Big data prediction model intelligent ammonia injection control method and system suitable for kiln flue gas denitration

Abstract

The invention discloses a big data prediction model intelligent ammonia spraying control method and system suitable for kiln flue gas denitration. According to the ammonia spraying control method, historical and real-time data of multiple parameters such as the fuel gas flow of the glass kiln are collected firstly, the NOx amount is predicted through modeling of the LSTM-GRU hybrid neural network, the ammonia spraying amount is calculated through self-adaptive control, and the accuracy is ensured through two-stage correction. And the calculated ammonia spraying amount is transmitted to a production master control system to adjust an ammonia supply valve, and automatic alarm and manual control are switched in case of failure. The system comprises a data acquisition module, an intelligent prediction module, a self-adaptive control module and the like, and further comprises a dynamic catalytic reaction simulation test platform for optimization. According to the invention, intelligent automatic control of the big data model is realized, the ammonia spraying amount is accurately adjusted, the denitration efficiency and the ammonia utilization rate are improved, the operation cost is reduced, and the system reliability and stability are enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial pollution control, and in particular to a big data prediction model intelligent ammonia injection control method and system applicable to denitrification of furnace flue gas. Background Art

[0002] Most of the nitrogen oxides in flue gas come from thermal NOx. When the combustion temperature is lower than 1300 °C, very little thermal NOx is generated; while when the temperature is higher than 1300 °C, with the increase of temperature, the generation amount of NOx increases sharply. The fuel gas of the glass melting furnace is premixed and then enters the regenerator to be heated, and then enters the furnace for combustion. Due to the requirements of the glass melting process in the glass furnace, the temperature of the flue gas generated by combustion exceeds 1500 °C, and a large amount of NOx is generated at high temperature. The NOx concentration in the flue gas of the glass furnace exceeds 2000 mg / Nm3. The production operation characteristics of the glass furnace also lead to the periodic fluctuation of the NOx concentration in the flue gas. Under normal circumstances, in order to prevent the furnace from being overheated, the furnace generally needs to perform a flame commutation operation every 20 - 30 minutes, and the commutation interval time is about half a minute. During the commutation period, the combustion needs to be stopped for a short time, resulting in a rapid decrease in the NOx concentration and flow rate of the flue gas. After commutation, in order to maintain the temperature in the furnace, the fuel gas input is increased for about two minutes, resulting in continuous fluctuation of NOx in the flue gas. The trend of the NOx concentration in the flue gas of the glass furnace shows a periodic pattern, such as Figure 1 At present, the medium and high temperature SCR denitrification process is generally used for the flue gas of glass furnaces. Due to the large fluctuation range of the NOx value in the flue gas, the ammonia addition amount cannot be automatically controlled by the PID proportional regulation method. The operator can only manually control the ammonia addition amount according to the hourly average value of NOx at the exhaust port, which often leads to an increase in the ammonia escape rate. Summary of the Invention

[0003] The purpose of the present invention is to provide a big data prediction model intelligent ammonia injection control method applicable to denitrification of furnace flue gas to solve the problems raised in the above background art.

[0004] To achieve the above purpose, the present invention provides the following technical solution: A big data prediction model intelligent ammonia injection control method applicable to denitrification of furnace flue gas, including the following steps: Collect the historical data and real-time data of the commutation operation time, fuel gas flow rate, combustion air volume, flue gas flow rate and NOx concentration of the glass furnace, collect data from the main control system through the modbusTCP protocol, and store them in the accurate ammonia injection control system database; An LSTM-GRU hybrid neural network training prediction model is used to model the historical data of flue gas flow rate and NOx concentration collected, and the prediction module uses sliding window prediction. Considering the influence of glass batching and melting process cycle on NOx emissions, it predicts the amount of NOx generated by the glass furnace in the next period of time, and the prediction frequency can be adjusted within every 15 seconds to 60 seconds; An adaptive control system is adopted. Taking the predicted NOx emission value of the LSTM-GRU hybrid neural network model as the benchmark, superimposing adaptive control calculation, automatically adjusting the control parameters according to the process stage, calculating the corresponding ammonia injection amount, and correcting the lag time caused by slow instrument response. The response period of the adaptive control module does not exceed 15 seconds; Based on the calculated ammonia injection amount, the system monitors the actual NOx emission value in real time, compares it with the predicted value. If there is an obvious deviation, the prediction model is corrected once within this cycle; A preset NOx value at the exhaust port is compared with the CEMS detection data at the exhaust port, and the average difference within a set period of time not less than ten minutes is calculated and corrected once every other period of time, which is used as the secondary correction to ensure the hourly average environmental protection requirements at the exhaust port; The calculated ammonia injection amount data is transmitted by the data transmission module of the intelligent precise ammonia injection control system to the production main control system, and the production main control system adjusts the ammonia supply valve. If the precise ammonia injection system fails, an alarm is immediately triggered and it automatically switches to the manual control mode.

