A denitration ammonia injection intelligent control method and system based on multi-model fusion and AI feedforward

Abstract

The application discloses a kind of based on multi-model fusion and AI feedforward's denitration intelligent control method and system of ammonia injection, it is related to industrial flue gas denitration environmental protection technical field, including obtaining the historical operation data of denitration system and using offline neural network training to establish feedforward AI model, obtain current working condition parameter input model and obtain feedforward ammonia injection reference quantity, real-time receive special working condition trigger signal, by PLC execution model hard switching and call corresponding special working condition correction model calculation correction bias, real-time acquisition outlet NOx actual value, calculate fine adjustment amount by double closed loop structure of cascade control, feedforward ammonia injection reference quantity, correction bias and fine adjustment amount are superimposed to generate final ammonia injection instruction;The application is fused by offline AI feedforward, special working condition hard switching and cascade feedback, realize the precise and rapid control of ammonia injection amount under the condition of violent fluctuation, avoid NOx emission, mainly used for coal-fired, glass kiln and chemical flue gas denitration.

Description

Technical Field

[0001] This invention relates to the field of industrial flue gas denitrification and environmental protection technology, and in particular to a denitrification ammonia injection intelligent control method and system based on multi-model fusion and AI feedforward. Background Technology

[0002] Flue gas from industrial processes such as coal combustion, chemical production, and glass kilns contains nitrogen oxides (NOx), which are one of the main sources of air pollution. In selective catalytic reduction (SCR) denitrification technology, precise control of ammonia injection is crucial to ensuring denitrification efficiency and preventing ammonia escape exceeding standards.

[0003] Traditional ammonia injection control often employs theoretical calculation models based on material balance or stoichiometry. However, in actual industrial environments, these models suffer from the following technical shortcomings: First, theoretical calculation models struggle to adapt to dynamic factors such as fuel changes, load fluctuations, and equipment aging, leading to significant deviations between the model and actual operating conditions. Second, for systems experiencing severe oscillations, proportional-integral-derivative feedback control exhibits lag in response, making the system prone to instability and resulting in instantaneous exceedances of nitrogen oxide emissions or a surge in ammonia escape. Third, existing AI-based control methods often require online learning, exhibiting the drawback of trial and error after initial corrections, which fails to meet the stringent environmental regulations that prohibit exceeding emission standards, and also poses cybersecurity risks due to reliance on cloud servers. Fourth, after the controller issues the theoretical ammonia injection command, the valves or pumps react slowly, resulting in a mismatch between the actual injection volume and the demand. Therefore, this invention proposes a denitrification ammonia injection intelligent control method and system based on multi-model fusion and AI feedforward to address the problems existing in the prior art. Summary of the Invention

[0004] To address the aforementioned problems, the present invention aims to propose an intelligent control method and system for denitrification ammonia injection based on multi-model fusion and AI feedforward. The present invention achieves precise and rapid control of ammonia injection under drastically fluctuating operating conditions by integrating offline AI feedforward, hard switching under special operating conditions, and cascade feedback, thereby avoiding excessive NOx emissions and solving the problems existing in the prior art.

[0005] To achieve the objectives of this invention, the invention is implemented through the following technical solution: a smart control method for denitrification ammonia injection based on multi-model fusion and AI feedforward, comprising the following steps: Step 1: Obtain historical operating data of the denitrification system, including inlet NOx concentration, flue gas flow rate, oxygen content, furnace temperature and actual ammonia injection rate. Then, use an offline neural network to train the historical operating data to establish a feedforward AI model. Next, obtain the current operating parameters of the denitrification system and input them into the feedforward AI model to calculate the feedforward ammonia injection rate. Step 2: Receive special working condition trigger signals in real time. When a special working condition trigger signal is received, the PLC performs a hard model switch, calls the special working condition correction model corresponding to the special working condition trigger signal, and calculates the correction bias. Step 3: Obtain the actual NOx value at the outlet in real time, and calculate the fine adjustment amount based on the deviation between the actual NOx value at the outlet and the preset NOx setpoint through a dual closed-loop structure of cascade control. Step 4: Superimpose the feedforward ammonia injection accuracy, the correction bias, and the fine adjustment to generate the final ammonia injection command.

[0006] A further improvement is that, in step one, the amount of training data for the offline neural network is no less than 640,000 sets of historical running data, and the number of training iterations is no less than 1,000 rounds.

[0007] A further improvement is that, in step two, the special operating condition trigger signal is the periodic fire switching signal of the glass furnace, the sudden pressure change signal of the ammonia water pipeline, or the sudden change signal of the inlet NOx instrument.

