Combustion control method and system for ammonia-coal mixed combustion boiler

Abstract

The invention discloses a combustion control method and system for an ammonia-coal co-combustion boiler, and relates to the technical field of thermal power generation, and the method comprises the following steps: carrying out feature extraction on a plurality of operation parameters collected in real time to obtain a plurality of features; a reinforcement learning model is constructed, a state space of the reinforcement learning model comprises a stability score, a NOx generation concentration predicted value and various characteristics, an action space is combustion control parameters of the ammonia-coal mixed combustion boiler, and a reward function is constructed through multiple optimization objectives of maximizing combustion efficiency and minimizing NOx emission and ammonia escape; and obtaining an optimal control strategy through the reinforcement learning model, and controlling the combustion control parameters through the optimal control strategy. According to the method, the states of different control actions under complex non-linear and time-varying characteristics are explored through the reinforcement learning model, and finally, an optimal control strategy capable of dynamically adapting to working condition changes is autonomously found and converged.

Description

Technical Field

[0001] This invention relates to the field of thermal power generation technology, and in particular to a combustion control method and system for an ammonia-coal co-fired boiler. Background Technology

[0002] With the increasingly severe global climate change problem, the "dual carbon" target has become an important direction for the development of the global energy industry. Coal-fired power plants, as one of the main sources of carbon emissions, face increasingly stringent carbon and pollutant emission limits. Ammonia-coal co-fired technology, as a promising low-carbon combustion technology, effectively reduces carbon dioxide emissions by mixing ammonia with pulverized coal during combustion, and has become a current research hotspot.

[0003] In existing technologies, the control of the ammonia-coal co-combustion process in boilers mostly adopts traditional fixed parameter adjustment methods. Traditional control methods cannot effectively cope with the nonlinear and time-varying characteristics of the ammonia-coal co-combustion process, resulting in combustion fluctuations and unstable pollutant emissions, making it difficult to accurately control the control parameters. Summary of the Invention

[0004] In view of the defects of the prior art, the present invention provides a combustion control method and system for an ammonia-coal co-fired boiler, which solves the existing problems.

[0005] The present invention adopts the following technical solution: In a first aspect, the present invention provides a combustion control method for an ammonia-coal co-fired boiler, comprising the following steps: Feature extraction was performed on various operating parameters of the ammonia-coal co-fired boiler collected in real time, resulting in a variety of features. These operating parameters included temperature field, pressure field, flame image, and various gas concentration values. Multiple features are input into a pre-trained scoring network to obtain the corresponding stability score; features corresponding to the temperature field, pressure field, and flame image are input into a pre-trained graph neural network to obtain the NO. x Generate concentration prediction values; Construct a reinforcement learning model whose state space includes a stability score and a NO. x The system generates predicted concentration values ​​and various features. The action space consists of the combustion control parameters of an ammonia-coal co-fired boiler. The reward function maximizes combustion efficiency and minimizes NO. x Construction of multiple optimization objectives for emissions and ammonia slip; The optimal control strategy is obtained through a reinforcement learning model, and the combustion control parameters are controlled by the optimal control strategy.

[0006] Preferably, the reward function is as follows: ; in, RFor the reward function, s t for t The state at any given moment, a t for t Actions at any moment = Combustion efficiency / Maximum efficiency (positive bonus) =−NO x Concentration / emission standards (negative incentives). =−Ammonia escape amount / Threshold (negative reward) , and These are the weighting coefficients.

[0007] Preferably, the multiple features include temperature features, pressure features, image features, and concentration features; temperature features include temperature extremes and temperature distribution trends, pressure features include pressure extremes and pressure distribution trends, image features include flame morphology, combustion state, and defect areas, and concentration features include concentration extremes and concentration distribution trends.

[0008] Preferably, the scoring network is based on a 3D-CNN+LSTM architecture.

[0009] Preferably, the combustion control parameters include the ammonia-to-coal ratio, the excess air coefficient, and the burner air distribution.

[0010] Preferably, combustion control parameters are executed by an actuator, which includes a fuel supply regulation system, an air distribution system regulation device, and a burner parameter regulation mechanism.

[0011] Preferably, a multi-scale feature extraction module is used to extract features from various operating parameters of the ammonia-coal co-fired boiler collected in real time. The multi-scale feature extraction module adopts a CNN and Transformer fusion architecture.

