Accurate ammonia spraying control method for SCR denitration system based on multi-mode perception

Abstract

The invention discloses an accurate ammonia injection control method for an SCR (Selective Catalytic Reduction) denitration system based on multi-modal perception, relates to the technical field of SCR denitration, and aims at solving the problems of hysteresis quality, insufficient precision, poor adaptability and the like of the existing denitration system, the method constructs a hybrid neural network model, and fuses multi-modal data such as coal quality characteristics and boiler operation conditions in real time. The model predicts the three-dimensional spatial and temporal distribution of the flue gas speed, temperature and NOx concentration at the inlet of the SCR reactor, so as to realize dynamic optimization and accurate control of ammonia injection. The denitration efficiency can be remarkably improved, ammonia consumption and ammonia escape are reduced, and the service life of the catalyst is prolonged.

Description

Technical Field

[0001] This invention relates to the field of SCR denitrification technology, and more specifically to a method for precise ammonia injection control in an SCR denitrification system based on multimodal sensing. Background Technology

[0002] Coal-fired power generation is one of the world's main methods of electricity production, but the flue gas it produces contains large amounts of nitrogen oxides (NOx). NOx is a major precursor to acid rain, photochemical smog, and PM2.5, posing a serious threat to environmental quality and human health. Selective catalytic reduction (SCR) technology has become the mainstream technology for flue gas denitrification in coal-fired power plants. SCR technology, under the action of a catalyst, injects a reducing agent into the flue gas, causing it to react with the NOx in the flue gas to produce harmless nitrogen and water.

[0003] The core objective of an SCR denitrification system is to ensure high denitrification efficiency while keeping ammonia slip at an extremely low level. High denitrification efficiency helps meet increasingly stringent environmental emission standards, while low ammonia slip prevents unreacted ammonia from being released with the flue gas, thus avoiding corrosion, blockage, and secondary pollution of downstream equipment.

[0004] Traditional SCR ammonia injection control strategies are mainly divided into two categories: feedback control and simple feedforward control. Feedback control strategies typically rely on measured NOx and / or ammonia slip concentrations at the SCR reactor outlet. The controller adjusts the total ammonia injection rate based on the deviation between these measurements and the setpoint using algorithms such as proportional-integral-derivative (PID). The advantage of this method is its relatively simple control logic and ability to correct for overall system deviations. However, its drawback lies in control lag. The entire process from the furnace outlet to the SCR reactor, through catalyst reaction, and finally to the outlet monitoring point involves a physical lag of several seconds or even tens of seconds. When changes in operating conditions cause problems with outlet NOx or ammonia slip, the controller can only adjust after the problem has occurred and been detected, making rapid response difficult. This is especially problematic under dynamic conditions such as drastic load fluctuations or frequent coal type changes, easily leading to excessive ammonia slip or decreased denitrification efficiency.

[0005] Simple feedforward control strategies attempt to reduce lag through prediction. Typically, the required total ammonia injection is estimated based on macroscopic operating parameters such as boiler load and total NOx at the SCR inlet using a pre-set mathematical model or lookup table. This method improves response speed to some extent, but its accuracy is limited by the representativeness of the parameters. More importantly, this simple feedforward control cannot handle the spatiotemporal nonuniformity of the flue gas flow field. It only focuses on matching the total amount, ignoring the degree of local spatial matching between ammonia and NOx. Even if the total ammonia and total NOx are matched in molar ratio, if the ammonia is not spatially uniform, it can still lead to ammonia excess in some areas causing ammonia escape, while insufficient ammonia in other areas results in incomplete denitrification.

[0006] In summary, existing SCR ammonia injection control technologies generally suffer from the following problems:

[0007] 1. Neither feedback control nor feedforward control based on macroscopic parameters can effectively eliminate control lag caused by physical processes.

[0008] 2. Most methods are unable or have difficulty effectively addressing the spatiotemporal nonuniformity of the flue gas flow field, temperature field, and NOx concentration field at the SCR inlet, resulting in poor local matching between ammonia injection and NOx.

[0009] 3. When coal type, load and other operating conditions change frequently, existing methods are difficult to maintain the stability of denitrification efficiency and ammonia slip.

[0010] 4. Existing methods fail to shift the starting point of control to the source of the combustion process, namely coal quality and combustion state, thus making it impossible to predict and adjust in advance before disturbances occur.

[0011] Therefore, it is necessary to propose a precise ammonia injection control method for SCR denitrification systems based on multimodal sensing to solve the above problems. Summary of the Invention

[0012] The purpose of this invention is to solve the problems of control lag and insufficient handling of the spatiotemporal nonuniformity of flue gas in the existing ammonia injection control method of SCR denitrification system.