[0008] A further improvement is made in step three, where the cascade control dual closed-loop structure includes a main controller and a secondary controller. The main controller takes the deviation between the actual outlet NOx value and the preset outlet NOx setpoint as input and outputs the theoretical ammonia injection flow rate. The secondary controller takes the deviation between the theoretical ammonia injection flow rate and the measured ammonia water flow rate as input and outputs the regulating valve opening command or the variable frequency pump speed command.

[0009] Further improvements are made in the following steps: In step two, the hard switching of the model includes: interrupting the current PID feedback calculation, calling the preset control parameters in the model for special operating conditions, and locking the decrement command of the secondary regulator.

[0010] A denitrification ammonia injection intelligent control system based on multi-model fusion and AI feedforward includes the following units: The data acquisition unit is used to acquire historical operating data and current operating parameters of the denitrification system; The feedforward AI modeling unit is connected to the data acquisition unit. It is used to train historical operating data using an offline neural network to build a feedforward AI model, and input the current operating parameters into the feedforward AI model to output the feedforward ammonia injection accuracy. The operating condition identification and model switching unit is connected to the data acquisition unit. It is used to receive special operating condition trigger signals from DCS or field instruments in real time, and to perform hard model switching when a special operating condition trigger signal is received. It calls the special operating condition correction model corresponding to the special operating condition trigger signal and outputs the correction bias. The cascade tracking control unit is used to acquire the actual value of NOx at the outlet in real time, and outputs a fine adjustment amount based on the deviation between the actual value of NOx at the outlet and the preset value of NOx at the outlet through the dual closed-loop structure of cascade control. The superposition calculation unit is connected to the feedforward AI modeling unit, the working condition recognition and model switching unit and the cascade tracking control unit, respectively, and is used to superimpose the feedforward ammonia injection accuracy, the correction bias and the fine adjustment to generate the final ammonia injection command. The PLC execution unit is used to carry and run the feedforward AI modeling unit, the working condition identification and model switching unit, the cascade tracking control unit and the superposition calculation unit, and to control the regulating valve or the variable frequency pump according to the final ammonia injection command.

[0011] A further improvement is that the feedforward AI modeling unit is constructed using a neural network toolbox, and the model parameters of the feedforward AI modeling unit are updated to the PLC execution unit at preset time intervals, and the update process does not interrupt online control.

[0012] A further improvement is that the special working condition correction model preset in the working condition identification and model switching unit includes time sequence control logic for the ammonia injection increase time, holding time and recovery time under the fire switching working condition.

[0013] A further improvement is that the PLC execution unit communicates with the DCS via the MODBUS TCP protocol, with the PLC acting as the client and the DCS as the server.

[0014] The beneficial effects of this invention are as follows: (1) The present invention combines a three-layer structure of feedforward coarse adjustment, hard switching under special working conditions and cascade feedback fine adjustment, which can effectively cope with the working condition oscillation at the minute or even second level and avoid exceeding the peak value of nitrogen oxide emissions.

[0015] (2) This invention uses offline neural network modeling, without online learning and trial and error process, which complies with the environmental protection industry's constraint that it does not allow mistakes to be made first, and does not rely on external networks, thus eliminating the risk of network attacks.

[0016] (3) This invention solves the problem of slow response of the actuator through cascade control, improves the tracking speed and accuracy of the ammonia injection quantity to the calculation command, and solves the problem that a single model cannot cover all working conditions by calling the optimal control strategy through the hard switching mechanism of the model.

[0017] (4) This invention reduces the impact of random errors by using large-scale historical data for neural network training, and the model has strong generalization ability. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the overall system structure of the present invention.

[0019] Figure 2 This is a schematic diagram of the hard switching logic flow of the model in this invention.

[0020] Figure 3 This is a schematic diagram of the cascade control dual closed-loop structure of the present invention.

[0021] Figure 4 This is a schematic diagram of the signal flow of the superposition calculation unit of the present invention. Detailed Implementation

[0022] To enhance understanding of the present invention, the present invention will be further described in detail below with reference to embodiments. These embodiments are only used to explain the present invention and do not constitute a limitation on the scope of protection of the present invention.

[0023] The following embodiments also illustrate the methods and systems described in this invention. Each embodiment is described in the order of the steps in the method claims, and the specific operation of each step is performed by the corresponding functional unit in the system claims.

