A low-nitrogen combustion process modeling and emission control method based on deep learning

Abstract

The application provides a low-nitrogen combustion process modeling and emission control method based on deep learning, and belongs to the technical field of combustion control and intelligent optimization of industrial processes based on deep learning; through multi-source industrial perception, heterogeneous data acquisition of the low-nitrogen combustion process and standardized state construction are realized, and standard combustion state samples are formed. Based on the coupling architecture of deep learning and mechanism constraint, a low-nitrogen combustion joint modeling network is built, the heterogeneous perception data is mapped into unified scene features by using a multi-modal feature extraction network, and the core combustion indicators and confidence scores are synchronously output by using a multi-task learning module. A combustion mechanism consistency constraint training strategy is adopted, a self-supervised feedback closed loop is formed by embedding a physical boundary constraint, and it is ensured that the prediction result conforms to the physical law. Finally, a control strategy is generated based on the confidence gating and reinforcement learning, and after safety verification, the control strategy is issued to a distributed control system, so that the pollutant emission is reduced, and the safety and self-evolution intelligent level of the combustion system are improved.

Description

Technical Field

[0001] This invention belongs to the field of combustion control and intelligent optimization technology of industrial processes based on deep learning, and particularly relates to a method for modeling and controlling low-NOx combustion processes and emissions based on deep learning. Background Technology

[0002] In coal-fired boilers, pulverized coal boilers, industrial kilns, and industrial combustion devices with staged air supply and fuel blending capabilities, low-NOx combustion process control is a crucial link in achieving pollutant emission reduction, improved combustion efficiency, and operational safety. The macroscopic control objectives for this type of process are mainly reflected in multiple aspects, including reducing NOx emission concentrations, improving combustion efficiency, decreasing fly ash carbon content, narrowing furnace thermal deviations, and suppressing the risk of abnormal carbon monoxide accumulation. These control objectives are not determined by a single measuring point or a single control variable, but are fundamentally influenced by the complex coupling relationships between furnace temperature distribution, local oxygen levels, fuel volatile matter release rhythm, primary air ratio, secondary air distribution method, staged air input intensity, burner swivel angle, boiler load changes, and flue gas emission feedback.

[0003] However, in actual industrial applications, the modeling and control of low-NOx combustion systems still largely rely on empirical rules, static operating condition curves, and distributed univariate loops. Operators typically manually adjust primary air, secondary air, staged air, and fuel distribution based on load ranges, empirical air distribution tables, and historical emission averages. They then rely on outlet flue gas measurement results to observe for exceedances and make readjustments. In this type of control mode, the control strategy mainly revolves around result feedback, making it difficult to prospectively model significant time delays, strong couplings, and nonlinear disturbances present in the combustion process. This results in the inaccurate capture of the dynamic correlation between pollutant generation processes and the combustion steady state.

[0004] Existing technologies exhibit several shortcomings in complex low-NOx combustion scenarios. First, boiler combustion processes are characterized by high inertia, significant hysteresis, and multivariate coupling. A nonlinear mapping exists between the formation of a local reducing atmosphere within the furnace, changes in burnout, and the amount of NOx generated. Traditional linear modeling methods struggle to effectively represent these coupled behaviors within a unified framework. Second, actual operational data comes from a wide range of sources, including furnace flame images, temperature measurement points, online flue gas analyzers, air-coal ratio signals, pulverizer operating status, and load commands. These different data sources commonly suffer from inconsistent time sampling frequencies, different signal dimensions, local missing data, and noise pollution, making it difficult to generate high-quality, unified state samples directly usable for modeling without thorough cleaning and alignment. Third, existing control methods typically focus on optimizing a single NOx index, lacking simultaneous constraints on side effects such as increased carbon monoxide, furnace slagging risk, combustion skew, and flue gas heat loss. This can easily lead to situations where combustion safety and economy are sacrificed in pursuit of NOx reduction.