[0012] Preferably, before performing feature extraction on the various operating parameters of the ammonia-coal co-fired boiler collected in real time, filtering, abnormal data detection, and repair preprocessing are performed on the various operating parameters.

[0013] Preferably, real-time acquisition of various operating parameters includes: Temperature measurement points are arranged at different heights and cross sections of the furnace to form a three-dimensional temperature field monitoring network, and the temperature field is collected through the three-dimensional temperature field monitoring network. Pressure transmitters are installed at key locations in the furnace, air duct, and flue to collect pressure data. The concentrations of various gases, including O2, CO, CO2, and NO, were collected using tunable semiconductor laser absorption spectroscopy. x and NH3; Images of the flames were captured using a high-speed camera.

[0014] Secondly, the present invention provides a combustion control system for an ammonia-coal co-fired boiler, comprising: The extraction module is used to extract features from various operating parameters of the ammonia-coal co-fired boiler collected in real time, and obtain various features; the various operating parameters include temperature field, pressure field, flame image and various gas concentration values; The input module is used to input multiple features into a pre-trained scoring network to obtain the corresponding stability score; features corresponding to the temperature field, pressure field, and flame image are input into a pre-trained graph neural network to obtain the NO. x Generate concentration prediction values; The building blocks are used to construct reinforcement learning models whose state space includes stability scores and NO. x The system generates predicted concentration values ​​and various features. The action space consists of the combustion control parameters of an ammonia-coal co-fired boiler. The reward function maximizes combustion efficiency and minimizes NO. x Construction of multiple optimization objectives for emissions and ammonia slip; The control module is used to obtain the optimal control strategy through a reinforcement learning model, and then control the combustion control parameters using the optimal control strategy.

[0015] Compared with the prior art, the above-mentioned at least one technical solution adopted by the present invention can achieve the following beneficial effects: This invention will characterize the stability score and NO emission of combustion fluctuations and pollutants. x The system generates predicted concentration values ​​and extracts various features based on boiler operating parameters as the state space, and uses the combustion control parameters of the ammonia-coal co-fired boiler as the action space. Through a reinforcement learning model, it explores the states of different control actions under complex nonlinear and time-varying characteristics, ultimately autonomously discovering and converging to an optimal control strategy that can dynamically adapt to changes in operating conditions. By dynamically optimizing combustion control parameters, the combustion process always approaches optimal operating conditions, fundamentally solving the problems of fluctuating combustion efficiency and unstable pollutant emissions, and improving the overall thermal efficiency of the boiler. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart of a combustion control method for an ammonia-coal co-fired boiler according to the present invention; Figure 2This is a schematic diagram of a combustion control method for an ammonia-coal co-fired boiler according to the present invention. Figure 3 This is a schematic diagram of the multi-scale feature extraction algorithm of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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 skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] This invention provides a combustion control method for an ammonia-coal co-fired boiler, referring to... Figure 1 and Figure 2 Specifically, it includes the following steps: S1: Collects various operating parameters during the operation of the ammonia-coal co-fired boiler.

[0020] Multiple operating parameters are monitored, including temperature field, pressure field, flame image, and various gas concentration values. Temperature field, pressure field, flame image, and various gas concentration values ​​are acquired through a multi-source heterogeneous data acquisition layer. Specifically, the multi-source heterogeneous data acquisition layer includes a temperature field sensor array, a pressure field sensor network, a multi-component gas analyzer, and a flame image monitoring system.

[0021] Temperature field sensor array: Employing distributed fiber Bragg grating temperature measurement technology, temperature measurement points are arranged at different heights and cross-sections of the furnace to form a three-dimensional temperature field monitoring network. Each sensor node contains multiple gratings, enabling simultaneous measurement of temperature at multiple locations.

[0022] Pressure field sensor network: Piezoresistive pressure transmitters are installed at key locations in the furnace, air duct, and flue to monitor pressure changes in real time.

[0023] Multi-component gas analyzer: Based on tunable semiconductor laser absorption spectroscopy (TDLAS) technology, simultaneously measuring O2, CO, CO2, and NO. x It measures the concentrations of various gases, including NH3. An extraction sampling method, coupled with a high-temperature heat tracing pipeline, ensures measurement accuracy.

[0024] Flame Image Monitoring System: Employs a high-speed industrial camera equipped with a dedicated lens and filter to capture flame images in real time. The system features automatic focusing and anti-dust accumulation functions, and can operate stably for extended periods in harsh environments.

[0025] S2: Data preprocessing and feature extraction.

