time model performance, avoiding the production interruption risks associated with traditional retraining methods. The intelligent classification strategy for error thresholds further enhances the system's ability to distinguish between steady-state and dynamic conditions, ensuring the accuracy and timeliness of control commands. Overall, a closed-loop intelligent "monitoring-learning-control" system is formed, providing a sustainable automated solution for clean combustion.

[0051] In some embodiments, the closed-loop control module dynamically adjusts the control target weight during the load change process to avoid system oscillation while ensuring that the emission limit is not exceeded.

[0052] The relative importance or priority coefficients assigned to different, and often conflicting, control objectives by the closed-loop control module when making multi-objective optimization decisions.

[0053] The variable load process can refer to the boiler power adjustment phase. For example, when the load changes at a rate of 2-5% / min, the matching of fuel and air volume lags (3-8 seconds). This is used to trigger dynamic control strategies to reduce combustion instability caused by inertia. Dynamic adjustment of control target weights can refer to adjusting the relative importance or priority coefficients assigned to different, often conflicting, control objectives. System oscillation can refer to unstable fluctuations in the control loop, such as the periodic swing of the damper opening within a ±15% range (frequency 0.1-0.3Hz). The introduction of a deadband compensation algorithm (threshold ±3%) can suppress oscillations to ensure a smooth transition in combustion efficiency.

[0054] ‌As a specific example: the closed-loop control module monitors the NOx concentration and steam pressure change rate in real time during the process of the boiler load increasing from 80% to 95% (rate 3% / min). When the pressure fluctuation exceeds ±0.2MPa / s, the dynamic adjustment control target weight module takes effect. The oscillation suppression algorithm automatically inserts a 0.5-second delay to cut off the positive feedback chain by analyzing the damper action frequency (>0.2Hz is judged as oscillation). The health protection module simultaneously records the number of limit relaxations (≤3 times per day). If the limit is exceeded, combustion optimization suggestions are triggered (such as adjustment of the premixed gas ratio). Finally, the system maintains the NOx concentration at 108-118mg / m 3 The steam pressure fluctuation is controlled within ±0.15MPa.

[0055] By dynamically adjusting the control target weights, a dynamic balance between environmental performance and stability is achieved during load changes. This not only avoids the risk of control instability caused by stringent low-emission requirements, but also ensures that pollutant concentrations remain within a controllable range. The oscillation suppression mechanism effectively reduces ineffective actuator movements and extends the service life of key equipment such as fans and valves. Dynamically adjusting the control target weights further optimizes energy consumption during the transition phase, reducing fluctuations in the overall boiler thermal efficiency to less than 1.5%. This solution provides both flexible and reliable technical support for thermal power units participating in deep peak regulation.

[0056] In some embodiments, when the health assurance module detects that a sudden change in infrared spectrometer data exceeds a preset standard deviation threshold, it automatically switches to a backup data source to ensure the continuity and stability of the monitoring data.

[0057] ‌Infrared spectrometer data mutation can refer to abnormal flue gas component monitoring, such as when the NOx concentration sampling value fluctuates by more than 3σ (σ=8mg / m 3), a sliding window algorithm (window width 15 seconds) is used to confirm non-noise interference, which is used to trigger the data credibility assessment process, and can promptly detect measurement distortion caused by probe fouling or optical path contamination. The preset standard deviation can refer to the data quality threshold parameter. For example, the CO concentration fluctuation threshold σ = 50ppm is set (based on historical 3-month data statistics). When 5 consecutive sampling cycles exceed this range, the spectrometer can be judged to be abnormal, so as to avoid a single instantaneous error data affecting the control decision. The backup data source can refer to a redundant measurement system, such as automatically switching to a laser absorption spectrometer (accuracy ± 5mg / m 3 ) or soft sensor model (NOx estimation based on bed temperature / air volume), and maintain monitoring continuity through data fusion algorithm (weight 0.7:0.3) to ensure the input reliability of the closed-loop control module.