[0013] To achieve the above objectives, the present invention specifically adopts the following technical solution:

[0014] A method for precise ammonia injection control in SCR denitrification systems based on multimodal sensing includes the following steps:

[0015] a. Real-time sensing layer: Real-time acquisition of coal quality characteristics parameters and flue gas combustion state field parameters at the furnace outlet of coal-fired power plant boilers;

[0016] b. Intelligent prediction layer: Construct and run a deep neural network model, which takes the coal quality characteristic parameters, the combustion state field parameters and the boiler operating condition parameters as inputs to predict the spatiotemporal distribution of flue gas velocity field, temperature field and NOx concentration field at the inlet section of the SCR reactor in the future.

[0017] c. Ammonia injection strategy generation layer: Based on the predicted spatiotemporal flow field at the SCR reactor inlet, dynamically determine the ammonia demand distribution in each region of the SCR reactor inlet section, and use intelligent optimization algorithms to calculate the optimal ammonia injection quantity allocation scheme for multiple nozzles of the SCR reactor.

[0018] d. Composite Control Execution Layer: Executes composite control strategies across multiple time scales, including:

[0019] i. Short-cycle predictive feedforward control based on the optimal ammonia injection quantity allocation scheme;

[0020] ii. Fast-cycle fine-tuning feedback control based on online monitoring of NOx and NH3 concentrations in the flue gas at the SCR reactor outlet;

[0021] Furthermore, the system for acquiring the coal quality characteristic parameters in the real-time sensing layer is an online coal quality analysis system. The online coal quality analysis system uses a laser Raman spectroscopy online analysis system, a near-infrared spectroscopy online analysis system, or a laser-induced breakdown spectroscopy system to acquire the moisture, ash, volatile matter, fixed carbon, nitrogen, sulfur content, and calorific value of the coal entering the furnace.

[0022] Furthermore, the system for acquiring the combustion state field parameters of the flue gas at the furnace outlet in the real-time sensing layer is a non-contact combustion state field sensor array, which employs an acoustic tomography system, an infrared optical tomography system, a fiber optic array temperature measurement system, or a laser gas analyzer array.

[0023] Furthermore, in the intelligent prediction layer, the boiler operating condition parameters include boiler load, primary air volume, secondary air volume, tertiary air volume, burnout air volume, supply air temperature, feed water flow rate, steam drum pressure or furnace pressure.

[0024] Furthermore, in the intelligent prediction layer, the deep neural network model adopts a hybrid neural network architecture that includes long short-term memory networks, gated recurrent units, convolutional neural networks, or Transformer structures.

[0025] Furthermore, in the ammonia injection strategy generation layer, the intelligent optimization algorithm is a genetic algorithm or a reinforcement learning algorithm.

[0026] Furthermore, in the composite control execution layer, the short-cycle predictive feedforward control executes ammonia injection by sending instructions to the electric ammonia valve or the pneumatic ammonia valve.

[0027] Furthermore, in the composite control execution layer, the online monitoring system for NOx and NH3 concentrations in the SCR reactor outlet flue gas on which the fast-cycle fine-tuning feedback control is based is a high-resolution online monitoring array. The array adopts a tunable diode laser absorption spectroscopy scanning system, a chemiluminescence gas analyzer array, or an ultraviolet differential absorption spectroscopy gas analyzer array.

[0028] Furthermore, in the composite control execution layer, the fast-cycle fine-tuning feedback control uses a multivariable PID controller or a model predictive controller to correct the short-cycle predictive feedforward control command.

[0029] Furthermore, in the composite control execution layer, the long-cycle adaptive control is achieved by periodically retraining and calibrating the deep neural network model offline or online to adapt to slowly changing processes such as catalyst activity decay, in-furnace coking, and sensor drift.

[0030] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0031] This invention constructs a hybrid neural network model that integrates multimodal data, including coal quality characteristics and boiler operating conditions, in real time. The model predicts the three-dimensional spatiotemporal distribution of flue gas velocity, temperature, and NOx concentration at the SCR reactor inlet, thereby achieving dynamic optimization and precise control of ammonia injection. This significantly improves denitrification efficiency, reduces ammonia consumption and slip, and extends catalyst life. Detailed Implementation

[0032] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0033] This invention provides a method for precise ammonia injection control of an SCR denitrification system based on multimodal perception. Its core lies in building a full-chain intelligent prediction and control system from the combustion source to the SCR ammonia injection actuator, so as to achieve advanced and precise control of the flue gas denitrification process.

[0034] 1. Construction of Multi-Source Real-Time Sensing Layer

[0035] This invention first deploys online monitoring equipment at key locations in coal-fired power plant boilers to acquire key parameters affecting SCR performance in real time and from multiple dimensions. It comprehensively captures all key factors influencing flue gas characteristics and NOx formation.