[0026] In this embodiment, data preprocessing includes an adaptive filtering module and an abnormal data detection and repair module.

[0027] Adaptive filtering module: This module employs a combination of wavelet transform and Kalman filtering for denoising temperature field data, pressure field data, gas concentration data, and flame image data. It automatically adjusts filtering parameters based on data characteristics to improve filtering effectiveness.

[0028] Anomaly detection and repair module: Employs the Isolation Forest algorithm to identify outlier data points, combined with time series interpolation to repair missing data. An anomaly knowledge base is established to classify and handle common anomalies.

[0029] Feature extraction is performed using a multi-scale feature extraction module: combining CNN and Transformer architectures, it extracts multi-scale features from the temperature field, pressure field, gas concentration, and flame image. CNN is used to extract local spatial features, while Transformer is used to capture temporal global features. Specifically, the temperature field outputs temperature features, including temperature extrema and temperature distribution trends; the pressure field outputs pressure features, including pressure extrema and pressure distribution trends; the flame image outputs image features, including flame morphology, combustion state, and defect areas; and the gas concentration outputs concentration features, including concentration extrema and concentration distribution trends.

[0030] It can be used for subsequent status monitoring, anomaly warning, or control decisions.

[0031] S3: Based on the results of S2, identify the combustion state and predict pollutant generation.

[0032] Combustion State Recognition Network: Based on a 3D-CNN+LSTM architecture, this network fuses multimodal features through an attention mechanism to achieve real-time assessment of combustion stability. Its input multimodal time-series data includes time-dimension image feature sequences and corresponding temperature, pressure, and concentration features.

[0033] The data processing workflow revolves around "multimodal data preprocessing, feature extraction, attention fusion, temporal modeling, and output scoring," ultimately achieving real-time evaluation. The network structure includes 3D convolutional layers, pooling layers, LSTM layers, and fully connected layers, outputting a stability score.

[0034] Pollutant formation prediction network: A graph neural network (GNN) is used to construct a graph structure containing chemical reaction pathways to predict NO formation. x The concentration is generated. The specific data processing flow includes: graph data preparation (defining the raw data into a standard graph format), feature initialization (graph data preprocessing to ensure the data meets the model input requirements), graph feature learning, and result output. The network input includes a time-dimensional image feature sequence and corresponding temperature, pressure, and chemical reaction relationships; the output is NO. x Generate concentration prediction values.

[0035] S4: Reinforcement learning is performed based on the results of S3.

[0036] Deep reinforcement learning decision-making module: based on the PPO algorithm, using combustion efficiency and NO... x Emissions and ammonia slip are multi-objective optimization functions, and control parameters are dynamically adjusted.

[0037] The combustion efficiency model describes the chemical reaction rate during pulverized coal combustion based on the Arrhenius equation, considering the effects of factors such as temperature, pressure, and reactant concentration on the reaction rate. The combustion efficiency model is as follows:

[0038] (1); in, r The reaction rate of pulverized coal mixed with ammonia during combustion. A Pre-exponential factor, E The activation energy is [Fuel], where [Fuel] is the coal powder concentration and [O2] is the oxygen concentration. R Let be the ideal gas constant. T The reaction temperature, a The first reaction order is... b This represents the second reaction order.

[0039] Fuel conversion rate x The reaction rate with ammonia-blended pulverized coal combustion r The relationship is: (2); in, nf 0 This is the initial fuel quantity. V This represents the reaction volume.

[0040] Combustion efficiency η For fuel conversion rate x The function of heat loss is: (3); in, Q loss To reduce heat loss due to heat dissipation and smoke extraction, Q theoretical The total heat of complete combustion of fuel, x This refers to the fuel conversion rate.

[0041] NO x The emission reduction model considers thermal NO x Based on the Zeldovich mechanism, this study describes the reaction of nitrogen and oxygen at high temperatures to produce NO. x The process of thermal NO xThe formation rate of NO is affected by the reaction temperature T, oxygen concentration [O2], and residence time t. x The empirical model is: (4); in: N fuel The mass fraction of nitrogen in the fuel is α, and the excess air coefficient is α. T The combustion temperature, t This represents the gas residence time.

[0042] NO x The emission reduction model is as follows: (5); Among them, [NO x ] is the generated NO x Concentration, [NO x ] 平衡 NO at this temperature x equilibrium concentration, k The reaction rate constant is... t This refers to the residence time of the gas in the high-temperature zone. This is the partial pressure of oxygen.