[0024] Example 1 according to Figures 1-4 As shown, this embodiment takes a glass production line with a daily melting capacity of 600 tons as an example, and provides a method and system for intelligent control of denitrification and ammonia injection based on multi-model fusion and AI feedforward. This production line switches the combustion direction (alternating between south and north fires) every 20 minutes, with the switching process lasting 45 seconds. During the switching period, the NOx concentration in the flue gas flues from a steady-state value of 280 mg / Nm³ to 450 mg / Nm³ or decreases to 120 mg / Nm³ within 30 seconds, with a variation range of ±60%. The process includes the following steps: Implementation of Step One: The data acquisition unit connects to the distributed control system via the MODBUS TCP protocol, with a sampling period of 1 second. The data acquisition unit obtains historical operating data of the denitrification system, specifically including: inlet NOx concentration (mg / Nm³, range 0-1000), flue gas flow rate (Nm³ / h, range 0-300000), oxygen content (%, range 0-25), furnace temperature (°C, range 800-1600), and actual ammonia injection rate (L / h, range 0-5000). This data was continuously collected at 1-second intervals for 60 days, resulting in 5.184 million sets of raw data points.

[0025] The feedforward AI modeling unit is a standalone engineering station with the following configuration: a 3.5GHz CPU, 32GB of memory, Windows 10 Professional operating system, and the Neural Network Toolkit installed.

[0026] The feedforward AI modeling unit incorporates a data preprocessing module. This module performs missing value removal, outlier linear interpolation replacement, and minimum-maximum normalization on the raw historical data. Only after preprocessing can the data be input into the neural network for training. The preprocessing of the 5.184 million sets of raw data was then carried out, with the specific steps as follows: The first step was to remove sampling points with missing data, that is, to delete the entire time period with more than 5 consecutive missing data points, resulting in the removal of 21,000 sets of data.

[0027] The second step involves outlier handling, which involves calculating a threshold of three times the standard deviation for each parameter and marking sampling points that exceed the mean by plus or minus three times the standard deviation as outliers. For each outlier, a linear interpolation of five normal sampling points before and after it is used to replace the outlier. A total of 18,000 outlier groups were processed.

[0028] The third step is to normalize the data, that is, to use the minimum-maximum normalization method to map the numerical range of each parameter to the interval [0,1].

[0029] After preprocessing, a total of 5.145 million valid data sets remained. These data were randomly divided into training, validation, and test sets in a ratio of 7:1.5:1.5.

[0030] Furthermore, the feedforward AI modeling unit also includes a neural network model, which adopts a three-layer feedforward neural network structure with the following specific parameters: Number of input layer nodes: 7, corresponding to inlet NOx concentration, flue gas flow rate, oxygen content, furnace temperature, kiln pressure, fuel quantity, and historical ammonia injection quantity (actual ammonia injection quantity in the previous second).

[0031] Number of hidden layers: 1, Number of nodes: 32, Activation function is hyperbolic tangent function tanh.

[0032] Number of output layer nodes: 1, corresponding to the predicted ammonia injection amount; the output layer activation function is a linear function.

[0033] The loss function is the mean squared error, and the optimization algorithm is the adaptive moment estimation (Adam) algorithm. The initial learning rate is set to 0.001, the first moment decay coefficient β1=0.9, and the second moment decay coefficient β2=0.999.

[0034] Training parameter settings: Batch size is 256 data sets, and the number of iterations is 2000 rounds. After every 100 iterations, the mean squared error of the current model is calculated using the validation set.

[0035] The training process is as follows: From round 1 to round 800, the training set loss decreased from 0.32 to 0.087, and the validation set loss decreased from 0.34 to 0.092. From round 800 to round 1500, the training set loss decreased to 0.041, and the validation set loss decreased to 0.045. From round 1500 to round 2000, the training set loss stabilized between 0.038 and 0.040, and the validation set loss stabilized between 0.042 and 0.044. The round with the smallest validation set loss, i.e., round 1780, was selected as the model parameters for the final feedforward AI model. The mean squared error of this model on the validation set was 0.042, and the mean squared error on the test set was 0.043.

[0036] After the feedforward AI modeling unit is trained, the weight matrix and bias vector of the neural network are extracted. Specifically, in the neural network training software, the weight matrix and bias vector of the network object are read through the interface function. That is, for the connection from the input layer to the hidden layer, its weight matrix W1 and bias vector b1 are read; for the connection from the hidden layer to the output layer, its weight matrix W2 and bias vector b2 are read. The above parameters are then exported as a binary file in the format of a 32-bit floating-point sequence, stored in the order of W1, b1, W2, b2.