[0005] While some data-driven methods have attempted to build emission prediction models using neural networks or time-series models, most remain at the level of single-output soft measurement, primarily used to predict outlet nitrogen oxide concentrations. They have not yet incorporated combustion stability, burnout status, localized oxygen deficiency risk, and controllable domain constraints into a unified framework. These methods are prone to prediction drift under conditions of operating deviation, coal quality fluctuations, or sensor anomalies. Furthermore, due to the lack of confidence level identification and safety boundary projection mechanisms, they struggle to directly output control adjustment commands that can be executed in industrial settings. Simultaneously, the lack of effective coupling between traditional control system rules based on human experience, combustion mechanism knowledge, and data-driven models results in models that, while possessing some fitting ability, are unable to meet the requirements for long-term online stable operation in complex industrial environments. Summary of the Invention

[0006] To address the above problems, this invention proposes a deep learning-based method for modeling and controlling low-NOx combustion processes and emissions, comprising the following steps: S1 collects flame image data, process variables and mechanism operation data, flue gas emission data, fuel data and operating condition data during the low-NOx combustion process, and preprocesses them to obtain flame image features, process variable features, flue gas and emission features, fuel features and operating condition context features. S2, construct a low-NOx combustion joint modeling network based on a deep learning and mechanism constraint coupling architecture, use a multimodal coding combustion state feature extraction network to map the feature data obtained in S1 into unified combustion scenario features, and output prediction vectors and confidence scores of multiple core combustion state indicators through a multi-task learning-based emission and combustion state joint prediction module. S3, based on the physical limits of each actuator and the single-step change rate limit, obtains the controllable feasible region; the features obtained in S1 and the prediction vector output in S2 are used together as the state input to the reinforcement learning policy network, and under the constraint of the controllable feasible region, the initial control quantity is output by maximizing the expected cumulative reward. A confidence-based gating security protection mechanism is implemented, which uses the confidence score output by S2 to intercept and reduce the intensity of the initial control quantity before execution, and outputs the final execution control quantity.

[0007] Preferably, the specific process for constructing the flame image features is as follows: First, acquire furnace flame image data of the flame area. And preprocessed flame image samples were obtained. ; Secondly, construct derived visual feature vectors. For preprocessed flame image samples Extracting flame brightness distribution Main combustion zone area ratio Flame center offset undulation of the flame edge and color distribution statistics They are then uniformly concatenated and encapsulated into derived visual feature vectors. ; Subsequently, a short temporal visual sample sequence was constructed. First, for the preprocessed flame image samples from multiple time points... Constructing a time sliding window from the continuous sequence formed The time sliding window The length is set to , the current moment and its predecessor Preprocessed flame image samples at each time point A short-temporal visual sample sequence is obtained by concatenating multidimensional tensors in chronological order. ; Finally, the short temporal visual sample sequence and its corresponding derived visual feature vectors Flame image features are obtained through normalization and feature concatenation operations. .

[0008] Preferably, the process for constructing the process variable characteristics is as follows: The primary air volume is read in real time from the distributed control system, field instrumentation system, and actuator feedback unit. Secondary air volume Total air volume air supply temperature Temperature of each layer of the furnace Flue temperature Oxygen measurement data Perform resampling, anomaly removal, and uniform time step slicing, with a fixed time window. Organizing the data into time series segments yields a preprocessed sequence of process variables. ,right Perform normalization and feature concatenation operations to obtain process variable features. .

[0009] Preferably, the process for constructing the flue gas and emission characteristics, fuel characteristics, and operating condition context characteristics is as follows: Nitrogen oxide emission concentrations are obtained from online flue gas monitoring devices and on-site analysis instruments. carbon monoxide concentration Oxygen content Simultaneously read the carbon content of fly ash The above four data points are combined to form the characteristics of flue gas and emissions. ; Collect coal quality variables Including moisture and fuel nitrogen content ,right Fuel features are obtained by performing nonlinear numerical mapping and multidimensional vector concatenation. ; Collect operating condition variables Including boiler load instructions Combined operation mode of coal mill ,right Perform nonlinear numerical mapping and multidimensional vector concatenation to obtain working condition context features. .