[0043] The ammonia escape suppression model is based on adsorption-desorption kinetics theory and describes the adsorption behavior of ammonia on the catalyst surface and its interaction with NO. x The reaction process and the escape pattern of unreacted ammonia gas were analyzed, taking into account temperature, catalyst activity, and the NH3 / NO3 ratio. x The influence of factors such as molar ratio on ammonia slip.

[0044] The ammonia escape inhibition model is as follows: (6); in, For the adsorption rate, [ NH [3] represents the NH3 concentration. For adsorption coverage, For parsing rate, k react For the reaction rate, [ NO x ]for NO x concentration, is the adsorption equilibrium constant.

[0045] The design incorporates a state space, action space, and reward function, continuously optimizing the control strategy through interaction with the environment. Based on the PPO algorithm, and considering combustion efficiency and NO... xThe optimization objectives, including emissions and ammonia slip, are defined by defining a state space (such as current combustion parameters and emission data), an action space (such as control parameters like ammonia-to-coal ratio, excess air coefficient, and burner air distribution), and a reward function (to measure the degree of achievement of the optimization objectives).

[0046] State space S: includes multidimensional state variables such as temperature features, pressure features, image features, and concentration features, stability score, and NO. x Generate concentration prediction values.

[0047] Action space A: includes control parameters such as ammonia-to-coal ratio, excess air coefficient, and burner air distribution.

[0048] Reward function R(s) t , a t Multi-objective weighted summation: (7); in: = Combustion efficiency / Maximum efficiency (positive bonus); =−NO x Concentration / emission standards (negative incentives); =−Ammonia escape amount / threshold (negative reward); + + =1 is the weighting coefficient.

[0049] Maximum efficiency is the efficiency value calculated based on rated operating conditions, while the threshold is set by combining three core factors: environmental protection standards, equipment corrosion risk, and economic cost.

[0050] A set of rules and methods for dynamically adjusting control parameters is developed through continuous interaction with the combustion environment, learning and iteratively updating to ultimately achieve multi-objective optimization.

[0051] S5: Knowledge Graph Reasoning System: Integrates combustion theory, expert experience, and historical cases to construct a knowledge graph. It uses an ontology model to represent domain knowledge and combines rule-based reasoning with machine learning reasoning to provide interpretable support for control decisions.

[0052] The modules operate through a collaborative mechanism, which includes: Data flow: 3D-CNN+LSTM outputs the burning status, GNN predicts NO. x The PPO adjusts control parameters based on the state and predictions, and the knowledge graph provides rule support and explanation.

[0053] Optimized closed loop: Reinforcement learning decision-making influences the combustion process by controlling parameters, and sensors feed back new data to the recognition network and prediction network, forming a "perception-reasoning-decision-execution" closed loop.

[0054] The reinforcement learning PPO algorithm learns the optimal policy through interaction with the environment and outputs a set of optimizations for multiple objectives (combustion efficiency, NO). x The adjustment amount of the core control parameters (emissions, ammonia escape) corresponds to the adjustable variables defined in the action space.

[0055] Fuel supply regulation system: This includes frequency converter control for the pulverized coal feeder, ammonia flow regulating valve control, and conveying system control. High-precision sensors and actuators are used to achieve precise control of the fuel supply.

[0056] The air distribution system regulating device includes secondary air damper opening control, burnout air ratio adjustment, and air box pressure control. Through a PID controller and a variable frequency fan, the air distribution system achieves rapid response and precise regulation.

[0057] Burner parameter adjustment mechanism: includes swirl intensity adjustment, primary air velocity adjustment, and burner tilt angle adjustment. It employs an electric actuator to achieve remote control and automatic adjustment of burner parameters.

[0058] Closed-loop feedback control system: Based on a DCS system, it enables rapid execution and feedback of control commands. A data interface is established between the control system and the AI ​​decision-making system to ensure accurate transmission of control commands and real-time feedback of execution results.

[0059] Multi-source data fusion and feature extraction.

[0060] Time synchronization algorithm: For sensor data with different sampling frequencies, a combination of linear interpolation and spline interpolation is used for time synchronization. A data timestamp index is established to ensure that data from the same time point can be accurately aligned.

[0061] Pollutant formation prediction and optimized control.

[0062] Physically Constrained Neural Networks: These networks embed chemical reaction kinetics into a neural network model to predict NOx formation under physical constraints. During model training, physical laws are incorporated to ensure that the prediction results conform to fundamental principles of combustion science.