[0058] As a specific example: when the boiler is running at 70% load, the health protection module analyzes the NOx concentration data output by the infrared spectrometer in real time (sampling frequency 1Hz). When it detects that the standard deviation of the data suddenly increases to 12mg / m3 within 10 seconds, 3 (Preset threshold 8mg / m 3 ), initiating a three-level verification process: first, comparing adjacent sensor data (allowable deviation ±15%), then verifying the light intensity attenuation coefficient (normal range 300-500 lux), and finally checking the ambient temperature (must be <50°C). If any verification fails, within the next control cycle (5 seconds), the system switches to a backup data source—a laser absorption spectrometer (with a 2-second warm-up time)—running in parallel with the LSTM soft-sensing model, with the laser measurement given 70% weight during data fusion. Simultaneously, an audible and visual alarm (priority level 2) is triggered, fault code E207 is recorded, and the maintenance terminal automatically generates a work order for cleaning the spectrometer lenses (estimated to take 30 minutes). The entire switching process ensures that the NOx concentration output interruption time is less than 8 seconds, and the control loop remains closed.

[0059] Through the intelligent diagnosis mechanism of infrared data mutations, the fault recognition rate of the flue gas monitoring system is significantly improved, avoiding control inaccuracies caused by the failure of a single sensor. The dynamic adjustment strategy of the preset standard deviation not only ensures the sensitivity of anomaly detection, but also effectively filters out false triggers caused by fluctuations in normal operating conditions. The rapid switching capability of the backup data source provides multiple guarantees, ensuring the continuity of key process parameter monitoring and providing a stable and reliable data foundation for ultra-low emission control. Overall, a complete health management closed loop is established from fault detection, redundancy switching to maintenance reminders, significantly reducing the risk of unplanned downtime.

[0060] In some embodiments, the LSTM timing correction network of the correction layer generates a time lag compensation factor by memorizing historical operating condition change patterns. The calculation formula for the time lag compensation factor includes:

[0061] wherein, is a time delay compensation factor, used to correct the phase delay between the predicted value and the actual value of NOx; is a history time window length, determined by the sliding time window mechanism of the dynamic characteristic construction module; is the generator power derivative at the t-i moment, calculated by the dynamic change rate of the generator power in the operating condition parameters; is the rate phase difference between the load change rate and the NOx generation rate; is the fuel-air mixing efficiency of the jth combustion area, calculated according to the coal supply particle size distribution and volatile content in the fuel characteristic data; is the primary air volume deviation of the jth combustion area, calculated by the difference between the real-time value and the reference value of the primary air volume collected by the distributed sensor network; is the spatial variation rate of the fuel volatile content of the jth combustion area, calculated according to the volatile content monitored online by the near-infrared spectrometer; is the exponential decay coefficient, is the non-linear strength of the control air volume change influence, is the suppression coefficient of the fuel gradient disturbance, wherein, , and is initially set to a preset value and is updated through the LSTM time sequence correction network training.