[0010] Preferably, the feature data obtained in S1 is mapped to unified combustion scene features using a multimodal coding combustion state feature extraction network, specifically as follows: A flame image feature extraction network is constructed to extract and output image branch latent features based on flame image features; a process variable temporal coding network is constructed to extract and output process branch latent features based on process variable features; an emission state coding network is constructed to extract and output emission branch latent features based on flue gas and emission features; an operating condition semantic and component attribute coding network is constructed by concatenating fuel features and operating condition context features to extract and output operating condition branch latent features; then, the image branch latent features, process branch latent features, emission branch latent features, and operating condition branch latent features are concatenated along the channel dimension and fused with attention weighting to output unified combustion scene features.

[0011] Preferably, the flame image feature extraction network consists of four two-dimensional convolutional modules; the first convolutional layer has a kernel size of 3×3, 1 input channel, 32 output channels, a stride of 1, and padding of 1; the second convolutional layer has a kernel size of 3×3, 32 input channels, 64 output channels, a stride of 1, and padding of 1; the third convolutional layer has a kernel size of 3×3, 64 input channels, 128 output channels, a stride of 1, and padding of 1; the fourth convolutional layer has a kernel size of 3×3, 128 input channels, 256 output channels, a stride of 1, and padding of 1; each convolutional module is followed by a batch normalization layer, a ReLU activation layer, and a 2×2 max pooling layer in sequence, ultimately outputting an image feature map. Subsequently, a global average pooling layer and a fully connected layer with an output dimension of 128 are used to... Mapped to image branch latent features .

[0012] Preferably, the emission and combustion state joint prediction module inputs the unified combustion scenario features into a shared feature backbone network and multiple prediction heads, and simultaneously outputs a prediction vector containing predicted values ​​for nitrogen oxide emission concentration, carbon monoxide emission concentration, burnout state characterization, furnace thermal deviation characterization, and combustion stability margin characterization. Subsequently, it performs prediction confidence assessment and model confidence interval generation, quantifies the uncertainty of the model in the forward inference process through a Monte Carlo random deactivation mechanism, and outputs a confidence score corresponding to the prediction vector.

[0013] Preferably, the specific method for obtaining the controllable feasible domain in S3 is as follows: Define control space For primary air regulation volume Secondary air distribution adjustment amount Graded wind input adjustment amount Coal supply and distribution adjustment amount and burner tilt adjustment The column vectors formed by the primary air regulation amount control the primary air fan frequency converter to adjust the momentum of the air-coal mixture; the secondary air distribution regulation amount control the secondary air damper actuator in the main combustion zone to change the tiered combustion air ratio; the staged air input regulation amount control the burnout air damper to accurately construct the reducing denitrification atmosphere zone; the coal feeding distribution regulation amount control the coal feeder drive motor to increase or decrease the actual fuel feed amount; the burner tilt angle regulation amount control the burner tilt angle drive device to change the flame center height and furnace heat absorption rate. By extracting operating condition variables Boiler load instructions Combined operation mode with coal mill The initial physical limits of each actuator are determined by linear numerical mapping. and Simultaneously, calculate the flame center offset. With oxygen content The deviation intensity is determined, and the safety margin of the current operating condition is quantified by comparing it with the distribution of historical stable operating data. The single-step rate of change limit of the control action is then calculated. Constructing basic action boundaries as follows: ; in, The scaling factor is determined by offline cross-validation.