[0063] Multi-objective optimization function: A multi-objective optimization function is designed to balance combustion efficiency, NOx emissions, and ammonia slip. A weighted summation method is used to transform the multiple objectives into a single-objective optimization problem, and the balance between the different objectives is achieved by adjusting the weighting coefficients.

[0064] Reinforcement learning control: The PPO algorithm is used to optimize control parameters, and the optimal control strategy is continuously learned through interaction with the environment. A reasonable reward function is designed to guide the agent to make decisions in a direction that improves combustion efficiency and reduces pollutant emissions.

[0065] System self-evolution mechanism.

[0066] Continuous Learning Framework: A continuous learning framework is designed to periodically collect new data to update the model. A combination of incremental learning and online learning methods is employed to continuously improve model performance while ensuring model stability.

[0067] Performance evaluation metrics: Develop a performance evaluation metric system for the algorithm, including multiple metrics such as prediction accuracy, control precision, and system stability. By monitoring these metrics in real time, the system automatically determines whether the model needs to be updated.

[0068] Fault Diagnosis and Self-Healing: Construct a fault diagnosis and self-healing system to monitor the system's operating status in real time. When a fault is detected, automatically locate and diagnose the fault, and take corresponding repair measures to ensure reliable system operation.

[0069] Example 1 System hardware deployment.

[0070] 1. Sensor Layout Scheme Temperature field sensor array: Utilizing distributed fiber Bragg grating temperature measurement technology, temperature measurement points are arranged at different heights and cross-sections within the furnace. A temperature measurement point is placed at 0.5m intervals in the burner area, forming a three-dimensional temperature field monitoring network. Each sensor node contains 5 gratings, enabling simultaneous measurement of the temperature at 5 locations, with a measurement range of 0~1600℃ and an accuracy of ±1℃.

[0071] Pressure field sensor network: Piezoresistive pressure transmitters are deployed at key locations in the furnace, air ducts, and flue. One pressure sensor is placed at each of the four corners of the furnace, pressure sensors are placed in the primary and secondary air ducts, and pressure sensors are placed in the flue, forming a pressure monitoring network. The sensor measurement range is -10 kPa to 10 kPa, and the response time is <10 ms.

[0072] Multi-component gas analyzer: Based on tunable semiconductor laser absorption spectroscopy (TDLAS) technology, a multi-component gas analyzer is installed in the flue. It employs an extraction sampling method with a sampling frequency set to 10Hz, simultaneously measuring the concentrations of multiple gases including O2, CO, CO2, NOx, and NH3, with an accuracy of ±0.01%. The analyzer is equipped with a high-temperature heating pipeline to ensure the sample gas temperature is above 180℃ to prevent condensation.

[0073] Flame Image Monitoring System: High-speed industrial cameras are installed at the observation ports, each aimed at a burner on a different floor. The cameras have a frame rate ≥100fps, a resolution ≥1920×1080 pixels, and are equipped with dedicated lenses and filters to capture flame images in real time. The system features automatic focusing and anti-dust accumulation functions, using compressed air to keep the lenses clean.

[0074] 2. Network architecture design.

[0075] Data transmission network: An industrial Ethernet network is used to construct the data transmission network, with a backbone network bandwidth of 1000Mbps and a branch network bandwidth of 100Mbps. Sensor data is transmitted to edge computing nodes via fieldbus, and the data processed by the edge computing nodes is transmitted to the core server via industrial Ethernet.

[0076] Edge computing nodes: Edge computing nodes are set up at the boiler site. These nodes are responsible for the preprocessing and feature extraction of sensor data, reducing the computing burden on the core server.

[0077] Core servers: The system employs a redundant design, is configured with high-performance servers, and uses a dual-machine hot standby mode to ensure system availability.

[0078] II. Algorithm Implementation Details 1. Multi-scale feature extraction algorithm (see...) Figure 3 ) CNN network architecture design: It adopts the ResNet architecture, containing 5 residual blocks, each consisting of multiple 3×3 convolutional layers. The input consists of temperature field, pressure field, and flame image data, and spatial features are extracted through convolutional kernels of different scales. The last layer of the network is a global average pooling layer, which compresses the feature map into a one-dimensional vector.