[0005] Furthermore, in the step S2, the historical data of the flue gas flow rate and NOx concentration of the glass furnace collected is preprocessed, and the preprocessing includes data cleaning and normalization processing to improve the training effect of the LSTM-GRU hybrid neural network training prediction model.

[0006] Furthermore, the specific content of the first correction is: when the deviation between the actual NOx emission value and the predicted value exceeds the preset first threshold, some parameters of the prediction model are adjusted according to the preset correction formula, and the correction formula is related to the deviation magnitude and the current process stage.

[0007] Furthermore, during the secondary correction, if the calculated average difference exceeds the preset second threshold, according to the positive or negative and magnitude of the difference, the correction coefficient of the ammonia injection amount is adjusted proportionally, and then the total ammonia supply amount is adjusted.

[0008] A big data prediction model intelligent ammonia injection control system applicable to flue gas denitrification of a kiln furnace includes: Data acquisition module: It conducts data interaction with the glass furnace control system, and in real time collects multi-dimensional data such as NOx concentration, flue gas temperature, flue gas flow rate, furnace pressure, and ammonia slip in the flue gas of the glass furnace at a sampling frequency of once every 2 seconds. It uses an anomaly detection algorithm to remove outliers from the data, and then stores the data in the time series database of the intelligent and precise ammonia injection control system; Intelligent prediction module: It uses a hybrid neural network of LSTM and GRU for modeling. The model evaluation indicators are RMSE (root mean square error) or MAPE (mean absolute percentage error). It predicts the NOx value of the flue gas at the inlet of denitration in the future. The model supports online learning and is automatically updated regularly every day; Adaptive control module: It takes the predicted NOx emission value of the LSTMGRU hybrid neural network model as a benchmark, superimposes adaptive control calculations, automatically adjusts control parameters according to the process stage, and calculates the corresponding ammonia injection amount; Calculation and correction module: Based on the calculated ammonia injection amount, it monitors the actual NOx value at the inlet of denitration in real time through the system, and the NOx data detected by the CEMS at the exhaust port, and makes primary and secondary corrections to the ammonia supply; Data transmission module: It transmits the final ammonia injection amount signal or fault signal to the main denitration control system, and the main control system directly controls the ammonia injection regulating valve to adjust the ammonia injection amount; Fault diagnosis module: It detects the system status in real time. Based on the anomaly detection algorithm, it identifies the fault modes of the sensor fault, prediction module fault, adaptive module fault, and network communication fault systems. When a fault occurs, it automatically switches to the standby control scheme; Kinetics catalytic reaction simulation test platform: It simulates according to the actual catalyst performance of the target denitration device, combines the experimental data of the catalyst manufacturer and the operation data of the glass furnace, and synthesizes a multi-dimensional and continuous reaction model platform through the neural network intelligent algorithm to simulate the actual production situation, provides simulation tests for the big data model intelligent and precise ammonia injection control system, and optimizes the system parameter settings.

[0009] Furthermore, the data acquisition module also has a data encryption function, and encrypts the multi-dimensional data such as NOx concentration and flue gas temperature in the flue gas of the glass furnace during storage and transmission to ensure data security.

[0010] Furthermore, during the prediction process, the intelligent prediction module dynamically adjusts the learning rate. When the decrease amplitude of the loss value during model training is less than the preset value in consecutive multiple training batches, it reduces the learning rate to improve the model convergence speed and prediction accuracy.

[0011] Furthermore, in the adaptive control module, multiple groups of control parameters corresponding to different process stages are set, and according to the process stage identifier obtained in real time, it quickly matches and calls the corresponding group of control parameters for ammonia injection amount calculation.

[0012] Furthermore, the kinetic catalytic reaction simulation test platform can also simulate emergencies under different working conditions, such as a large fluctuation in the fuel gas flow rate and catalyst poisoning within a short period of time, to test the response ability and control effect of the system under extreme conditions.