[0014] Preferably, the reinforcement learning policy network in S3 is used to predict vectors. The various predictive indicators in the definition of the comprehensive reward function as follows: ; in, This represents the predicted concentration of nitrogen oxide emissions. This represents the predicted carbon monoxide emission concentration, directly extracted from the prediction vector. ; This is a risk item calculated based on the furnace thermal deviation characterization value. For risk terms calculated based on burnout state characterization values, the weighting coefficients are... to and reward coefficient All were determined by offline multi-objective equilibrium experiments; Score the confidence level; By maximizing the expected cumulative reward, the policy network... Directly output preliminary control quantity under constraints The preliminary control quantity That is, a set of instantiated feature column vectors within the control space.

[0015] Preferably, the confidence-based gating security protection mechanism implemented in S3 specifically includes: High confidence threshold based on offline training set statistics Set to 0.85, low confidence threshold Set to 0.60; when the confidence score At that time, the preliminary control quantity was issued directly. ;when At that time, a scaling factor is introduced. Calculate the final execution control quantity ;when If the prediction fails, the policy network output is forcibly blocked, and the system switches to a conservative control template built based on expert experience.

[0016] Compared with the prior art, the present invention has the following innovative features and beneficial effects: (1) Design of joint modeling network architecture for low-NOx combustion based on multi-source heterogeneous perception and multi-task prediction: A joint modeling network architecture based on independent multimodal coding and shared backbone multi-task learning is proposed. The architecture first constructs a combustion state feature extraction network based on multimodal coding. For standardized low-NOx combustion state samples, it independently constructs flame image feature extraction networks, process variable temporal coding networks, emission state coding networks, and operating condition semantic and component attribute coding networks. It extracts the corresponding four-branch latent features and performs channel-dimensional concatenation and multi-head attention weighted fusion to form a unified combustion scene feature with strong physical correlation. Subsequently, it constructs a joint emission and combustion state prediction module based on multi-task learning. It abandons the traditional single-objective soft measurement island mode and inputs the unified combustion scene features into a shared feature backbone network and five sets of parallel prediction heads. It simultaneously outputs a prediction vector including NOx emission concentration, carbon monoxide emission concentration, burnout state, furnace thermal deviation, and combustion stability margin characterization values. At the same time, it introduces a confidence assessment module based on Monte Carlo random deactivation mechanism to extract the variance and mean ratio features in the forward inference process in real time to quantify uncertainty, thereby achieving comprehensive measurement and high-confidence perception of complex combustion states. (2) Embedded Combustion Mechanism Consistency Verification and Mechanism Data Co-constraint Optimization Strategy Design: An end-to-end mechanism constraint coupling training strategy based on embedded industrial physical laws is proposed to eliminate the defects of pure data-driven black boxes. During the model supervised training process, prior industrial combustion mechanism knowledge is directly extracted and transformed into a differentiable penalty term and incorporated into the network loss function. By constructing a combustion mechanism consistency constraint loss term that includes oxygen boundary consistency constraints, temperature upper limit constraints, component negative correlation constraints, and burnout coordination constraints, dynamic optimization is performed to coordinate the updating of supervised prediction error and physical boundary verification. Under this mechanism, the prediction bias generated by the joint modeling network that violates physical common sense is transformed into a gradient flow that can guide weight adjustment through backpropagation. This gradient flow is directly fed back to the multi-task prediction head and multi-modal feature encoding network to drive its directional correction of latent feature mapping logic, thereby completely avoiding the prediction divergence phenomenon that occurs in traditional data-driven models under boundary conditions or without data, and ensuring that the final stable output prediction vector accurately anchors the physical constraint boundary of real furnace combustion while meeting the visual-level fitting accuracy. (3) A closed-loop system for safe reinforcement learning execution based on confidence gating and feasible domain boundary mapping: In response to the core pain points of the difficulty in implementing deep learning control strategies in real industrial sites and the low fault tolerance rate, a closed-loop system for the entire link of safety bottom line support and physical mapping of adjustment commands is established. First, the feasible control domain is constructed and the operational boundary is mapped. The absolute physical limits and single-step change rate limits of each physical actuator are calculated using real-time operating condition context features to construct a hard feasible control domain for defining the boundary of control action exploration. Second, a confidence-gated safety protection mechanism is designed. The confidence score output by the joint modeling network is directly reused to perform hard interception of the preliminary control quantity generated by the near-end strategy optimization algorithm before execution. Based on the high confidence threshold and low confidence threshold, the mechanism of direct output of control quantity, the mechanism of proportional scaling and reduction of control quantity, or the mechanism of prediction blocking and forced switching to conservative control template are triggered respectively. Most importantly, the final execution control quantity after interception and regulation is aligned and bundled with the multi-dimensional prediction vector and confidence score in a high-dimensional data manner to form a complete closed-loop output result and directly sent to the underlying distributed control system to execute the physical mechanism intervention action. This realizes a complete reverse control closed loop from multi-source state perception, multi-dimensional dynamic prediction to action safety verification and then to the physical implementation of control commands. Attached Figure Description