[0036] 1.1. Online sensing of coal quality characteristic parameters

[0037] An online coal quality analysis system is deployed at the coal feeder inlet or pulverized coal pipeline of a coal-fired power plant. This system can acquire various key physicochemical properties of the coal entering the furnace in real time and continuously. Technologies that can be used include, but are not limited to:

[0038] Online coal quality analysis system using laser Raman spectroscopy: This system uses a high-energy laser beam to irradiate a moving coal sample and collects the resulting Raman scattering spectral signals. By analyzing the frequency shift and intensity of the scattered light, it can obtain key combustion characteristic parameters of coal in real time, non-contactly, and non-destructively, including industrial analysis (such as moisture, ash, volatile matter, and fixed carbon content), elemental analysis (such as carbon, hydrogen, oxygen, nitrogen, and sulfur content), and calorific value. Its advantages include fast response speed (typically within seconds), high measurement accuracy, and minimal impact from the surface condition of the coal sample and ambient humidity. In particular, the nitrogen content in coal is a major determinant of fuel-type NOx formation.

[0039] Near-infrared spectroscopy (NIR) online coal quality analysis system: Utilizing the characteristic absorption bands of near-infrared light in coal samples, and through the establishment of a pre-calibration model, it rapidly analyzes components such as moisture, ash, and volatile matter in coal. This technology is characterized by its fast analysis speed and relatively low cost.

[0040] Laser-Induced Breakdown Spectroscopy (LIBS) system: This system uses a high-energy pulsed laser to instantaneously heat and vaporize the surface of a coal sample, forming a plasma. By analyzing the spectrum emitted by the plasma, the elemental composition of the coal sample, including carbon, hydrogen, oxygen, nitrogen, and sulfur, can be obtained rapidly and online.

[0041] The aforementioned analysis system can provide real-time data streams of coal fed into the furnace at a frequency of seconds or minutes. This data includes, but is not limited to:

[0042] Moisture content: directly affects the combustion temperature of coal, flame propagation speed, and flue gas volume.

[0043] Ash content: affects heat transfer in the furnace, coking tendency and fly ash characteristics.

[0044] Volatile matter content: determines the ignition characteristics of coal, NOx formation during the rapid combustion stage, and flame stability.

[0045] Fixed carbon content: mainly affects the combustion characteristics and main heat release of coal.

[0046] Nitrogen content (N): The most direct and critical parameter for predicting fuel-borne NOx formation.

[0047] Sulfur content (S): Affects SO2 formation, which in turn may affect catalyst activity, SO3 conversion rate, and the risk of blockage caused by the formation of ammonium bisulfate (ABS) after ammonia escape.

[0048] Calorific value: This determines the amount of fuel required by the boiler under a specific load. These coal quality characteristics are the primary dependent variable in the entire predictive control chain, and their changes are the most fundamental cause of variations in the combustion process and downstream flue gas characteristics.

[0049] 1.2. Combustion State Field Sensing

[0050] A non-contact sensor array is deployed in the boiler furnace outlet area, approximately several meters to tens of meters upstream of the SCR reactor inlet, to acquire the direct results of the combustion process in real time and at high resolution, namely the spatial distribution characteristics of the flue gas field. This is crucial for capturing combustion non-uniformity caused by factors such as coal type, air distribution, and load variations. Possible technologies include, but are not limited to:

[0051] Acoustic Tomography System: This system uses multiple acoustic transmitters and receivers evenly arranged around the perimeter of the furnace outlet section. The transmitters emit sound waves, and the receivers receive the sound wave signals as they pass through the flue gas medium. Since the propagation speed of sound waves in flue gas is closely related to the flue gas temperature and composition, by measuring the propagation time of sound waves along different paths, tomographic algorithms can be used to reconstruct the two-dimensional or three-dimensional temperature and velocity field distributions within the section in real time. This technology has advantages such as strong penetration, high temperature resistance, and strong resistance to dust contamination.

[0052] Infrared optical tomography (IRT) systems utilize the characteristic absorption or emission properties of specific components in flue gas, such as CO2 and H2O, in the infrared band. By arranging multiple infrared beams at the furnace outlet cross-section, such as through fiber optic transmission, and measuring the attenuation or emission intensity of the beams after passing through the flue gas, combined with tomographic imaging algorithms, the two-dimensional temperature field and the concentration distribution of specific components within the cross-section can be retrieved.

[0053] Fiber optic array temperature measurement system: A high-density, high-temperature-resistant fiber optic probe array is arranged at the furnace outlet section to directly measure the local flue gas temperature. Although it is a contact measurement, it can provide temperature data with high spatial resolution.

[0054] Laser gas analyzer (TDLAS or DOAS) array: Multiple laser beam paths are arranged at the furnace outlet section. Utilizing tunable diode laser absorption spectroscopy (TDLAS) or differential absorption spectroscopy (DOAS) technology, the concentration distribution of gases such as NOx and O2 in different regions is measured in real time. Through array arrangement, the non-uniform distribution of NOx can be preliminarily understood. These sensor arrays can provide real-time, high-resolution two-dimensional or three-dimensional velocity field, temperature field, and initial NOx concentration field data of the flue gas at the furnace outlet section (upstream of the SCR inlet). These data reflect the direct impact of coal quality changes during combustion, providing information for predicting the fine flow field at the SCR inlet.