[0037] The weight matrix W1 from the input layer to the hidden layer has a size of 32×7, and the hidden layer bias vector b1 has a size of 32×1. The weight matrix W2 from the hidden layer to the output layer has a size of 1×32, and the output layer bias vector b2 has a size of 1×1.

[0038] The weight matrix and bias vector are converted to 32-bit floating-point format and packaged into a binary file in the following order: file header (4 bytes, identifying the model version number), W1 (32×7×4=896 bytes), b1 (32×4=128 bytes), W2 (1×32×4=128 bytes), b2 (4 bytes). The MD5 checksum of the binary file is calculated and appended to the end of the file (16 bytes).

[0039] The engineering station is connected to the PLC execution unit via Ethernet cable. The aforementioned binary file is transferred to the PLC execution unit's memory card using a file transfer protocol. After the transfer is complete, the PLC execution unit reads the file and verifies its MD5 value. If the verification is successful, the file is loaded into the running memory. This process does not interrupt the online control of the PLC execution unit; the original model continues to run until the new model is loaded. Once the new model is loaded, the PLC execution unit automatically switches to using the new model in the next control cycle (200 milliseconds).

[0040] In this embodiment, the PLC execution unit is an S7-1500 series, the central processing unit is a CPU 1515-2PN, the working memory capacity is 1MB, and the load memory is 12MB. The program blocks of the PLC execution unit are written in structured text language.

[0041] During the operation of the denitrification system, the data acquisition unit acquires current operating parameters every 200 milliseconds, including the current inlet NOx concentration (measured value, unit mg / Nm³), current flue gas flow rate (measured value, unit Nm³ / h), current oxygen content (measured value, unit %), current furnace temperature (measured value, unit ℃), current kiln pressure (measured value, unit Pa), current fuel quantity (measured value, unit kg / h), and the actual ammonia injection quantity in the previous control cycle (measured value, unit L / h).

[0042] The above seven current operating parameters are first normalized to obtain a normalized vector. Then, the normalized vector is input into the feedforward AI model to obtain the output result. Finally, the output result is reverse normalized to obtain the feedforward ammonia injection standard quantity, in L / h.

[0043] In this embodiment, under steady-state conditions before ignition switching, the feedforward ammonia injection rate fluctuates between 320 L / h and 350 L / h.

[0044] Implementation of Step Two: The operating condition identification and model switching unit is hardwired to the distributed control system and receives discrete signals. The distributed control system sends a fire start signal (24V DC level signal, active high) 5 seconds before the fire start, and the fire start signal lasts for 60 seconds.

[0045] Furthermore, the scan cycle of the operating condition identification and model switching unit is 50 milliseconds. When the fault start signal is detected to change from low to high, the operating condition identification and model switching unit performs the following hard switching operation: First: An interrupt command is sent to the cascade tracking control unit to pause the current proportional-integral-derivative feedback calculation. This interrupt command enables an interrupt flag, and the integral term of the main controller stops accumulating.

[0046] Second: Call the "Special Condition Correction Model for Fire Switching" preset in the PLC execution unit storage area. This model contains the following time-series control logic: In the first 10 seconds after the fire switch start signal is valid, the correction bias is +160L / h (that is, 160L / h is added to the feedforward ammonia injection standard); from the 10th to the 40th second after the fire switch start signal is valid, the correction bias decreases linearly to +80L / h at a rate of 2.67L / h / second; from the 40th to the 60th second after the fire switch start signal is valid, the correction bias remains unchanged at +80L / h.

[0047] Third: A latch-down command is sent to the secondary controller of the cascade tracking control unit. The latch-down command enables a logic flag. When the flag is true, the secondary controller only allows the output to increase or remain unchanged, and prohibits the output from decreasing. The latch-down duration is 60 seconds.

[0048] After the fire start signal ends (i.e., the signal transitions from high to low), the operating condition identification and model switching unit maintains the corrected model for 30 seconds. During this period, the bias correction is gradually reduced to zero at a rate of 10 L / h every 5 seconds. After 30 seconds, the operating condition identification and model switching unit resumes the default model (normal operating condition model) and clears the interrupt flag and the latch decrement flag.

[0049] Implementation of Step Three: The cascade tracking control unit includes a main controller and a secondary controller, both of which employ a proportional-integral-derivative control algorithm.

[0050] Furthermore, the parameters of the main controller are set as follows: proportional gain Kp1 = 1.2, integral time Ti1 = 60 seconds, derivative time Td1 = 10 seconds, and control period is 200 milliseconds. The input of the main controller is the first deviation e1, where: e1 = Preset export NOx setting - Real-time acquired actual export NOx value The preset outlet NOx setting is 35 mg / Nm³, while the actual outlet NOx value is measured in real time by an online monitoring instrument installed at the flue outlet. The value is transmitted to the analog input module of the PLC execution unit via a 4-20mA current signal, with a range of 0-100 mg / Nm³ and an accuracy of ±1.5%.