[0017] To more clearly illustrate the technical solutions 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 following description is only one embodiment of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating the overall technical route of the present invention.

[0019] Figure 2 This is a flowchart of the multi-source heterogeneous data acquisition and standardized sample construction process of this invention.

[0020] Figure 3 This is a diagram of the joint modeling network architecture and mechanism constraint logic architecture of the present invention.

[0021] Figure 4 This is a flowchart illustrating the reinforcement learning strategy generation and confidence-gated decision-making process of this invention.

[0022] Figure 5 The figure shows the error comparison test results of different algorithms in predicting multiple objectives of low-NOx combustion in the examples.

[0023] Figure 6 The figure shows a comparison of the emission boundary distribution before and after the introduction of combustion mechanism constraints in the example.

[0024] Figure 7This is a correlation distribution diagram showing the relationship between prediction confidence and the security degradation interval of the control strategy in the example. Detailed Implementation

[0025] To address the control challenges posed by large time delays and multivariate coupling in complex industrial combustion processes and to achieve multi-objective optimization of low-NOx combustion, this invention proposes a deep learning-based method for low-NOx combustion process modeling and emission control. The overall process is as follows: Figure 1 As shown: By systematically acquiring heterogeneous data on low-NOx combustion processes based on multi-source industrial sensing and constructing standardized states, standardized low-NOx combustion state samples with aligned physical properties and unified data modalities are established. A joint low-NOx combustion modeling network is built based on a deep learning and mechanism-constrained coupling architecture. A multi-modal encoded combustion state feature extraction network accurately maps heterogeneous sensing data to unified combustion scenario features. A joint emission and combustion state prediction module based on multi-task learning simultaneously outputs multiple core combustion state indicators and quantified confidence scores. Furthermore, a constraint-coupled training strategy based on combustion mechanism consistency is implemented, reusing embedded physical boundary constraints to construct a multi-objective self-supervised feedback loop, ensuring that the prediction vectors generated by the network are accurately anchored to physical laws across multiple dimensions. Finally, a control strategy based on confidence gating and reinforcement learning is executed to generate and verify the safety baseline. Through control feasible domain mapping and single-step change rate limit extraction, the final control quantity is distributed to the underlying distributed control system, thereby significantly reducing pollutant emission concentrations and improving the safety and self-evolving intelligence level of the combustion control system.

[0026] The implementation process of the present invention will be further described below with reference to specific embodiments.