[0055] 2. Construction of the Intelligent Prediction Model Layer

[0056] The core innovation of this invention lies in constructing a deep neural network model capable of accurately predicting the future spatiotemporal flow field at the SCR inlet. This model learns from multi-source heterogeneous data to establish a complex causal mapping from source perturbations to downstream effects.

[0057] 2.1. Model Input and Output Model Input:

[0058] Real-time coal quality characteristic parameters: Data from the aforementioned online coal quality analysis system, including moisture, ash, volatile matter, fixed carbon, nitrogen content, sulfur content, calorific value, etc. These are one-dimensional time series data characterizing fuel properties.

[0059] Boiler operating parameters include, but are not limited to, boiler load, primary air volume, secondary air volume, tertiary air volume, burnout air volume, supply air temperature, feedwater flow rate, steam drum pressure, and furnace pressure. These are also one-dimensional time series data.

[0060] Model output: The three-dimensional spatiotemporal distribution of flue gas velocity field, temperature field, and NOx concentration field at the inlet cross-section of the SCR reactor over the next tens of seconds to several minutes (e.g., the next 60 to 180 seconds). Here, three-dimensional refers to the two-dimensional spatial distribution within the cross-section plus the evolution over time.

[0061] 2.2. Deep Neural Network Model Architecture

[0062] Considering that the input data has multiple modalities and the output data has high-dimensional spatiotemporal characteristics, this invention adopts a hybrid neural network architecture to fully capture the complex relationships between different data modalities and the dependencies of spatiotemporal sequences.

[0063] Feature extraction layer:

[0064] For one-dimensional time series data such as coal quality characteristics and boiler operating parameters, Long Short-Time Memory (LSTM) networks or Gated Recurrent Units (GRUs) are used to capture their time dependencies. These recurrent neural network (RNN) variants are good at processing sequential data, can remember information from longer time steps, and filter out unimportant noise.

[0065] For combustion state field data, such as furnace outlet temperature field, flow velocity field, and NOx concentration field, due to their spatial structure, convolutional neural networks (CNNs) can be used to extract their spatial features. CNNs perform feature mapping in the spatial dimension through convolutional kernels, which can effectively capture local patterns and structures in the flow field.

[0066] Fusion and Prediction Layer:

[0067] Features extracted from different modalities (using LSTM / GRU and CNN) are fused. This can be achieved through fully connected layers or more complex attention mechanisms (such as the self-attention mechanism in Transformers). Attention mechanisms can dynamically assign weights to different input features based on the importance of the current prediction task, thereby better fusing multimodal information.

[0068] The fused features are input into a deep regression network, which is responsible for outputting the three-dimensional velocity, temperature, and NOx concentration fields of the SCR inlet section at future time steps. Since the output is high-dimensional, such as the velocity, temperature, and NOx values ​​of 30x30 grid points at multiple time steps, this part may require multiple parallel output branches or a subnetwork with a complex output structure.

[0069] 2.3. Model Training and Optimization

[0070] Training dataset:

[0071] Historical operating data: Collect more than one year of operating data from the power plant, including real-time data from all sensing layers, boiler operating parameters, and NOx and NH3 concentration data at the SCR inlet and outlet.

[0072] Computational Fluid Dynamics (CFD) Simulation Dataset: Since it is difficult to obtain high-precision spatiotemporal flow field data for the entire cross-section of the SCR inlet in real power plants as labels, it is necessary to combine CFD simulation data. Using CFD software such as Fluent and CFX, the boiler-SCR system is modeled in detail, simulating flue gas flow, heat transfer, and NOx generation / reduction processes under different coal types, loads, and air distribution conditions. By calibrating the CFD simulation results with actual operating data, a comprehensive dataset that conforms to physical laws and closely approximates actual operating conditions can be constructed.

[0073] Training objective: The training objective of the model is to minimize the error between the predicted future spatiotemporal flow field and the true label. Commonly used loss functions include mean squared error (MSE) or mean absolute error (MAE).

[0074] Training strategy: Offline pre-training is performed using a large-scale dataset, followed by online fine-tuning and continuous learning using real-time data from actual power plant operations. This enables the model to quickly adapt to new operating conditions and environmental changes.

[0075] Hardware platform: Training and inference of deep neural network models require high-performance computing resources, which are typically deployed on servers equipped with GPUs or industrial edge computing devices to meet real-time requirements.

[0076] 3. Dynamic Virtual Mesh Generation and Optimal Ammonia Injection Strategy Generation

[0077] 3.1. Dynamic Virtual Mesh Generation

[0078] Based on the future flue gas velocity field, temperature field, and NOx concentration field of the SCR reactor inlet section output by the intelligent prediction model (e.g., in future seconds T1, T2, ..., Tn), the control system dynamically divides the SCR reactor inlet section into multiple small virtual grids in each control cycle. For example, the entire section is divided into 30x30 grids, totaling 900 virtual grids.