[0062] The historical time window length can refer to a dynamic time window length, for example, set to 60-180 seconds (corresponding to 120-360 sampling points) depending on boiler capacity. A sliding window mechanism updates the historical data queue every 5 seconds to capture the complete inertial characteristics of load changes and adapt to operating conditions with varying load change rates. The generator power derivative can refer to the differential term of power change. For example, the first-order derivative of generator power (in MW / s) is calculated in real time by a central controller. A five-point difference method (with a step size of 0.5 seconds) is used to eliminate measurement noise, accurately reflecting the instantaneous trend of boiler load changes and quantifying the inertial effects of the thermal system. The fuel-air mixing efficiency can refer to the combustion zone mixing efficiency coefficient. For example, based on coal feeder speed feedback (accuracy of ±2 rpm) and online coal quality analysis (volatile matter 28-35%), a radial basis function (RBF) model is used to calculate an efficiency value of 0.65-0.92 for each combustion zone to compensate for fuel distribution unevenness. Primary air volume deviation can refer to the dynamic offset of air volume, for example, the percentage deviation (±15%) between the real-time measurement value and the set value of an array of air pressure sensors (16 points / burner). Using a Kalman filter (process noise Q = 0.01) to eliminate fluctuations, this method can characterize the instantaneous equilibrium state of the air distribution system. The spatial variation rate of fuel volatile content can refer to the fuel property gradient. For example, the spatial distribution of volatile content can be obtained through near-infrared spectroscopy (resolution 0.5 nm). The Sobel operator is used to calculate the variation rate (unit: % / m) between adjacent burners. This can be used to detect pulverized coal pipeline blockage or uneven distribution. The exponential decay coefficient can refer to a time-decay weighting factor, for example, with an initial value of 0.85 (range 0.7-1.0) and automatically adjusted every 30 minutes (step size 0.005) through an LSTM network. For example, when the load change rate exceeds 3% / min, it increases to 0.93 to emphasize the contribution of recent power changes. The nonlinear intensity of the effect of air volume changes can be referred to as the air volume nonlinear coefficient, with an initial value of 1.2 (range 0.8-1.5) and dynamic optimization based on historical air-to-coal ratio data (optimal range 2.8-3.2). For example, it can be reduced to 1.05 when burning high-moisture coal. This adjusts the sensitivity of air volume deviation to the compensation factor. The fuel gradient disturbance suppression coefficient can be referred to as the fuel disturbance suppression factor, with an initial value of 0.3 (range 0.1-0.5) and automatically increased to 0.45 when ▽F_j>5% / m is detected. A smoothing correction factor is generated in combination with volatile gradient data to eliminate prediction jitter caused by coal seam thickness fluctuations.

[0063] The ‌Time Lag Compensation Factor‌ can refer to a phase correction parameter, which is used to superimpose a dynamic adjustment amount within a preset range (e.g., -5% to +7%) on the predicted value, thereby eliminating the lag in NOx measurement values ​​caused by boiler thermal inertia. The ‌Phase Delay‌ can refer to the difference in signal response time. For example, there is a 40-90 second delay in the change in NOx concentration after fuel adjustment (depending on the boiler capacity). By correcting the predicted value 3 control cycles (15 seconds) in advance through the compensation factor output by the LSTM network, the control command can be synchronized with the actual emission changes. The ‌Rate Phase Difference‌ can refer to the difference in the timing of the change of dynamic parameters. For example, there is an 8-12 second time offset between the load increase rate and the NOx generation rate. Based on the rate phase difference, an exponential decay term (exp(- / α)), which is used to quantify the weight coefficient of inertia effect under different working conditions.

[0064] The correction layer, constructed using an LSTM time-series correction network, effectively addresses the time lag problem inherent in traditional prediction models when handling boiler dynamics, significantly improving emission prediction accuracy under variable load conditions. The multi-parameter fusion calculation of the time-lag compensation factor considers both macro factors such as equipment inertia and microscopic heterogeneity within the combustion zone, adapting the compensation effect to varying operating conditions. The active phase delay compensation mechanism enables predictive control, significantly reducing overshoot caused by response delays. Overall, a complete predictive optimization chain, from feature extraction and time-series learning to real-time correction, is formed, providing a more precise time-dimensional solution for clean combustion control.

[0065] In some embodiments, the formula for calculating the rate phase difference includes:

[0066] in: is the instantaneous value of the primary air volume at the kth sampling time, which is collected in real time by the distributed sensor network; is the phase angle of the wind-coal coupling characteristic, which is generated by the ratio correlation of the dynamic characteristic construction module; is the standard deviation of the furnace temperature field distribution, which is calculated from the multi-band data of the infrared spectrometer array; is the mean value of the temperature field, The sequence is obtained by taking the sliding average; is the flame morphology fluctuation rate, obtained by analyzing the time series data of the infrared spectrometer array; is the ambient humidity correction coefficient, which is obtained by normalizing the flue gas humidity recorded by the temperature and humidity sensor; M is the number of sampling points in the infrared spectrometer space.