[0027] S1. Heterogeneous Data Acquisition and Standardized State Construction of Low-NOx Combustion Process This step aims to collect multi-source heterogeneous data from complex industrial environments and perform standardized preprocessing to establish a strict alignment relationship between cross-modal combustion process operating data, providing a uniformly formatted data input for subsequent low-NOx combustion co-modeling networks; the specific process is as follows: Figure 2As shown: First, flame image feature construction is performed. Furnace flame image data is acquired through image acquisition equipment. After preprocessing, feature extraction, and time-series stitching, flame image features are output. Second, process variable and actuator operation data acquisition is performed. Real-time data of air volume, temperature, and oxygen are read. After resampling and time-series slicing, process variable features are output. Third, flue gas emission, fuel, and operating condition data acquisition is performed. Corresponding process indicators are acquired and numerical mapping is performed to output flue gas and emission features, fuel features, and operating condition context features. Finally, heterogeneous data encapsulation is performed. The flame image features, process variable features, flue gas and emission features, fuel features, and operating condition context features acquired above are encapsulated using multi-dimensional tensor stitching to output standardized low-NOx combustion state samples. S1.1 Flame Image Feature Construction. First, this invention continuously acquires furnace flame image data of the flame area using image acquisition devices deployed at furnace observation ports, burner areas, and industrial endoscopic vision terminals. The original image was then subjected to denoising, distortion correction, brightness equalization, and image cropping to obtain preprocessed flame image samples. .

[0028] Secondly, construct derived visual feature vectors. For preprocessed flame image samples The system extracts the flame brightness distribution. Main combustion zone area ratio Flame center offset undulation of the flame edge and color distribution statistics They are then uniformly concatenated and encapsulated into derived visual feature vectors. Specifically, the present invention first addresses... The image is converted to grayscale, and the weighted average of the grayscale values ​​of all pixels is calculated to obtain the flame brightness distribution. This is used to characterize the overall ignition intensity within the furnace; the luminance threshold is determined using the maximum inter-class variance method and then... Perform binarization segmentation to identify the flame pixel region. Then, the area ratio of the main combustion zone was calculated. ; then calculate The geometric centroid coordinates are determined, and the Euclidean distance of the centroid relative to the image center point is calculated to obtain the flame center offset. Finally, utilize Operator extraction By analyzing the flame edge contour and calculating the perimeter-to-area ratio of edge pixels, the degree of undulation of the flame edge can be obtained. ;Will Switch to Color space, extraction Histogram features of hue components and The mean of the saturation component yields the color distribution statistics. .

[0029] Subsequently, a short temporal visual sample sequence was constructed. First, for the preprocessed flame image samples from multiple time points... Constructing a time sliding window from the continuous sequence formed The time sliding window The length is set to This invention will use the current moment... and its predecessor The preprocessed flame image samples at each time point A short-temporal visual sample sequence is obtained by concatenating multidimensional tensors in chronological order. .

[0030] Finally, the short temporal visual sample sequence and its corresponding derived visual feature vectors Flame image features are obtained through normalization and feature concatenation operations. .

[0031] S1.2 Process Variable and Actuator Operation Data Acquisition. This step is used to establish a dataset of process variables directly related to low-NOx combustion regulation. This invention reads primary air volume in real time from the distributed control system, field instrumentation system, and actuator feedback unit. Secondary air volume Total air volume air supply temperature Temperature of each layer of the furnace Flue temperature Oxygen measurement data Resampling, outlier removal, and uniform time step slicing were performed on the above variables, using a fixed time window. Organizing the data into time series segments yields a preprocessed sequence of process variables. ,right Perform normalization and feature concatenation operations to obtain process variable features. .

[0032] S1.3 Flue gas emission, fuel, and operating condition data acquisition. This step is used to construct labeled variables and auxiliary state variables. First, the nitrogen oxide emission concentration is obtained from the online flue gas monitoring device and field analysis instruments. carbon monoxide concentration Oxygen content Simultaneously read the carbon content of fly ash The above four data points are combined to form the characteristics of flue gas and emissions. .

[0033] Simultaneously, coal quality variables were collected. Including moisture and fuel nitrogen content ,right Fuel features are obtained by performing nonlinear numerical mapping and multidimensional vector concatenation. ; Finally, collect operating condition variables. Including boiler load instructions Combined operation mode of coal mill ,right Perform nonlinear numerical mapping and multidimensional vector concatenation to obtain working condition context features. .