[0079] Unlike traditional fixed grid division, the dynamism here is reflected in the fact that although the number and approximate location of the grids can be preset, the internal properties of the flue gas microparticles represented by each virtual grid are updated and predicted in real time. Furthermore, the grid density can be dynamically adjusted based on the local gradient and rate of change of the predicted flow field. For example, finer grids can be used in areas with large NOx concentration gradients to capture more details.

[0080] Each virtual grid is linked to its predicted future flow rate, temperature, and NOx flux. Flux here refers to the amount of NOx passing through the grid per unit time, combining NOx concentration and flow rate information.

[0081] 3.2. Calculation of Theoretical Ammonia Requirement

[0082] For each dynamic virtual grid, the control system calculates the theoretical ammonia injection rate required for that grid based on the predicted NOx flux, temperature, and flow rate at the future moments it is bound to.

[0083] The calculation formula is based on the stoichiometric ratio of the SCR reaction; for example, the molar ratio of NOx to NH3 is typically between 0.8:1 and 1.2:1, and takes into account the effects of catalyst activity and flue gas temperature on the reaction rate. For example:

[0084] Required NH3 molar amount = predicted NOx molar amount × ammonia-nitrogen molar ratio (AMR).

[0085] Among them, AMR can be dynamically adjusted according to the characteristics of the catalyst, the reaction temperature, and the desired denitrification efficiency.

[0086] By calculating all virtual grids, a high-resolution ammonia demand distribution map is generated. This clearly depicts the local ammonia demand at different locations along the SCR inlet section.

[0087] 3.3. Generation of the optimal ammonia injection strategy

[0088] The ammonia injection grid of an SCR reactor typically has M nozzles with fixed physical positions. For example, a typical 600MW unit may have 80 nozzles, arranged in multiple layers and rows. Each nozzle has its specific injection characteristics and physical position, such as injection angle, coverage area, and flow rate adjustment range.

[0089] This invention utilizes intelligent optimization algorithms to solve the inverse problem of transforming an ammonia demand distribution map into an optimal flow allocation scheme for M physical nozzles. Possible algorithms include:

[0090] Genetic Algorithm (GA): A genetic algorithm is a global optimization search algorithm based on biological evolution. It searches for the optimal solution in a multi-dimensional search space by simulating natural selection and genetic mechanisms (such as selection, crossover, and mutation). In this application, each "individual" represents a nozzle flow allocation scheme, and its "fitness" is determined by indicators such as the degree of matching between the scheme and the ammonia demand distribution map, ammonia slip, and denitrification efficiency. The genetic algorithm can effectively avoid getting trapped in local optima and find the globally or near-globally optimal ammonia injection scheme.

[0091] Reinforcement Learning (RL): Reinforcement learning learns the optimal strategy through the interaction between an agent and its environment. An ammonia injection control system can be viewed as an agent whose "action" is adjusting the flow rate of M nozzles, whose "environment" is the boiler-SCR system and flue gas flow field, and whose "reward" function is linked to indicators such as denitrification efficiency, ammonia slip rate, and NOx emission concentration. Through extensive trial and error and learning, the reinforcement learning model can autonomously discover the optimal ammonia injection strategy.

[0092] The goal of the optimization algorithm is:

[0093] To maximize the matching of ammonia demand distribution: so that the amount of ammonia injected by the M nozzles can cover and satisfy the ammonia demand of each dynamic virtual grid as much as possible in space.

[0094] Minimize ammonia escape: Avoid localized ammonia excess.

[0095] Maximize denitrification efficiency: Ensure NOx is fully reduced.

[0096] Meet the physical constraints of the nozzle: the nozzle flow rate must be within its adjustable range, and the total ammonia content must be within a reasonable range.

[0097] The optimization algorithm completes calculations within milliseconds to hundreds of milliseconds, generating real-time flow distribution instructions for M nozzles. These instructions are then sent to the ammonia injection actuator.

[0098] 4. Execution of multi-timescale composite control strategies

[0099] This invention employs a hierarchical, multi-time-scale composite control strategy that organically combines predictive foresight, feedback accuracy, and adaptive robustness to form an intelligent control closed loop that prioritizes prediction, supplements feedback, and optimizes adaptively.

[0100] 4.1. Short-period predictive feedforward control layer

[0101] Based on real-time sensed coal quality parameters and combustion state field data, the system runs an intelligent prediction model to generate the future SCR inlet spatiotemporal flow field. Then, based on this predicted flow field, the optimal ammonia injection command, i.e., the flow distribution scheme of M nozzles, is calculated through dynamic virtual mesh generation and intelligent optimization algorithms.