[0067] The instantaneous primary air volume can refer to real-time data on the burner's inlet air. For example, an array of differential pressure sensors (accuracy ±0.5 Pa) collects data from 16 measurement points across the air duct every 0.2 seconds. The flow distribution is reconstructed using a cubic spline interpolation algorithm to accurately calculate the actual oxygen supply to each combustion zone. The characteristic phase angle of air-coal coupling can refer to the dynamic response parameter of fuel-air mixing. For example, a lag angle within the range of 0-π radians can be calculated through cross-spectral analysis (window length 5 seconds) of the feeder speed (sampling rate 10 Hz) and the damper opening (resolution 0.1%). This can characterize the mixing delay characteristics of different coal types. The standard deviation of the furnace temperature field distribution can refer to a combustion uniformity indicator. For example, a three-dimensional temperature field (grid accuracy 50 mm) is generated using kriging interpolation based on data from a 32-channel infrared spectrometer (wavelength range 2-5 μm). The standard deviation of the cross-sectional temperature difference (unit: °C) is then calculated to detect flame center deviation or coking anomalies. The average temperature field value can refer to the overall furnace heat load parameter. For example, a sliding average (with a 30-second window width) of infrared scan data (update frequency 2Hz) can eliminate transient fluctuations of ±15°C caused by combustion pulsation, reflecting the boiler's steady-state thermal power output. The flame shape fluctuation rate can be an indicator of combustion stability. For example, by extracting the flame profile envelope from an infrared video stream (frame rate 25fps), the area change rate between adjacent frames is calculated (threshold 0.5% / s). If this value exceeds the threshold for 10 consecutive seconds, a warning of combustion oscillation risk can be issued. The ambient humidity correction factor can be a compensation factor for atmospheric conditions. For example, using a dew point sensor (accuracy ±0.3°C) to measure flue gas moisture content, an exponential normalization formula can be used to map humidity levels between 30% and 80% RH to a correction range of 0.8-1.2 to eliminate the impact of air density changes on air volume measurements.

[0068] Through the rate phase difference calculation model with multi-dimensional parameter fusion, the precise quantification of the air-coal-temperature coupling effect in the dynamic process of the boiler was achieved for the first time, solving the problem that traditional methods are difficult to capture complex time-varying characteristics. The real-time data acquisition architecture based on the distributed sensor network ensures the synchronization and spatial resolution of the multi-physical field parameters required for the calculation. The collaborative analysis method of infrared spectroscopy and fluid dynamics parameters effectively overcomes the limitations of a single signal source. The introduction of the environmental humidity factor significantly improves the calculation stability under different seasonal conditions. Overall, a complete technical chain from data perception, feature extraction to timing compensation has been formed, which provides a more accurate dynamic process description capability for ultra-low emission control, and at the same time lays the key parameter foundation for the digital twin modeling of the combustion system.

[0069] In some embodiments, the operating condition mapping model calibrates the nonlinear mapping relationship between NOx concentration and operating parameters through experiments, and preferentially adjusts the secondary air valve opening.

[0070] As a specific example: when the boiler load increases from 75% to 82%, the control system first queries the working condition mapping model, matches the latest 5 groups of historical working condition data (similarity > 92%) under the current coal quality characteristics (volatile matter 32%, ash 18%), and extracts the corresponding optimal secondary air damper opening suggestion value (upper air damper 58°±2°, lower air damper 42°±1°). Then, the system sends adjustment instructions to the field PLC through the Modbus protocol, and the 12 secondary air dampers are adjusted to the target position within 8 seconds, while the oxygen content sensor feedback data in real time (current value 3.1%→3.4%). After the adjustment is completed, the system compares the actual NOx concentration (from 82mg / m 3 to 76mg / m 3 ) with the predicted value of the curved surface library (78mg / m 3 ), and automatically updates the response surface coefficient of the working condition point (correction weight 0.15). The whole process makes the combustion efficiency increase by 0.8 percentage points, and does not cause the steam parameter fluctuation (pressure deviation <±0.05MPa).