[0051] The output of the main controller is the theoretical ammonia injection flow rate, which is calculated using the following formula: Accordingly, the integral term is discretized using the trapezoidal rule, and the differential term is discretized using the backward difference method. The theoretical ammonia injection flow rate is limited to the range of 0 to 600 L / h.

[0052] Furthermore, the parameters of the secondary controller are set as follows: proportional coefficient Kp2 = 2.5, integral time Ti2 = 10 seconds, derivative time Td2 = 2 seconds, and control cycle is 100 milliseconds. The input of the secondary controller is the second deviation e2, where e2 = theoretical ammonia injection flow rate - measured ammonia water flow rate. The measured ammonia water flow rate is measured by an electromagnetic flowmeter installed on the ammonia water pipeline, transmitted via a 4-20mA current signal, with a range of 0-800L / h and an accuracy of ±0.5%.

[0053] The output of the secondary controller is the valve opening command, ranging from 0% to 100%, corresponding to the valve closing to fully opening. The valve is a pneumatic diaphragm control valve, with an input signal of 4-20mA current. The opening degree is linearly related to the input current (4mA corresponds to 0%, 20mA corresponds to 100%).

[0054] The fine adjustment range output by the cascade tracking control unit, i.e., the valve opening command output by the secondary controller, is 0% to 100%, corresponding to the valve closing to fully opening. The control valve is a pneumatic diaphragm control valve, with an input signal of 4-20mA current. The opening degree is linearly related to the input current (4mA corresponds to 0%, 20mA corresponds to 100%). Accordingly, the fine adjustment range output by the cascade tracking control unit is the valve opening command output by the secondary controller.

[0055] Implementation of Step Four: The superposition calculation unit calculates the final ammonia injection command according to the following formula: Final ammonia injection command = Feedforward ammonia injection accuracy × K f + Correction bias × K c + Fine adjustment amount × K fb in, K f The feedforward coefficient is set to 1.0 in this embodiment; K c As a correction factor, it is set to 1.0 in this embodiment; K fb The feedback coefficient is set to 0.8 in this embodiment (to prevent feedback overshoot). The above coefficient can be modified online by the operator on the touch screen human-machine interface of the PLC execution unit. The modification range is 0.5 to 1.2, and the step size is 0.01.

[0056] The final ammonia injection command is in L / h. This command value, after range conversion (0-600L / h corresponds to 4-20mA), is converted into a 4-20mA current signal via an analog output module and output to the electric positioner of the regulating valve. The regulating valve adjusts its opening according to the input current, ensuring the actual ammonia flow rate follows the command value.

[0057] This embodiment operated for 30 days (including approximately 2160 flame changeover cycles), and the data recorded is as follows: During the flame changeover period, the highest instantaneous peak value of NOx at the outlet was 48 mg / Nm³, and the average peak value was 42 mg / Nm³. There were no instances where the NOx concentration exceeded the set value of 35 mg / Nm³ for more than 3 seconds. The ammonia slip concentration remained below 2.5 mg / Nm³. Compared with historical data using traditional proportional-integral-derivative feedback control, this embodiment reduced the number of NOx exceedance events (defined as outlet NOx > 50 mg / Nm³ for more than 5 seconds) during flame changeover periods from an average of 18 times per month to 0 times.

[0058] Example 2 This embodiment applies to the denitrification system of a coal-fired boiler in a chemical plant. The system handles a flue gas volume of 180,000 Nm³ / h. A downstream induced draft fan tripped due to a frequency converter malfunction, causing the flue gas pressure to plummet from -500 Pa to -3500 Pa within 10 seconds, and the inlet NOx concentration to jump from 300 mg / m³ to 800 mg / m³. The steps include: Implementation of Step One: The difference between this embodiment and Embodiment 1 is that the training data for the feedforward AI model in this embodiment consists of 700,000 sets of historical operating data (including inlet NOx concentration, flue gas flow rate, oxygen content, furnace temperature, kiln pressure, fuel quantity, and historical ammonia injection quantity over the past 40 days), and the number of iterations is 1000 rounds. After training, the weight matrix is ​​extracted and deployed to the PLC execution unit.