[0013] Furthermore, the fault diagnosis module adopts a method of fusing multiple anomaly detection algorithms to comprehensively judge whether the system has a fault. Only when different detection algorithms all detect anomalies is it determined that the system has a fault, so as to improve the accuracy of fault diagnosis and reduce false alarms.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: The present invention uses the LSTM-GRU hybrid neural network prediction algorithm, which can deeply mine the laws in historical data and accurately predict the change trend of the NOx quantity in the flue gas at the inlet of denitration. Based on the accurate prediction results, the system can accurately adjust the ammonia injection amount to better match the ammonia injection amount with the NOx generation amount. Compared with the traditional manual adjustment method, it effectively avoids the problems of low denitration efficiency and ammonia escape caused by untimely or inaccurate adjustment, significantly improves the denitration efficiency and the utilization rate of ammonia, reduces the usage amount of denitration agents, and thus saves the denitration cost.

[0015] The adaptive control module can automatically adjust the control parameters according to different stages of the production process and quickly respond to changes in the production process. Moreover, the system is provided with a two-stage correction mechanism to monitor the actual NOx emission value in real time and compare it with the predicted value. Once a deviation is found, it can timely correct the prediction model and the ammonia injection amount, greatly improving the response ability of the system to changes in the actual working conditions. This ensures the accuracy and stability of intelligent and precise ammonia injection, and can ensure the efficient and stable operation of the denitration system regardless of the operating state of the glass furnace.

[0016] The fault diagnosis module monitors the system status in real time. Based on advanced anomaly detection algorithms, it can accurately identify various fault modes such as sensor faults, prediction module faults, adaptive module faults, and network communication faults. Once a fault is detected, the system immediately triggers an alarm and automatically switches to the manual control mode or the standby control scheme, ensuring that the denitration system can still maintain basic operation in case of a fault, avoiding denitration interruption and environmental pollution problems caused by system faults, and improving the reliability and safety of the system.

[0017] By establishing a kinetic catalytic reaction simulation test platform, combining the actual catalyst performance of the target denitration device, the experimental data of the catalyst manufacturer, and the operation data of the glass furnace, the actual production situation is simulated. This provides a comprehensive simulation test environment for the intelligent and precise ammonia injection control system of the big data model, enabling the optimization of system parameters before the system is put into actual production, discovering and solving potential problems in advance, and effectively improving the overall operation stability and performance of the system.

[0018] The big data AI intelligent model has the characteristics of self-learning and self-driven optimization. As the production data continues to accumulate, the model can automatically optimize its own parameters, continuously improving the accuracy of prediction and control. This significantly improves the automation level of the precise ammonia injection system, reduces the dependence on manual intervention, lowers labor costs, and at the same time enhances the intelligent level of the production process, providing strong support for the intelligent upgrading of industrial production. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a trend chart of the NOx concentration in the flue gas of the glass furnace; Figure 2 It is a schematic diagram of the system flow of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0020] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention. Embodiment

[0021] Please refer to Figure 1 -2, the present invention provides a big data prediction model intelligent ammonia injection control method and system applicable to the denitration of flue gas from kilns, aiming to achieve precise control of the total ammonia regulation of the denitration system and improve the performance of the denitration ammonia injection system in terms of timeliness, accuracy, reliability, and stability.

[0022] In an actual industrial scenario, a representative large-scale regenerative glass melting furnace is used as the implementation object. The daily glass melting capacity of this glass furnace reaches

[100] tons, and high-calorific natural gas is used as the main fuel. During the combustion process, the temperature in the furnace is as high as 1550°C - 1650°C, which causes the concentration of thermal NOx in the flue gas to be at a high level for a long time, usually fluctuating violently between 2500mg / Nm³ - 3800mg / Nm³, posing great challenges to the denitration work.

[0023] The data acquisition module is the basic support of the entire system, and its stable and accurate data acquisition ability is crucial. This module establishes a reliable data transmission link with the main control system (PLC or DCS) of the glass furnace through the Modbus TCP protocol. Every 2 seconds, the system automatically collects a variety of key data during the operation of the glass furnace, including multi-dimensional information such as fuel gas flow rate, combustion air volume, glass furnace flue gas flow rate, NOx concentration, flue gas temperature, furnace pressure, and ammonia slip.