[0102] These ammonia injection commands are generated and sent to the ammonia injection actuator before a change in the flow field is predicted, or just as the change begins to appear at the furnace outlet. For example, when the online coal quality analyzer detects a signal that the coal type is about to change, the predictive model immediately starts, calculates the changing trends of the flow field in the furnace and SCR inlet over the next 60-180 seconds, and adjusts the ammonia injection commands in advance. This ensures that when the disturbed flue gas arrives at the SCR reactor, ammonia is ready and injected in an optimal manner.

[0103] Ammonia injection commands are sent to field-operated or pneumatic ammonia valves via industrial control networks (such as Ethernet / IP or Modbus TCP). These valves precisely regulate the ammonia flow rate according to the commands. Electric ammonia valves typically offer higher control accuracy and repeatability, while pneumatic ammonia valves offer faster response times.

[0104] 4.2. Fast Cycle Fine-Tuning Feedback Control Layer

[0105] Although the predictive model has high accuracy, problems such as model errors, unmodeled dynamics, or sensor drift may still exist in actual operation, such as localized catalyst wear and nozzle clogging. To eliminate these effects, this invention deploys a high-resolution online monitoring array at the SCR reactor outlet. Possible technologies include, but are not limited to:

[0106] Tunable Diode Laser Absorption Spectroscopy (TDLAS) Scanning System: This system deploys multiple TDLAS probes to scan the SCR outlet cross-section with a laser beam, acquiring the two-dimensional concentration distribution of NOx and NH3 in real time. TDLAS features fast response, high accuracy, and strong anti-interference capability.

[0107] Chemiluminescence gas analyzer array: Multiple sampling points and chemiluminescence NOx analyzers are arranged at different positions on the outlet section to obtain local NOx concentration.

[0108] Ultraviolet Differential Absorption Spectrometry (UV-DOAS) Gas Analyzer Array: Similar to TDLAS, but operates in the ultraviolet band and can simultaneously measure NOx and SO2.

[0109] The real-time two-dimensional concentration distribution data of NOx and NH3 provided by the SCR outlet monitoring array will serve as feedback signals. The control system (e.g., using a multivariable PID controller or model predictive controller (MPC)) will compare these measured values ​​with preset emission target values ​​(e.g., NOx < 35 mg / Nm³). 3 The NOx concentration is compared with that of NH3 (<3ppm). If a local area is found to have a high NOx concentration or a slight increase in ammonia escape, the feedback correction module will make a small adjustment to the ammonia quantity of the corresponding nozzle at a speed of seconds.

[0110] The main function of this layer is to compensate for prediction errors and ensure the stability and robustness of the system in the short term. It is a closed loop of "prediction as the main method and feedback as the secondary method," and the magnitude of feedback correction is usually small to avoid drastic conflicts with feedforward control.

[0111] 4.3. Long-Period Adaptive Optimization Layer

[0112] During long-term operation, SCR systems face slow-changing processes such as catalyst activity decay, in-furnace coking, burner wear, and sensor drift. To ensure the long-term accuracy and robustness of the system, a long-cycle adaptive optimization layer is introduced.

[0113] The system continuously collects all operational data from the power plant (including coal quality, combustion status, boiler operating conditions, and SCR inlet / outlet data). At regular intervals (e.g., weekly, monthly, or quarterly), the intelligent prediction model is retrained and calibrated offline or online using new data accumulated over a period of time.

[0114] The goal of retraining is to update the model's internal parameters so that it can better adapt to the actual state of the current system. For example, when catalyst activity begins to decline, the model will automatically adjust its predicted ammonia demand by learning new operating data to compensate for the decrease in catalyst efficiency, thereby maintaining denitrification efficiency without increasing ammonia slip.

[0115] Furthermore, this layer is responsible for health monitoring and drift calibration of the sensing layer sensors to ensure the accuracy of input data. This guarantees that the entire intelligent control system maintains optimal performance throughout its lifecycle, possessing strong self-learning and adaptive capabilities.

[0116] Example

[0117] This embodiment uses a 600MW ultra-supercritical coal-fired power generating unit as an example to illustrate in detail the specific application process and effects of the method of the present invention. The unit is equipped with a three-layer ammonia injection grid SCR denitrification system with a total of 80 nozzles, designed denitrification efficiency of 90%, and a NOx emission target of 35mg / Nm³. 3 The ammonia escape target is 3 ppm.

[0118] 1. Deployment of multi-source real-time perception layer:

[0119] An online coal quality analyzer based on laser Raman spectroscopy is deployed at the coal feeder inlet. This instrument provides data on the moisture, ash, volatile matter, fixed carbon, nitrogen content, and calorific value of the coal fed into the furnace every 5 seconds. When the coal type is switched from bituminous coal to lean coal, the system can detect the decrease in nitrogen content and volatile matter in real time.