[0071] The predictive adjustment mechanism constructed by the working condition mapping model significantly improves the control response speed in the variable working condition process, and avoids the emission fluctuation caused by the traditional trial-and-error method. The precise coordinated adjustment of the secondary air dampers ensures the NOx emission reduction effect and maintains the combustion stability. The continuous self-updating function of the database makes the model always adapt to the equipment aging and coal type change, forming an intelligent control closed loop from data accumulation, optimization calculation to execution feedback, and providing reliable real-time decision support for clean combustion.

[0072] In some embodiments, the safety degradation control module switches to a preset operation curve when the bed temperature or the wind pressure exceeds the safety range, and the curve is generated according to the similarity matching of historical safe working conditions.

[0073] As a specific example: when the air distribution plate pressure sensor detects that the wind pressure drops to 1.2kPa (lower than the safety threshold of 1.5kPa), the safety degradation control module completes the following actions within 80ms: first, call the historical database of the last 1 hour, and filter out 5 groups of similar working conditions (load rate 82±3%, same coal type), calculate the matching degree by dynamic time warping algorithm (optimal matching 92.7%); then load the corresponding preset operation curve (secondary air damper opening gradient increases 5% / s, coal supply decreases in steps), and trigger the audible and light alarm (frequency 2Hz); the system monitors the bed temperature change rate in real time during the switching process (controlled within-3℃ / s), and successfully stabilizes in the safety range (wind pressure recovers to 1.6kPa, bed temperature drops from 880℃ to 830℃) after 15 seconds, and finally generates an event report recording the whole process of abnormal handling.

[0074] The hardware-level emergency control module establishes a millisecond-level safety protection system, effectively eliminating the potential delay risks associated with traditional software control. An intelligent curve matching mechanism based on historical operating conditions ensures the reliability of the control strategy while retaining adaptability to specific abnormal operating conditions. The entire solution implements a closed-loop protection system from anomaly detection, rapid decision-making, and safe execution, significantly improving the safety margin of boiler operation and providing an intelligent solution for equipment protection under extreme operating conditions.

[0075] The preferred embodiments disclosed above are intended only to help illustrate this specification. The optional embodiments do not exhaustively describe all details, nor do they limit the invention to the specific embodiments described. Obviously, many modifications and variations are possible based on the content of the present invention. These embodiments are selected and described in detail in this specification to better explain the principles and practical applications of the present invention, thereby enabling those skilled in the art to better understand and utilize this specification. This specification is limited only by the claims and their full scope and equivalents.

Claims

1. A circulating fluidized bed boiler NOx emission concentration prediction system, characterized in that: include: A distributed sensor network synchronously collects operating parameters, fuel property data, and environmental variables of the target circulating fluidized bed boiler. The operating parameters include dynamic data such as exhaust pressure and generator power acquired in real time by the DCS system. The fuel property data includes coal particle size distribution obtained through coal sampling and screening tests or online particle size monitoring devices, and volatile matter content measured through coal sampling and testing. Environmental variables include flue gas humidity and intake air temperature continuously recorded by temperature and humidity sensors. The dynamic feature construction module uses a sliding time window mechanism to calculate the dynamic change rate of operating parameters, and associates the instantaneous ratio of coal feed rate to primary air volume to generate air-coal coupling characteristics. Under variable load conditions, it generates time-lag compensation characteristics through phase difference analysis. A hybrid prediction model, comprising a base layer and a correction layer. The base layer outputs NOx concentration predictions using an LC mean model, while the correction layer activates an LSTM timing correction network to compensate for combustion delay effects when load fluctuations are detected. The closed-loop control module uses a preset operating condition mapping model to adjust the secondary air damper opening and bed temperature setpoints based on the deviation between the predicted NOx concentration and the emission limit; The health assurance module implements a triple protection mechanism of sensor abnormality diagnosis, prediction credibility assessment, and safety degradation control.