[0059] When a sudden pressure change occurs, the feedforward AI modeling unit receives the current operating parameters within the current control cycle (200 milliseconds): current inlet NOx concentration (785 mg / m³), current flue gas flow rate (178000 Nm³ / h), current oxygen content (4.2%), current furnace temperature (1120℃), current kiln pressure (-3480 Pa), current fuel quantity (8.2 t / h), and the actual ammonia injection rate from the previous control cycle (310 L / h). After normalization, these parameters are input into the feedforward AI model, which calculates and outputs a feedforward ammonia injection rate of 680 L / h (approximately 2.2 times the steady-state value). The entire feedforward calculation process takes 18 milliseconds and is completed within the 200-millisecond control cycle.

[0060] Implementation of Step Two: The operating condition identification and model switching unit receives the pressure signal from the ammonia water pipeline in real time. This signal is measured by a pressure transmitter installed on the ammonia water pipeline, with a range of 0-1.6MPa, and outputs a 4-20mA current signal. The pressure signal sampling period is 100 milliseconds.

[0061] The operating condition identification and model switching unit internally sets a pressure change rate threshold. When the pressure change rate (pressure drop per unit time) exceeds 0.05 MPa / s, it is determined to be a sudden pressure change signal triggered. In this embodiment, the pressure drops from 0.8 MPa to 0.6 MPa within 2 seconds, with a change rate of 0.1 MPa / s, which exceeds the threshold.

[0062] Upon detecting a sudden pressure change signal, the operating condition identification and model switching unit performs the following hard switch: First: Interrupt the proportional-integral-derivative calculation of the main controller and lock the integral term of the main controller to the value of the last step before triggering.

[0063] Second: The "Pressure Sudden Change Special Condition Correction Model" is invoked. This model calculates the correction bias based on the inlet NOx change rate. The inlet NOx change rate is calculated by collecting five inlet NOx concentration values ​​within one second and fitting the slope using the least squares method. In this embodiment, the inlet NOx change rate is 50 mg / m³ / s. The correction bias calculation formula is: Correction bias = Change rate × 2.5 L·s / m³, which yields 125 L / h.

[0064] Third: Send a lock-down command to the secondary controller for 40 seconds.

[0065] Implementation of Step Three: Within one second of the sudden pressure change, the actual NOx value at the outlet increased from 38 mg / m³ to 85 mg / m³. The main controller calculated the theoretical ammonia injection flow rate. Since the integral term of the main controller was locked, only the proportional and derivative terms participated in the calculation, and the theoretical ammonia injection flow rate jumped from the original value of 320 L / h to 520 L / h.

[0066] The deviation between the theoretical ammonia injection flow rate (520 L / h) and the measured ammonia flow rate (310 L / h) received by the secondary controller is 210 L / h. The secondary controller outputs the valve opening command, which increases from 42% to 78%. At the same time, due to the lockout reduction command taking effect, the secondary controller output only increases and does not decrease, preventing the valve from reverting.

[0067] Implementation of Step Four: The superimposed calculation unit calculates: Final ammonia injection command = 680L / h (feedforward) × 1.0 + 125L / h (correction) × 1.0 + the flow rate corresponding to 78% opening (approximately 490L / h) × 0.8 = 680 + 125 + 392 = 1197L / h. After range limiting (maximum 600L / h), the final output is 600L / h.

[0068] Within 0.8 seconds of receiving the command, the regulating valve moved from 42% opening to 98% opening, and the actual ammonia flow rate increased from 310L / h to 580L / h.

[0069] The entire process from the occurrence of a sudden pressure change to the return of the outlet NOx concentration to the set value was recorded: 3 seconds after the disturbance occurred, the outlet NOx reached a peak of 98 mg / Nm³; 8 seconds later, the outlet NOx decreased to 42 mg / Nm³ and continued to decrease; 12 seconds later, the outlet NOx stabilized at 35 mg / Nm³ ± 3 mg / Nm³. Historical data from traditional proportional-integral-derivative control showed that under the same disturbance, the peak concentration reached 210 mg / Nm³, with a recovery time of 43 seconds.

[0070] Furthermore, in this embodiment, the feedforward AI model is updated monthly, and the specific process is as follows: Step S1: Export the newly added running data (approximately 50,000 sets) from the past 30 days as a CSV file and store it on the local hard drive of the engineer's station.

[0071] Step S2: Launch the Neural Network Toolbox on the engineer's station and load the existing feedforward AI model as the initial model. Merge the new data with the original training data (total data volume of 750,000 sets) and re-normalize them.

[0072] Step S3: Set training parameters: batch size is 256, number of iterations is 500 (because it is a fine-tuning based on the original model, there is no need to train from scratch). The initial learning rate is set to 0.0005.