[0120] An acoustic tomography system was deployed in the boiler furnace outlet flue (approximately 15 meters upstream of the SCR reactor inlet). This system consists of 16 transmitters and 16 receivers, providing 20x20 resolution two-dimensional temperature and velocity field data per second for the furnace outlet cross-section. Simultaneously, six TDLAS laser gas analyzers were positioned at key locations along the cross-section to monitor local NOx concentrations in real time.

[0121] 2. Construction of the intelligent prediction model layer:

[0122] We collected operational data from the power plant over the past year and a half, including the aforementioned sensing data, boiler load, and various air volumes, and combined this data with the unit's CFD simulation model for flow field, temperature field, and NOx concentration field under various operating conditions. We then constructed a hybrid deep neural network model incorporating LSTM, CNN, and Transformer attention mechanisms.

[0123] The model is trained on a high-performance computing platform to learn the complex mapping relationship between "coal quality characteristics + boiler operating conditions + furnace outlet state field → 30x30 resolution flue gas velocity field, temperature field, and NOx concentration field at the SCR inlet for the next 120 seconds". For example, the model learns that when the nitrogen content decreases and the load decreases, the NOx concentration peak will appear in a specific region at the furnace outlet, and predicts its evolution path at the SCR inlet.

[0124] The trained model can receive signals of coal quality and load changes in real time with an inference time of approximately 0.5 seconds. Before coal type switching, the online coal quality analyzer can detect changes in the coal quality of the coal entering the furnace. The model immediately starts and predicts how the flue gas flow field at the SCR inlet section will transition from bituminous coal combustion to lean coal combustion over the next 60 to 180 seconds, including the velocity field, temperature field, and NOx concentration field. For example, it predicts that the NOx concentration in certain areas of the furnace outlet will increase, while the flue gas velocity distribution will change.

[0125] 3. Dynamic Virtual Mesh and Ammonia Injection Optimization:

[0126] Based on the predicted future flue gas flow field, the control system dynamically divides the SCR inlet section into 900 virtual grids of 30x30 pixels per second. Each virtual grid is accompanied by the predicted flow velocity, temperature, and NOx flux for future moments.

[0127] The system calculates the required theoretical ammonia amount for each virtual grid based on the predicted data. For example, if a significant increase in NOx concentration is predicted for a certain area, the system will calculate the required increase in ammonia for the corresponding virtual grid in that area.

[0128] The optimization module, based on a genetic algorithm, receives the ammonia demand from these 900 virtual grids and, combined with the positions, injection characteristics, and adjustable ranges of 80 physical nozzles, calculates the optimal flow distribution scheme for these 80 nozzles within milliseconds. This scheme aims to ensure that ammonia molecules "follow" the predicted high-NOx flue gas clusters, precisely injecting them into the desired locations, rather than blindly and evenly distributing them. For example, if it is predicted that a local NOx concentration peak caused by lean coal combustion will reach the left side of the SCR region, the optimization algorithm will instruct the nozzles in the left region to increase the ammonia supply, while the nozzles on the right side will maintain or decrease the ammonia supply.

[0129] 4. Composite control execution:

[0130] Before predicting changes in the flue gas flow field caused by coal type switching and load changes, the system has already sent the optimal ammonia injection command to the ammonia injection valve. For example, 5 seconds after the online coal quality analyzer detects the coal type switching signal, the prediction model has already given the prediction results for the next 60 seconds and adjusted the ammonia injection rate accordingly to ensure that ammonia is ready when the NOx concentration peak reaches the SCR reactor.

[0131] The TDLAS scanning system deployed at the SCR outlet monitors the two-dimensional concentration distribution of NOx and NH3 in real time. If a slight increase in ammonia slip or a higher NOx concentration is detected in a localized area, the feedback correction module uses a multivariable PID controller to make minute adjustments to the ammonia flow rate of the corresponding nozzle at a rate of seconds, eliminating the impact of prediction errors. For example, if ammonia slip is found to be high in the lower right corner of the SCR outlet, the system will fine-tune the ammonia flow rate of the relevant nozzles in the lower right corner area.

[0132] The system continuously collects daily operational data from the power plant. Every weekend, the predictive model is retrained and calibrated offline using data from the previous week. For example, after a year of catalyst operation, when the catalyst activity begins to decline, the model automatically adjusts its internal parameters by learning new operational data, thereby better matching the ammonia injection rate to actual demand and maintaining denitrification efficiency and ammonia slip at optimal levels. The model can also compensate for slight sensor drift through adaptive learning.

[0133] Through the above implementation method, the NOx emission concentration at the SCR outlet of this power plant remained stable within the target range (e.g., 35 mg / Nm³) during coal type switching and load fluctuations. 3 Fluctuation range ±2mg / Nm 3 The ammonia slip rate was controlled below 3 ppm (average 1.5-2.5 ppm), far superior to the performance of traditional control methods under similar operating conditions. With traditional methods, NOx emissions could surge to 50-60 mg / Nm³ during coal type switching. 3 Ammonia slip can reach 5-8 ppm. The method of this invention significantly improves the robustness of the system and its environmental benefits.