[0024] Considering the complexity of the industrial production site environment, data anomalies occur frequently. To ensure the quality of the collected data, an anomaly detection algorithm combining statistical analysis and machine learning is adopted. On the one hand, the 3σ rule is used to preliminarily screen the data, and data points that deviate from the mean by more than 3 standard deviations are marked; on the other hand, the Isolation Forest algorithm is introduced to identify data that deviates from the normal pattern from the data distribution structure. For example, at a certain moment, a maximum value appears in the collected flue gas flow rate data. The 3σ rule initially judges it as an outlier, and the Isolation Forest algorithm further confirms that this data point is isolated in the data space, thus determining it as abnormal data and eliminating it.

[0025] After the outlier processing, in order to make the data smoother and more stable, the moving window average method is used to optimize the data. Taking the NOx concentration data as an example, the size of the moving window is set to a fixed value, that is, the NOx concentration data at consecutive fixed time points is taken each time for arithmetic average calculation, and the obtained average value is used to replace the middle data point within the window. This can effectively eliminate the short-term fluctuations of the data and make the data better reflect the true change trend of the NOx concentration. At the same time, due to the possible problem of time asynchronization of data from different data sources during transmission, through a high-precision time synchronization device, based on the clock of the main control system, the data time of each data source is accurately calibrated to ensure the consistency of all data in time, providing an accurate and reliable data basis for subsequent data analysis and model training. The collected data is stored in an orderly manner in the time series database of the intelligent and precise ammonia injection control system for easy access and in-depth analysis at any time.

[0026] The intelligent prediction module, as a core component of the system, undertakes the crucial task of predicting future NOx emission values. This module uses a hybrid LSTM-GRU neural network for modeling. The LSTM (Long Short-Term Memory network), with its unique gating mechanism, can effectively handle long-term dependencies in long sequence data and has significant advantages in capturing the long-term trends of NOx emissions during the production process of glass furnaces. In this embodiment, the LSTM is set to 2 layers, with 128 nodes configured in each layer, and these nodes can fully learn the complex features in historical data. The GRU (Gated Recurrent Unit), while maintaining the model performance, simplifies the network structure and improves the computational efficiency. It is also set to 2 layers, with 64 nodes in each layer. The fully connected layer is set to 3 layers, with the number of nodes being 128, 64, and 1 respectively. Its role is to integrate the features extracted by the LSTM and GRU and output the final prediction result.

[0027] Considering that the periodic changes in technological processes such as glass batching and melting have an important impact on NOx emissions, the prediction frequency can be flexibly adjusted between every 15 seconds and 60 seconds according to the actual production situation. During the regular production stage of the glass furnace, the process is relatively stable and the change in NOx emissions is relatively gentle. At this time, setting the prediction frequency length to every 30 seconds can not only ensure the prediction accuracy but also reasonably balance the consumption of computing resources. During special technological stages such as the fire-switching operation of the glass furnace, the NOx emissions will show violent fluctuations. To capture these changes in a timely manner, the prediction time length is shortened to 15 seconds to provide more real-time and accurate prediction data for ammonia injection control.

[0028] To evaluate the prediction performance of the model, RMSE (Root Mean Square Error) and MAPE (Mean Absolute Percentage Error) are used as evaluation indicators. RMSE can intuitively reflect the average error degree between the predicted value and the true value, and MAPE measures the accuracy of the prediction from the perspective of relative error. During the model training process, by continuously adjusting the hyperparameters of the model, such as the learning rate and the number of iterations, the values of RMSE and MAPE are minimized as much as possible to improve the prediction accuracy of the model. In addition, the model supports the online learning function. During the low-peak period after the production ends every day, the system will automatically collect the new data generated during the production process of the day and integrate it into the training dataset to retrain the model. In this way, the model can automatically learn new laws and trends as the production data accumulates continuously, and continuously optimize its own parameters to better adapt to the dynamic changes in the production process of the glass furnace.

[0029] The adaptive control module predicts the NOx emission value based on the LSTM-GRU hybrid neural network model, and combines with the adaptive control algorithm to achieve accurate calculation and real-time adjustment of the ammonia injection amount. This module can automatically adjust the control parameters according to different process stages of the glass furnace. During the heating stage of the glass furnace, the fuel gas flow rate and the combustion air flow rate will increase significantly, resulting in a significant increase in the NOx generation amount. At this time, the adaptive control module increases the calculation coefficient of the ammonia injection amount and correspondingly increases the ammonia injection amount to ensure that there is enough ammonia to react with NOx and achieve efficient denitrification. During the heat preservation stage of the glass furnace, the NOx generation amount is relatively stable and low, and the adaptive control module reduces the calculation coefficient of the ammonia injection amount to avoid excessive injection of ammonia, thereby reducing the occurrence of ammonia slip.