[0034] S1.4 Heterogeneous data encapsulation. The preprocessed data includes... , , , , The samples were assembled and packaged into standardized low-NOx combustion state samples. .

[0035] S2. Construct a dynamic prediction model for low-NOx combustion based on the coupling of deep learning and mechanistic constraints. This step aims to establish a dynamic prediction model for complex combustion processes using standardized low-NOx combustion state samples, providing accurate predicted states and quantitative confidence metrics for subsequent control strategy generation; the specific process is as follows: Figure 3As shown: First, a combustion state feature extraction network based on multimodal coding is constructed. For the flame image features in the standardized low-NOx combustion state samples output by S1, a flame image feature extraction network is constructed to extract and output image branch latent features. For the process variable features in the standardized low-NOx combustion state samples, a process variable temporal coding network is constructed to extract and output process branch latent features. For the flue gas and emission features in the standardized low-NOx combustion state samples, an emission state coding network is constructed to extract and output emission branch latent features. For the fuel features and operating condition context features in the standardized low-NOx combustion state samples, an operating condition semantic and component attribute coding network is constructed by concatenating them to extract and output operating condition branch latent features. Then, the image is segmented... The hidden features of the branch, process branch, emission branch, and operating condition branch are concatenated and fused with attention weighting along the channel dimension to output unified combustion scenario features. Next, a joint emission and combustion state prediction module based on multi-task learning is constructed. The unified combustion scenario features are input into a shared feature backbone network and multiple prediction heads, simultaneously outputting prediction vectors containing predicted values ​​for nitrogen oxide emission concentration, carbon monoxide emission concentration, burnout state characterization, furnace thermal deviation characterization, and combustion stability margin characterization. Finally, prediction confidence assessment and model confidence interval generation are performed. The uncertainty of the model in the forward inference process is quantified through a Monte Carlo random deactivation mechanism, and a confidence score corresponding to the prediction vector is output. S2.1 Combustion State Feature Extraction Network Based on Multimodal Coding: This step is used to extract standardized low-NOx combustion state samples. Different types of information are mapped to a unified high-dimensional feature representation.

[0036] First, regarding the characteristics of flame images This invention constructs a flame image feature extraction network. The network consists of four two-dimensional convolutional layers. The first convolutional layer has a 3×3 kernel, 1 input channel, 32 output channels, a stride of 1, and padding of 1. The second convolutional layer has a 3×3 kernel, 32 input channels, 64 output channels, a stride of 1, and padding of 1. The third convolutional layer has a 3×3 kernel, 64 input channels, 128 output channels, a stride of 1, and padding of 1. The fourth convolutional layer has a 3×3 kernel, 128 input channels, 256 output channels, a stride of 1, and padding of 1. Each convolutional layer is followed by a batch normalization layer, a ReLU activation layer, and a 2×2 max pooling layer, ultimately outputting an image feature map. Subsequently, a global average pooling layer and a fully connected layer with an output dimension of 128 are used to... Mapped to image branch latent features .

[0037] Secondly, regarding time series data of process variables. A process variable temporal coding network was established, and dynamic dependency features in the air-coal ratio, temperature distribution, oxygen fluctuation, and load change were extracted by jointly using temporal convolutional units and gated recurrent units. The temporal convolutional unit uses a one-dimensional convolutional layer with 64 kernels, a kernel length of 5, a stride of 1, and padding of 2. It is followed by two layers of gated recurrent units, with each layer containing 128 hidden units, ultimately outputting the process branch hidden features. .

[0038] For flue gas and emission characteristic vectors This invention constructs an emission status coding network. The emission status coding network uses two fully connected layers to extract a deep representation of the current emission level. The first fully connected layer has a dimension of 64, and the second fully connected layer has a dimension of 64, outputting the latent features of the emission branches. .