[0079] The Transformer module design incorporates positional encoding and multi-head attention mechanisms to capture temporal features. Positional encoding uses sine and cosine functions to integrate positional information into the input features. The multi-head attention mechanism maps input features to multiple subspaces and computes attention weights in parallel, enhancing the model's ability to capture different features.

[0080] Feature fusion strategy: A weighted fusion method is adopted, adaptively adjusting the weights of each modality feature according to different operating conditions. The weights of each modality feature are learned through a fully connected layer, and then the features are summed using weighted summation. The specific formula is as follows:

[0081] F = w1F1 + w2F2 + w3F3; Where F1, F2, and F3 are the feature vectors of the temperature field, pressure field, and flame image, respectively, and w1, w2, and w3 are the corresponding weights, which are obtained by normalization using the softmax function.

[0082] The feature fusion vector F is applied in the following two aspects: (1) Combustion stability The temperature field (F1) reflects the uniformity of heat distribution in the combustion zone, the pressure field (F2) reflects the combustion intensity and pressure fluctuations, and the flame image (F3) contains intuitive combustion features such as flame shape and brightness. The fused feature vector F can integrate the information from these three sources to more accurately determine whether combustion is stable (e.g., if F1 shows local high temperature, F2 shows large pressure fluctuations, and F3 shows distorted flame shape, the weights after fusion will emphasize abnormal features to quickly identify an "unstable combustion" state).

[0083] (2) Combustion efficiency The uniformity of the temperature field (F1) is directly related to combustion efficiency, while the energy conversion efficiency of the pressure field (F2) and the combustion completeness of the flame image (such as flame color and transparency, F3) are auxiliary indicators. The fusion feature F can quantify the combustion efficiency level (such as "high efficiency", "low efficiency", "uneven combustion"), providing a basis for real-time adjustment (e.g., by observing changes in weights, an increase in F1 weight corresponding to local low temperature indicates the need to optimize the fuel injection position).

[0084] III. Deep Reinforcement Learning Decision Algorithm.

[0085] State-space design: includes multi-dimensional state variables such as temperature, pressure, gas composition, and flame characteristics.

[0086] Action space design: including control parameters such as ammonia-coal ratio, excess air coefficient, and burner air distribution, specifically as follows: ammonia-coal ratio (0~30%, step size 0.1%), excess air coefficient (1.0~1.5, step size 0.01%), secondary air damper opening (0~100%, step size 1%), and burnout air ratio (0~30%, step size 1%).

[0087] Reward function R(s) t , a t Taking into account multiple objectives such as combustion efficiency, NOx emissions, and ammonia slip, a weighted summation of these objectives is used, as shown in the following formula: ; in: = Combustion efficiency / Maximum efficiency (positive bonus); =−NOx concentration / emission standard (negative incentive); =−Ammonia escape amount / threshold (negative reward); w1+w2+w3=1 is the weighting coefficient.

[0088] IV. Knowledge Graph Construction Methods Knowledge extraction: Knowledge is extracted from combustion science literature, expert experience, and historical cases. Natural language processing techniques are used to analyze the text of documents and reports, extracting entities, relationships, and attributes. For example, causal relationships such as "increased temperature leads to increased NOx formation" can be extracted from literature.

[0089] Knowledge Representation: Domain knowledge is represented using an ontology model, establishing semantic relationships between concepts. Classes, attributes, and relationships are defined to construct the architecture of the knowledge graph. For example, a "burner" class is defined, with attributes such as "type" and "location," and relationships such as "consumption" and "mixing" exist with classes such as "fuel" and "air."

[0090] Knowledge Reasoning: This approach combines rule-based reasoning and machine learning reasoning to apply knowledge. Rule-based reasoning uses production rules, such as "If the temperature > 1300℃ and the excess air coefficient > 1.2, then NOx generation increases." Machine learning reasoning uses graph neural networks to learn implicit relationships in a knowledge graph for prediction and reasoning.

[0091] V. System Integration and Testing 1. Software platform development Microservice architecture design: The system adopts the Spring Cloud microservice framework, dividing it into multiple microservices such as data acquisition, data processing, model training, decision-making, and execution. Each microservice is developed, deployed, and scaled independently, improving the system's maintainability and reliability.

[0092] Data visualization interface: Develop a data visualization interface to display the system's operating status and control effects in real time. The interface includes modules such as real-time data monitoring, historical data query, combustion status assessment, and control parameter adjustment, supporting multi-dimensional data display and analysis.