[0030] The response speed of the adaptive control module is crucial, and its response period is strictly controlled within no more than 15 seconds. Taking the flame reversal operation of the glass furnace as an example, when the reversal operation occurs, the system can quickly capture the process change within 15 seconds, re-adjust the control parameters, calculate the appropriate ammonia injection amount, and ensure the continuity and stability of the denitrification process. In order to compensate for the lag time caused by the slow response of the instrument, a time compensation algorithm is introduced into the adaptive control module. By accurately measuring and analyzing the instrument response time and combining with the actual data in the production process, a time compensation model is established. When calculating the ammonia injection amount, the output of the ammonia injection amount is adjusted in advance according to the lag time of the instrument, so that ammonia can be injected into the flue gas at the optimal time point and fully react with NOx, thereby improving the denitrification efficiency. Embodiment

[0031] The difference from Implementation 1 is that the calculation and correction module further improves the accuracy of the ammonia injection amount through a two-stage correction mechanism to ensure that the denitrification effect always meets the environmental protection requirements. During the first correction process, the system continuously and real-time monitors the actual NOx value at the denitrification inlet and dynamically compares it with the predicted value. When the deviation between the actual NOx emission value and the predicted value exceeds the preset first threshold, the first correction program is immediately triggered. Suppose the preset first threshold is set at ±60 mg / Nm³. If at a certain moment, the actual NOx value at the denitrification inlet is more than 60 mg / Nm³ higher than the predicted value, the system determines that there is a significant deviation. At this time, the prediction model is adjusted according to the preset correction formula. The correction formula comprehensively considers factors such as the deviation magnitude, the current process stage, and the historical data of the system operation. For example, the correction coefficient = deviation value / (predicted value × process stage correction factor), where the process stage correction factor is preset according to the influence degree of different process stages on NOx emissions. By adjusting the correction coefficient, some parameters of the prediction model are optimized, such as adjusting parameters like weights and biases in the LSTM-GRU hybrid neural network model, enabling the model to more accurately fit the actual production data and improving the prediction accuracy. After completing the first correction of the prediction model, the system recalculates the ammonia injection amount according to the corrected model to ensure that the ammonia injection amount can more accurately match the actual NOx emission situation.

[0032] The secondary correction is mainly to ensure that the hourly average value of NOx at the exhaust port strictly complies with the environmental protection standards. A target value of NOx at the exhaust port is preset, and this target value is determined based on local environmental protection regulations and the actual production situation of the glass furnace, for example, set at 90 mg / Nm³. The system compares the CEMS detection data at the exhaust port with the preset target value every once in a while and calculates the average difference during this period. If the average difference exceeds the preset second threshold (such as ±15 mg / Nm³), the secondary correction process is started. According to the positive or negative and magnitude of the difference, the correction coefficient of the ammonia injection amount is adjusted proportionally. If the average difference is positive, that is, the average value of the actual NOx at the exhaust port is higher than the target value, indicating that the ammonia injection amount is insufficient, the system appropriately increases the correction coefficient of the ammonia injection amount; conversely, if the average difference is negative, the correction coefficient of the ammonia injection amount is correspondingly decreased. In this way, the total ammonia supply is secondarily corrected to ensure that the hourly average value of NOx at the exhaust port is always stable within the range required by environmental protection.

[0033] The data transmission module is responsible for accurately transmitting the final ammonia injection amount signal and the fault signal to the main denitration control system. During the data transmission process, in order to ensure the integrity, accuracy, and reliability of the data, multiple data verification and redundant transmission technologies are adopted. Each time data is sent, the system calculates multiple checksums simultaneously, such as CRC checksum, parity checksum, etc., and sends these checksums together with the data. After receiving the data, the main denitration control system calculates and compares different checksums respectively. If all checksums match, it is considered that the data transmission is correct; if any one checksum does not match, the system determines that an error has occurred during the data transmission and immediately sends a retransmission request to the intelligent precise ammonia injection control system. After receiving the retransmission request, the intelligent precise ammonia injection control system resends the data until the data transmission is successful. In addition, in order to improve the stability of data transmission, a redundant transmission link is adopted, that is, data is transmitted through multiple network links simultaneously. When a certain link fails, the system automatically switches to other normal links to ensure the continuity of data transmission.