[0134] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. The scope of patent protection of the present invention shall be determined by the claims. Similarly, any equivalent structural changes made based on the content of the present invention's specification shall also be included within the scope of protection of the present invention.

Claims

1. A method for precise ammonia injection control in an SCR denitrification system based on multimodal sensing, characterized in that, Includes the following steps: a. Real-time sensing layer: Real-time acquisition of coal quality characteristics parameters and flue gas combustion state field parameters at the furnace outlet of coal-fired power plant boilers; b. Intelligent prediction layer: Construct and run a deep neural network model, which takes the coal quality characteristic parameters, the combustion state field parameters and the boiler operating condition parameters as inputs to predict the spatiotemporal distribution of flue gas velocity field, temperature field and NOx concentration field at the inlet section of the SCR reactor in the future. c. Ammonia injection strategy generation layer: Based on the predicted spatiotemporal flow field at the SCR reactor inlet, dynamically determine the ammonia demand distribution in each region of the SCR reactor inlet section, and use intelligent optimization algorithms to calculate the optimal ammonia injection quantity allocation scheme for multiple nozzles of the SCR reactor. d. Composite Control Execution Layer: Executes composite control strategies across multiple time scales, including: i. Short-cycle predictive feedforward control based on the optimal ammonia injection quantity allocation scheme; ii. Fast-cycle fine-tuning feedback control based on online monitoring of NOx and NH3 concentrations in the flue gas at the SCR reactor outlet; iii. Long-period adaptive control that optimizes the deep neural network model based on long-term operating data.

2. The method for precise ammonia injection control in an SCR denitrification system based on multimodal sensing according to claim 1, characterized in that, The system for acquiring the coal quality characteristics parameters in the real-time sensing layer is an online coal quality analysis system. The online coal quality analysis system uses a laser Raman spectroscopy online analysis system, a near-infrared spectroscopy online analysis system, or a laser-induced breakdown spectroscopy system to acquire the moisture, ash, volatile matter, fixed carbon, nitrogen, sulfur content, and calorific value of the coal entering the furnace.

3. The method for precise ammonia injection control in an SCR denitrification system based on multimodal sensing according to claim 1, characterized in that, The system for acquiring the combustion state field parameters of the flue gas at the furnace outlet in the real-time sensing layer is a non-contact combustion state field sensor array. The array adopts an acoustic tomography system, an infrared optical tomography system, a fiber optic array temperature measurement system, or a laser gas analyzer array.

4. The method for precise ammonia injection control in an SCR denitrification system based on multimodal sensing according to claim 1, characterized in that, In the intelligent prediction layer, the boiler operating parameters include boiler load, primary air volume, secondary air volume, tertiary air volume, burnout air volume, supply air temperature, feed water flow rate, steam drum pressure or furnace pressure.

5. The method for precise ammonia injection control in an SCR denitrification system based on multimodal sensing according to claim 1, characterized in that, In the intelligent prediction layer, the deep neural network model adopts a hybrid neural network architecture that includes long short-term memory networks, gated recurrent units, convolutional neural networks, or Transformer structures.

6. The method for precise ammonia injection control in an SCR denitrification system based on multimodal sensing according to claim 1, characterized in that, In the ammonia injection strategy generation layer, the intelligent optimization algorithm is a genetic algorithm or a reinforcement learning algorithm.

7. The method for precise ammonia injection control in an SCR denitrification system based on multimodal sensing according to claim 1, characterized in that, In the composite control execution layer, the short-cycle predictive feedforward control executes ammonia injection by sending instructions to the electric ammonia valve or the pneumatic ammonia valve.

8. The method for precise ammonia injection control in an SCR denitrification system based on multimodal sensing according to claim 1, characterized in that, In the composite control execution layer, the online monitoring system for NOx and NH3 concentrations in the SCR reactor outlet flue gas, upon which the fast-cycle fine-tuning feedback control is based, is a high-resolution online monitoring array. The array employs a tunable diode laser absorption spectroscopy scanning system, a chemiluminescence gas analyzer array, or an ultraviolet differential absorption spectroscopy gas analyzer array.

9. The method for precise ammonia injection control in an SCR denitrification system based on multimodal sensing according to claim 1, characterized in that, In the composite control execution layer, the fast-cycle fine-tuning feedback control uses a multivariable PID controller or a model predictive controller to correct the short-cycle predictive feedforward control command.

10. The method for precise ammonia injection control in an SCR denitrification system based on multimodal sensing according to claim 1, characterized in that, In the composite control execution layer, the long-cycle adaptive control is achieved by periodically retraining and calibrating the deep neural network model offline or online to adapt to slowly changing processes such as catalyst activity decay, coking in the furnace, and sensor drift.