[0093] User Access Control System: Design a user access control system that divides users into three roles: administrator, operator, and observer. Administrators have the highest privileges and can configure the system and manage users; operators can adjust control parameters and log operations; observers can only view the system's running status and cannot perform any operations.

[0094] 2. System Testing Plan Unit testing: Independently test each functional module to ensure proper functioning. For example, test whether the data acquisition module can accurately acquire sensor data, and test whether the data processing module can correctly perform data preprocessing and feature extraction.

[0095] Integration testing: Verifies the data interaction and collaborative capabilities between modules. For example, it tests whether data transmission between the data acquisition module and the data processing module is normal, and whether instruction transmission between the decision-making module and the execution module is accurate.

[0096] On-site testing: Tests are conducted on an actual boiler to evaluate the system's performance under real-world operating conditions. Tests include assessing system stability, control accuracy, and response speed. During testing, the load and coal type are gradually adjusted to verify the system's adaptability to different operating conditions.

[0097] Example 2 Taking a 300MW coal-fired boiler renovation project at a power plant as an example, the implementation process and effects of this invention are explained in detail: 1. System deployment and configuration.

[0098] According to the above sensor arrangement scheme, a temperature field sensor array, a pressure field sensor network, a multi-component gas analyzer, and a flame image monitoring system are installed on the boiler.

[0099] Build a network architecture, configure edge computing nodes and core servers to ensure stable system operation.

[0100] Develop and deploy the software platform, and configure user permissions and system parameters.

[0101] 2. Model training and optimization.

[0102] Six months of historical operating data were collected, covering different loads, different coal types, and different operating conditions, to construct a sample dataset.

[0103] The dataset is cleaned and labeled, and divided into training, validation and test sets.

[0104] We employed a transfer learning approach, pre-training a combustion state recognition model using publicly available combustion datasets, and then fine-tuning it on the dataset from this project.

[0105] Reinforcement learning is used to optimize the control strategy. The training period is set to 10,000 episodes, and each episode simulates the control process under different operating conditions.

[0106] 3. System operation and effect evaluation.

[0107] After the system is officially put into operation, it collects sensor data in real time, performs preprocessing and feature extraction through edge computing nodes, and then transmits it to the core server for decision analysis.

[0108] The core server generates a control command every 2 seconds and sends it to each actuator through the DCS system to achieve real-time control of the boiler combustion process.

[0109] Establish an operation log recording system to record the system's operating status and control effectiveness in real time. Evaluate the system weekly and adjust model parameters and control strategies based on the evaluation results.

[0110] Based on the same concept, the present invention also provides a combustion control system for an ammonia-coal co-fired boiler, including an extraction module, an input module, a construction module and a control module.

[0111] The extraction module is used to extract features from various operating parameters of the ammonia-coal co-fired boiler collected in real time, and obtain various features; the various operating parameters include temperature field, pressure field, flame image and various gas concentration values.

[0112] The input module is used to input multiple features into a pre-trained scoring network to obtain the corresponding stability score; features corresponding to the temperature field, pressure field, and flame image are input into a pre-trained graph neural network to obtain the NO. x Generate concentration prediction values.

[0113] The building blocks are used to construct reinforcement learning models whose state space includes stability scores and NO. x The system generates predicted concentration values ​​and various features. The action space consists of the combustion control parameters of an ammonia-coal co-fired boiler. The reward function maximizes combustion efficiency and minimizes NO. x Construction of multiple optimization objectives for emissions and ammonia escape.

[0114] The control module is used to obtain the optimal control strategy through a reinforcement learning model, and then control the combustion control parameters using the optimal control strategy.

[0115] 1. Intelligent multimodal information fusion and intelligent decision control optimize the parameters of each module in real time according to the system's operating status, breaking through the limitations of traditional control in nonlinear, multi-objective, and strongly coupled scenarios, improving fuel combustion efficiency, reducing operating costs, and reducing nitrogen oxide emissions.

[0116] 2. The ammonia-coal co-firing system reduces dependence on a single fossil fuel, improves the stability and sustainability of energy supply, and its rational utilization helps to promote the transformation of the energy structure towards green and low-carbon.

[0117] 3. The system can flexibly adjust the ammonia-coal mixing ratio according to different energy demands, environmental conditions and policy requirements, and has strong adaptability and flexibility, making it suitable for the low-carbon retrofit of thermal power units of different scales and types.