[0034] The fault diagnosis module monitors the operating status of the system in real time and accurately identifies the fault modes of the system by integrating multiple anomaly detection algorithms. Multiple technical means are adopted, such as the 3σ rule based on statistical analysis, the isolation forest algorithm based on machine learning, and the autoencoder anomaly detection algorithm based on deep learning. When different detection algorithms all detect anomalies, it is determined that the system has a fault, thereby effectively improving the accuracy of fault diagnosis and reducing the occurrence of false alarms. For example, when detecting a sensor fault, if it is found based on the 3σ rule that the data collected by a certain sensor exceeds the normal range continuously for multiple times, and at the same time the isolation forest algorithm and the autoencoder anomaly detection algorithm also determine that this data point is an anomaly point, it is considered that the sensor may have a fault. Once a fault is detected, the system immediately triggers a comprehensive alarm mechanism, including audible and visual alarms, SMS notifications, and email reminders, etc., to inform the operator in a timely manner. At the same time, the system automatically switches to the backup control plan, such as switching to the manual control mode or enabling the pre-set emergency control strategy. In the manual control mode, the operator can manually adjust the ammonia injection amount according to the real-time data and detailed alarm information provided by the system to ensure that the denitration system can still maintain the basic denitration function during the fault period and avoid excessive NOx emissions caused by system failures.

[0035] Before the big data model intelligent and precise ammonia injection system is put into actual production, it is necessary to conduct comprehensive and in-depth simulation tests through the kinetic catalytic reaction simulation test platform. According to the performance parameters of the actual catalyst of the target denitrification device, accurate simulation is carried out, combined with the experimental data provided by the catalyst manufacturer and a large amount of historical operation data of the glass kiln, and a multi-dimensional continuous reaction model platform is constructed using the neural network intelligent algorithm. In the simulation process, not only the normal production conditions are simulated, but also various emergencies under different conditions are covered, such as a large fluctuation of the fuel gas flow rate by ±40% in a short period of time, catalyst poisoning (simulated catalyst activity decreases by 60%), and sudden power outages. Through simulation tests of these extreme conditions, the system's response capability and control effect under various complex conditions are fully tested. According to the test results, the optimization algorithm is used to fine-tune the parameters of the big data model intelligent and precise ammonia injection control system, such as optimizing the structure and parameters of the prediction model, improving the control strategy of the adaptive control module, and improving the correction logic of the calculation correction module, etc., to ensure that the system can cope with various complex and changeable situations in actual operation and achieve stable and efficient denitrification control. Example

[0036] With the expansion of glass furnace production scale and the increase of process complexity, the lag of spectral analysis data acquisition speed can no longer meet the needs of precise control. Therefore, on the basis of existing data acquisition, a rapid detection method is added. Introducing in-situ analyzers with shorter response time to monitor the changes of NOx, O2, and NH3 in flue gas in real time, and obtaining more accurate NOx and NH3 concentration information is conducive to the establishment of more accurate prediction models and more accurate correction optimization.

[0037] In terms of data processing, multi-source data fusion technology is used. Data from different sensors are correlated and analyzed to build a data fusion model. For example, by combining fuel gas flow, combustion air volume, flue gas temperature and extracted in-situ analytical instrument data, the principal component analysis (PCA) method is used to extract key features, reduce data dimensions, and improve data processing efficiency. At the same time, the autoencoder in deep learning is used to learn the features of the fused data, explore the potential complex relationships between the data, and provide more valuable information for subsequent prediction and control.

[0038] In the intelligent prediction module, we continue to explore more advanced model architectures and optimization algorithms. The introduction of the attention mechanism enables the model to pay more attention to data features closely related to NOx emissions and improve the accuracy of predictions. For example, when processing data from different process stages of a glass kiln, the attention mechanism can automatically adjust the degree of attention to features such as fuel gas flow and combustion-supporting air volume, thereby more accurately capturing the changing trend of NOx emissions.

[0039] In the adaptive control module, develop more refined control strategies. According to factors such as the different production loads, product types, and ambient temperatures of the glass furnace, subdivide the process stages and formulate personalized control parameters for each subdivided stage. For example, when producing glass products of different thicknesses, since the heat demand and combustion process of the furnace are different, correspondingly adjust the calculation coefficients and control parameters of the ammonia injection amount to achieve more accurate ammonia injection control.

[0040] At the same time, strengthen the collaborative optimization between the intelligent ammonia injection control system and other production links of the glass furnace. Conduct data interaction with the furnace temperature control system, and adjust the ammonia injection amount in advance according to the change trend of the furnace temperature to avoid the problem of untimely denitrification caused by the change in the NOx generation amount due to temperature fluctuations. Establish a linkage mechanism with the glass batching system and adjust the ammonia injection strategy in real time according to the change of raw material components to ensure efficient denitrification under different raw material conditions.