2. The system according to claim 1, wherein: The distributed sensing network also includes an embedded infrared spectrometer array for capturing the combustion flame morphology and temperature field distribution in the furnace in real time through multi-band optical sensors.

3. The system according to claim 1, wherein: The dynamic feature construction module uses the Pearson correlation coefficient to screen static strong correlation features under steady-state conditions.

4. The system according to claim 1, wherein: The hybrid prediction model has an online self-updating mechanism. When it is detected that the continuous prediction error exceeds a preset threshold, an incremental learning process is triggered to update the model parameters, thereby improving the adaptability of the model under dynamic working conditions.

5. The system according to claim 1, wherein: The closed-loop control module avoids system oscillation by dynamically adjusting the control target weight during the load change process while ensuring that the emission limit is not exceeded.

6. The system according to claim 1, wherein: When the health protection module detects that the infrared spectrometer data suddenly changes and exceeds a preset standard deviation threshold, it automatically switches to the backup data source to ensure the continuity and stability of the monitoring data.

7. The system according to claim 1, wherein: The LSTM timing correction network of the correction layer generates a time lag compensation factor by memorizing the historical working condition change pattern. The calculation formula for calculating the time lag compensation factor includes: in, is the time lag compensation factor, which is used to correct the phase delay between the predicted value and the actual value of NOx; is the length of the historical time window, which is determined by the sliding time window mechanism of the dynamic feature construction module; is the generator power derivative at the time ti, which is the dynamic rate of change calculated by the generator power in the operating condition parameters; is the phase difference between the load change rate and the NOx generation rate; is the fuel-air mixing efficiency of the jth combustion zone, which is calculated based on the coal feed particle size distribution and volatile matter content in the fuel characteristic data; is the primary air volume deviation of the j-th combustion zone, which is calculated based on the difference between the instantaneous value of the primary air volume collected in real time by the distributed sensor network and the reference value; is the spatial variation rate of the fuel volatile content in the jth combustion zone, calculated based on the volatile content monitored online by the near-infrared spectrometer; is the exponential decay coefficient, In order to control the nonlinear intensity of the influence of air volume change, is the suppression coefficient of the fuel gradient disturbance, where 、 and The initial setting is a preset value and is updated through the LSTM timing correction network training.

8. The system according to claim 7, characterized in that The calculation formula for calculating the rate phase difference includes: in: is the instantaneous value of the primary air volume sampled at the kth time, collected in real time by the distributed sensor network; is the phase angle of the wind-coal coupling feature, generated by the ratio correlation of the dynamic feature building module; is the standard deviation of the furnace temperature field distribution, calculated from the multi-band data of the infrared spectrometer array; is the mean value of the temperature field, The sequence is obtained by taking the sliding average; is the flame morphology fluctuation rate, obtained by analyzing the time series data of the infrared spectrometer array; is the ambient humidity correction coefficient, which is obtained by normalizing the flue gas humidity recorded by the temperature and humidity sensor; M is the number of sampling points in the infrared spectrometer space.

9. The system according to claim 1, wherein: The operating condition mapping model calibrates the nonlinear mapping relationship between NOx concentration and operating parameters through experiments, and preferentially adjusts the secondary air door opening.

10. The system according to claim 1, wherein: The safety degradation control module switches to a preset operating curve when the bed temperature or wind pressure exceeds a safety range. The curve is generated based on similarity matching of historical safety operating conditions.

Images