[0118] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0119] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A combustion control method for an ammonia-coal co-fired boiler, characterized in that, Includes the following steps: Feature extraction was performed on various operating parameters of the ammonia-coal co-fired boiler collected in real time, resulting in a variety of features. These operating parameters included temperature field, pressure field, flame image, and various gas concentration values. Multiple features are input into a pre-trained scoring network to obtain the corresponding stability score; features corresponding to the temperature field, pressure field, and flame image are input into a pre-trained graph neural network to obtain the NO. x Generate concentration prediction values; Construct a reinforcement learning model whose state space includes a stability score and a NO. x The system generates predicted concentration values ​​and various features. The action space consists of the combustion control parameters of an ammonia-coal co-fired boiler. The reward function maximizes combustion efficiency and minimizes NO. x Construction of multiple optimization objectives for emissions and ammonia slip; The optimal control strategy is obtained through a reinforcement learning model, and the combustion control parameters are controlled by the optimal control strategy.

2. The combustion control method for an ammonia-coal co-fired boiler as described in claim 1, characterized in that, The reward function is as follows: ； in, R For the reward function, s t for t The state at any given moment, a t for t Actions at any moment = Combustion efficiency / Maximum efficiency (positive bonus) =−NO x Concentration / emission standards (negative incentives). =−Ammonia escape amount / Threshold (negative reward) , and These are the weighting coefficients.

3. The combustion control method for an ammonia-coal co-fired boiler as described in claim 1, characterized in that, Multiple features include temperature features, pressure features, image features, and concentration features; Temperature features include extreme temperature values ​​and temperature distribution trends; pressure features include extreme pressure values ​​and pressure distribution trends; image features include flame morphology, combustion state, and defect areas; and concentration features include extreme concentration values ​​and concentration distribution trends.

4. The combustion control method for an ammonia-coal co-fired boiler as described in claim 1, characterized in that, The scoring network is based on a 3D-CNN+LSTM architecture.

5. The combustion control method for an ammonia-coal co-fired boiler as described in claim 1, characterized in that, The combustion control parameters include the ammonia-to-coal ratio, excess air coefficient, and burner air distribution.

6. The combustion control method for an ammonia-coal co-fired boiler as described in claim 1, characterized in that, Combustion control parameters are executed by actuators, which include a fuel supply regulation system, an air distribution system regulation device, and a burner parameter regulation mechanism.

7. The combustion control method for an ammonia-coal co-fired boiler as described in claim 1, characterized in that, The multi-scale feature extraction module extracts features from various operating parameters of the ammonia-coal co-fired boiler collected in real time. The multi-scale feature extraction module adopts a CNN and Transformer fusion architecture.

8. The combustion control method for an ammonia-coal co-fired boiler as described in claim 1, characterized in that, Before extracting features from various operating parameters of the ammonia-coal co-fired boiler collected in real time, filtering, abnormal data detection, and repair preprocessing are performed on various operating parameters.

9. The combustion control method for an ammonia-coal co-fired boiler as described in claim 1, characterized in that, Real-time acquisition of various operating parameters includes: Temperature measurement points are arranged at different heights and cross sections of the furnace to form a three-dimensional temperature field monitoring network, and the temperature field is collected through the three-dimensional temperature field monitoring network. Pressure transmitters are installed at key locations in the furnace, air duct, and flue to collect pressure data. The concentrations of various gases, including O2, CO, CO2, and NO, were collected using tunable semiconductor laser absorption spectroscopy. x and NH3; Images of the flames were captured using a high-speed camera.

10. A combustion control system for an ammonia-coal co-fired boiler, characterized in that, include: The extraction module is used to extract features from various operating parameters of the ammonia-coal co-fired boiler collected in real time, and obtain various features; Multiple operating parameters include temperature field, pressure field, flame image, and various gas concentration values; The input module is used to input multiple features into a pre-trained scoring network to obtain the corresponding stability score; features corresponding to the temperature field, pressure field, and flame image are input into a pre-trained graph neural network to obtain the NO. x Generate concentration prediction values; The building blocks are used to construct reinforcement learning models whose state space includes stability scores and NO. x The system generates predicted concentration values ​​and various features. The action space consists of the combustion control parameters of an ammonia-coal co-fired boiler. The reward function maximizes combustion efficiency and minimizes NO. x Construction of multiple optimization objectives for emissions and ammonia slip; The control module is used to obtain the optimal control strategy through a reinforcement learning model, and then control the combustion control parameters using the optimal control strategy.

Images