[0041] The fault diagnosis module further introduces fault prediction technology. Through the real-time monitoring and analysis of system operation data, use long short-term memory networks (LSTM) or gated recurrent units (GRU) in machine learning to predict the performance of key components such as sensors and actuators, and discover potential fault hazards in advance. For example, by monitoring the signal fluctuations and drift trends of sensors, predict whether the sensors are about to fail, and arrange maintenance and replacement in advance to avoid control errors caused by sensor failures.

[0042] In terms of fault-tolerant control, in addition to switching to the manual control mode, develop more intelligent fault-tolerant control algorithms and increase redundant design. When a certain control module fails, the system automatically adjusts the control logic and uses the redundant functions of other normal modules to achieve partial control functions. For example, when some functions of the adaptive control module fail, through preset backup control rules and calculation methods, combined with the current production data, maintain basic ammonia injection control to ensure that the denitrification system can still maintain a certain operating efficiency during the fault period and reduce the impact on production.

[0043] To facilitate the operator's intuitive understanding and monitoring of the system's operating status, develop a powerful visualization interface. Through real-time data visualization technology, display the operating parameters of the glass furnace, NOx emission concentration, ammonia injection amount, and the system's fault alarm information in the form of charts and graphs. Use geographic information system (GIS) technology to centrally display the denitrification systems of multiple glass furnaces to achieve unified monitoring of distributed production equipment.

[0044] Meanwhile, a remote monitoring and diagnosis platform is established to support the access of mobile terminals such as mobile phones and tablets. Operators can view the operating status of the system, receive alarm information, and perform remote operations and controls through mobile devices anytime and anywhere. When the system fails, remote experts can obtain system data in real time through the platform for remote diagnosis and guiding repairs, improving the efficiency and timeliness of fault handling.

[0045] In terms of data privacy protection, the collected production data of glass furnaces and the operating data of denitration systems are encrypted for storage and transmission. Encryption algorithms such as the Advanced Encryption Standard (AES) are used to encrypt the data to ensure the security of the data during storage and transmission. Meanwhile, a data access audit mechanism is established to track and analyze the access records of the data in detail to prevent data leakage and abuse.

Claims

1. A big data prediction model intelligent ammonia injection control method suitable for furnace flue gas denitrification, characterized in that: The steps include: Collect historical and real-time data of the glass furnace's reversing operation time, fuel gas flow, combustion-supporting air volume, glass furnace flue gas flow, and NOx concentration, collect data from the main control system through the modbusTCP protocol, and store them in the precise ammonia injection control system database; The LSTMGRU hybrid neural network training prediction model is used to model the collected historical data of flue gas flow and NOx concentration. The prediction module uses sliding window prediction to consider the impact of glass batching and melting process cycle on NOx emissions, and predict the amount of NOx generated by the glass furnace in the future. The prediction frequency can be adjusted within every 15 seconds to 60 seconds. Adopting an adaptive control system, taking the predicted NOx emission value of the LSTMGRU hybrid neural network model as a benchmark, superimposing adaptive control calculations, automatically adjusting control parameters according to the process stage, calculating the corresponding ammonia injection amount, and correcting the lag time caused by the slow response of the instrument. The response cycle of the adaptive control module does not exceed 15 seconds; Based on the calculated ammonia injection amount, the system monitors the actual NOx emission value in real time and compares it with the predicted value. If there is a significant deviation, the prediction model will be corrected once within this cycle; The preset outlet NOx value is compared with the outlet CEMS detection data, and the average difference within a set period of time of not less than ten minutes is calculated. It is corrected once every certain period of time as a secondary correction to ensure the environmental protection requirements of the outlet hourly average value; The calculated ammonia injection amount data is transmitted to the production main control system by the data transmission module of the intelligent precision ammonia injection control system, and the production main control system adjusts the ammonia supply valve. If the precision ammonia injection system fails, an alarm will be triggered immediately and it will automatically switch to manual control mode.