[0073] Step S4: After training, the mean squared error is evaluated on the validation set to be 0.041, which is better than the original model's 0.043. The round with the smallest mean squared error (round 320) is selected as the new model.

[0074] Step S5: Extract the weight matrix and bias vector of the new model, and calculate the MD5 checksum. Generate a new binary model file.

[0075] Step S6: Transfer the model file to the PLC execution unit's memory card via Ethernet. During the transfer, the PLC execution unit continues to run the original model, and the denitrification control remains unaffected.

[0076] Step S7: After the transmission is complete, the PLC execution unit reads the new model file, verifies that the MD5 value matches the value recorded at the sending end, and then loads the new model into the spare memory area. The loading process takes 1.2 seconds.

[0077] Step S8: At the start of the next control cycle, the PLC execution unit switches the model pointer from the original model to the new model. The switching process is completed within one scan cycle (200 milliseconds) and does not affect the control output.

[0078] Step S9: The old model is retained in the original memory area as a rollback version. The PLC execution unit records the activation time of the new model and monitors the deviation between the actual NOx value at the outlet and the model prediction value for the next 48 hours. If the deviation exceeds ±15% for 10 consecutive control cycles, the system automatically rolls back to the old model.

[0079] The model update process in this embodiment does not require system downtime or switching to manual control, thus achieving a non-disruptive update.

[0080] Example 3 This embodiment illustrates the comprehensive control effect of the present invention when both periodic firing and non-periodic pressure fluctuations exist simultaneously, and demonstrates the hourly average adjustment function, including the following steps: Implementation of Step One: Same as Example 1.

[0081] Implementation of Step Two: This embodiment integrates an hourly average adjustment module into the PLC execution unit. This module uses a 1-hour sliding window to calculate the arithmetic mean of the actual NOx values ​​at the outlet over the past 30 minutes, denoted as AVG. 30min .

[0082] If AVG 30min If the NOx level is more than 5 mg / Nm³ below the preset outlet NOx setting (35 mg / Nm³), the outlet NOx setting will be automatically increased by 3 mg / Nm³ to 38 mg / Nm³ for the next 30 minutes. At the same time, the feedforward ammonia injection rate will be multiplied by a coefficient of 0.95 to reduce the amount of ammonia injected and save ammonia water consumption.

[0083] If AVG 30min If the NOx setting exceeds the preset outlet value by more than 5 mg / Nm³, the outlet NOx setting will be automatically reduced by 3 mg / Nm³ to 32 mg / Nm³ for the next 30 minutes. At the same time, the feedforward spray amount of amino acid will be multiplied by a coefficient of 1.05 to reduce the risk of exceeding the standard.

[0084] The set value can be adjusted from 25 mg / Nm³ to 45 mg / Nm³. Adjustment will be locked if the value is outside the range.

[0085] Furthermore, in this embodiment, within 10 seconds after the glass furnace fire switching signal is triggered, the induced draft fan frequency fluctuation causes a short-term change in flue gas pressure. The operating condition identification and model switching unit then simultaneously receives both the fire switching start signal and the ammonia water pipeline pressure fluctuation signal (pressure change rate 0.03 MPa / s, below the 0.05 MPa / s threshold, but duration exceeding 3 seconds).

[0086] The operating condition identification and model switching unit then sets priorities: the ignition switching condition has a higher priority than the pressure fluctuation condition. When multiple special operating condition trigger signals exist simultaneously, the highest priority special operating condition correction model (ignition switching condition correction model) is invoked, and the pressure fluctuation signal is used as an additional correction term for the bias. The formula for calculating the additional correction term is as follows: Additional bias = Pressure change rate × 1.2 L·s / (MPa).

[0087] In this embodiment, the pressure change rate is 0.03 MPa / s, and the additional correction is 0.036 L / h, which has a negligible impact.

[0088] Implementation of Step Three: Same as Example 1.

[0089] Implementation of Step Four: The superposition calculation formula is the same as in Example 1. Furthermore, this example includes a parameter adjustment page on the touchscreen human-machine interface, allowing the operator to modify the following parameters online: feedforward coefficient K_feedforward (0.5-1.2), feedback coefficient K_feedback (0.5-1.2), main controller proportional coefficient Kp1 (0.5-2.0), main controller integral time Ti1 (30-120 seconds), and secondary controller proportional coefficient Kp2 (1.0-4.0), etc.

[0090] The above parameter modifications take effect immediately without requiring compilation or program download. The touchscreen records the operator account, modification time, and values ​​before and after each parameter modification, storing these records on the PLC execution unit's memory card for export and analysis.