2. A big data prediction model intelligent ammonia injection control system suitable for furnace flue gas denitrification, using the method described in claim 1, characterized in that: include: Data acquisition module: It interacts with the glass furnace control system and collects multi-dimensional data such as NOx concentration, flue gas temperature, flue gas flow, furnace pressure, and ammonia escape in the flue gas of the glass furnace in real time at a sampling frequency of once every 2 seconds. It uses an anomaly detection algorithm to remove abnormal values ​​in the data and then stores them in the time series database of the intelligent and precise ammonia spray control system. Intelligent prediction module: LSTMGRU hybrid neural network is used for modeling. The model evaluation index uses RMSE root mean square error or MAPE mean absolute percentage error to predict the future denitrification imported flue gas NOx value. The model supports online learning and is automatically updated every day. Adaptive control module: Taking the predicted NOx emission value of the LSTMGRU hybrid neural network model as a benchmark, superimposing adaptive control calculation, automatically adjusting control parameters according to the process stage, and calculating the corresponding ammonia injection amount; Calculation correction module: Based on the calculated ammonia injection amount, the system monitors the actual denitrification inlet NOx value in real time, and the NOx data detected by the exhaust CEMS, to make primary and secondary corrections to the ammonia supply; Data transmission module: transmits the final ammonia injection amount signal or fault signal to the denitrification main control system, and the main control system directly controls the ammonia injection regulating valve to adjust the ammonia injection amount; Fault diagnosis module: Real-time detection of system status, based on the abnormal detection algorithm, to identify the failure mode of sensor failure, prediction module failure, adaptive module failure, network communication failure system, and automatically switch to the backup control scheme when a failure occurs; Kinetic catalytic reaction simulation test platform: Based on the actual catalyst performance of the target denitrification device, the simulation is carried out, combined with the experimental data of the catalyst manufacturer and the operation data of the glass kiln, and a multi-dimensional continuous reaction model platform is fitted through a neural network intelligent algorithm to simulate the actual production conditions, provide simulation tests for the big data model intelligent and precise ammonia injection control system, and optimize the system parameter settings.

3. The big data prediction model intelligent ammonia injection control method suitable for furnace flue gas denitrification according to claim 1 is characterized in that: In step S2, the collected historical data of glass furnace flue gas flow and NOx concentration are preprocessed, and the preprocessing includes data cleaning and normalization to improve the training effect of the LSTM-GRU hybrid neural network training prediction model.

4. The big data prediction model intelligent ammonia injection control method suitable for furnace flue gas denitrification according to claim 1 is characterized in that: The primary correction is specifically: when the deviation between the actual NOx emission value and the predicted value exceeds a preset first threshold, some parameters of the prediction model are adjusted according to a preset correction formula, and the correction formula is related to the deviation size and the current process stage.

5. The big data prediction model intelligent ammonia injection control method suitable for furnace flue gas denitrification according to claim 1 is characterized in that: During the secondary correction, if the calculated average difference exceeds a preset second threshold, the correction coefficient of the ammonia injection amount is adjusted proportionally according to the sign and size of the difference, thereby adjusting the total ammonia supply amount.

6. The big data prediction model intelligent ammonia injection control system suitable for furnace flue gas denitrification according to claim 2 is characterized in that: The data acquisition module also has a data encryption function, which encrypts the collected multi-dimensional data of NOx concentration and flue gas temperature in the glass furnace flue gas during storage and transmission to ensure data security.

7. The big data prediction model intelligent ammonia injection control system suitable for furnace flue gas denitrification according to claim 2 is characterized in that: The intelligent prediction module dynamically adjusts the learning rate during the prediction process. When the loss value of the model training decreases by less than a preset value in multiple consecutive training batches, the learning rate is reduced to improve the model convergence speed and prediction accuracy.

8. The big data prediction model intelligent ammonia injection control system suitable for furnace flue gas denitrification according to claim 2 is characterized in that: In the adaptive control module, a plurality of control parameter groups corresponding to different process stages are set, and according to the process stage identifiers obtained in real time, the corresponding control parameter groups are quickly matched and called to calculate the ammonia injection amount.

9. The big data prediction model intelligent ammonia injection control system suitable for furnace flue gas denitrification according to claim 2 is characterized in that: The kinetic catalytic reaction simulation test platform can also simulate emergencies under different working conditions, such as large fluctuations in fuel gas flow and catalyst poisoning in a short period of time, to test the system's response capability and control effect under extreme conditions.

10. The big data prediction model intelligent ammonia injection control system suitable for furnace flue gas denitrification according to claim 2, characterized in that: The fault diagnosis module adopts a fusion method of multiple abnormality detection algorithms to comprehensively judge whether the system has a fault. Only when different detection algorithms detect an abnormality, the system is determined to have a fault, so as to improve the accuracy of fault diagnosis and reduce false alarms.

Images