[0091] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the present invention without departing from its framework and scope of application, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A smart control method for denitrification ammonia injection based on multi-model fusion and AI feedforward, characterized in that: Includes the following steps: Step 1: Obtain historical operating data of the denitrification system, including inlet NOx concentration, flue gas flow rate, oxygen content, furnace temperature and actual ammonia injection rate. Then, use an offline neural network to train the historical operating data to establish a feedforward AI model. Next, obtain the current operating parameters of the denitrification system and input them into the feedforward AI model to calculate the feedforward ammonia injection rate. Step 2: Receive special operating condition trigger signals in real time. When a special operating condition trigger signal is received, the PLC performs a hard model switch, calls the special operating condition correction model corresponding to the special operating condition trigger signal, and calculates the correction bias. In Step 2, the special operating condition trigger signal is a periodic fire switching signal of the glass furnace, a sudden pressure change signal of the ammonia water pipeline, or a sudden change signal of the inlet NOx instrument. In Step 2, the hard model switch includes: interrupting the current PID feedback calculation, calling the preset control parameters in the special operating condition correction model, and locking the decrement command of the secondary controller. Step 3: Obtain the actual NOx value at the outlet in real time, and calculate the fine adjustment amount based on the deviation between the actual NOx value at the outlet and the preset NOx setpoint through a cascade control dual closed-loop structure. In Step 3, the cascade control dual closed-loop structure includes a main controller and a secondary controller. The main controller takes the deviation between the actual NOx value at the outlet and the preset NOx setpoint as input and outputs the theoretical ammonia injection flow rate. The secondary controller takes the deviation between the theoretical ammonia injection flow rate and the measured ammonia water flow rate as input and outputs the regulating valve opening command or the variable frequency pump speed command. Step 4: Superimpose the feedforward ammonia injection accuracy, the correction bias, and the fine adjustment to generate the final ammonia injection command.

2. The intelligent control method for denitrification ammonia injection based on multi-model fusion and AI feedforward as described in claim 1, characterized in that: In step one, the training data for the offline neural network consists of no less than 640,000 sets of historical running data, and the number of training iterations is no less than 1,000 rounds.

3. A denitrification ammonia injection intelligent control system based on multi-model fusion and AI feedforward, applied to the denitrification ammonia injection intelligent control method based on multi-model fusion and AI feedforward as described in any one of claims 1-2, characterized in that: Includes the following units: The data acquisition unit is used to acquire historical operating data and current operating parameters of the denitrification system; The feedforward AI modeling unit is connected to the data acquisition unit. It is used to train historical operating data using an offline neural network to build a feedforward AI model, and input the current operating parameters into the feedforward AI model to output the feedforward ammonia injection accuracy. The operating condition identification and model switching unit is connected to the data acquisition unit. It is used to receive special operating condition trigger signals from DCS or field instruments in real time, and to perform hard model switching when a special operating condition trigger signal is received. It calls the special operating condition correction model corresponding to the special operating condition trigger signal and outputs the correction bias. The special operating condition correction model preset in the operating condition identification and model switching unit includes time series control logic for the ammonia injection increase time, holding time and recovery time under the fire switching condition. The cascade tracking control unit is used to acquire the actual value of NOx at the outlet in real time, and outputs a fine adjustment amount based on the deviation between the actual value of NOx at the outlet and the preset value of NOx at the outlet through the dual closed-loop structure of cascade control. The superposition calculation unit is connected to the feedforward AI modeling unit, the working condition recognition and model switching unit and the cascade tracking control unit, respectively, and is used to superimpose the feedforward ammonia injection accuracy, the correction bias and the fine adjustment to generate the final ammonia injection command. The PLC execution unit is used to carry and run the feedforward AI modeling unit, the working condition identification and model switching unit, the cascade tracking control unit and the superposition calculation unit, and to control the regulating valve or the variable frequency pump according to the final ammonia injection command.

4. The intelligent control system for denitrification ammonia injection based on multi-model fusion and AI feedforward as described in claim 3, characterized in that: The feedforward AI modeling unit is built using a neural network toolbox. The model parameters of the feedforward AI modeling unit are updated to the PLC execution unit at preset time intervals, and the update process does not interrupt online control.

5. The intelligent control system for denitrification ammonia injection based on multi-model fusion and AI feedforward as described in claim 3, characterized in that: The PLC execution unit communicates with the DCS via the MODBUS TCP protocol, with the PLC acting as the client and the DCS as the server.

Images