[0039] For operating condition context data With fuel characteristics A semantic encoding network for operating conditions and component attributes is constructed. First, these two are concatenated into a context input vector. Then, two fully connected mapping layers are used to extract semantic representations of coal quality status, operating mode, and equipment boundary conditions. The first fully connected layer has an output dimension of 64, and the second fully connected layer also has an output dimension of 64. A ReLU activation function is used in between to output the latent features of the operating condition branches. .

[0040] After encoding each branch, the latent features from different branches are concatenated along the channel dimension and then weighted by attention to form a unified combustion scene feature. This serves as the basis for subsequent multi-task prediction and control decision support. The unified state characteristic is denoted as... The calculation expression is as follows: ; in, This indicates a feature fusion operation. Specifically, the feature fusion operation consists of a four-head multi-head attention layer and a fully connected mapping layer with an output dimension of 256, where each attention head has a feature dimension of 64.

[0041] S2.2 Joint Prediction of Emissions and Combustion States Based on Multi-Task Learning: This step is used to simultaneously output multiple state variables directly related to low-NOx combustion control. This invention sets up multiple prediction heads on the basis of a shared feature backbone network, each used to output predicted NOx emission concentration values. Predicted carbon monoxide emission concentration Burnout status characterization value Furnace thermal deviation characterization value and combustion stability margin characterization value Each prediction head is... The input consists of two fully connected layers. The first layer has an output dimension of 64, and the second layer has an output dimension of 1. A ReLU activation function is applied after the first layer. Let the... The model output prediction vector at time t is denoted as . Then its expression is as follows: ; in, This represents the predicted concentration of nitrogen oxide emissions. This represents the predicted carbon monoxide emission concentration. This represents the value indicating the burnout state. This represents the value characterizing the thermal deviation of the furnace. This represents the combustion stability margin characterization value. Through the above joint output method, the system can avoid the problem of a narrow control perspective caused by traditional single-index modeling, so that emission control no longer focuses only on nitrogen oxides, but also considers multiple state variables that have a direct impact on combustion quality and operational safety.

[0042] S2.3 Prediction Confidence Assessment and Model Confidence Interval Generation: This step is used to provide a confidence assessment for subsequent emission control. This invention enables a Monte Carlo random deactivation mechanism during the inference phase, and generates a confidence interval for the prediction vector at each time step. Generate the corresponding confidence score Specifically, the confidence evaluation head consists of two fully connected layers: the first layer has an output dimension of 32, the second layer has an output dimension of 1, and finally... The function maps the output to the interval between 0 and 1. During computation, the same input sample is processed... The algorithm performs a forward propagation, randomly shutting down some neurons with a probability of 0.2 in each propagation. This is achieved through calculation... The uncertainty of the prediction is quantified by the ratio of the variance to the mean of the output results, thus obtaining the confidence score. The confidence score Used to describe the stability and reliability of the current model's prediction results under a given operating condition.

[0043] S2.4 Constraint-Coupled Training Based on Combustion Mechanism Consistency: This step aims to improve the physical plausibility and engineering usability of the model's prediction results. In addition to the supervised loss, the system introduces a combustion mechanism consistency constraint loss term. The constraints include oxygen boundary consistency constraints. Temperature upper limit constraint Component negative correlation constraint and burnout coordination constraints Among them, the monotonicity constraint of oxygen content This is used to ensure that the predicted value of nitrogen oxides increases monotonically with increasing oxygen content; upper temperature limit constraint. ,in The theoretical upper limit of generation is determined based on the Arrhenius law; component negative correlation constraint. This is used to constrain the negative correlation coupling characteristics of carbon monoxide and nitrogen oxides under low-oxygen and high-temperature conditions; burnout coordination constraint. This is used to limit the model from producing predictions that violate the constraints of emission reduction and burnout.

[0044] Let the total loss function of the model be denoted as... Then its expression is as follows: ; in, This represents the loss term in multi-task supervision and prediction. This represents the loss term due to the consistency constraint of combustion mechanism. The smoothing constraint term represents the state change between adjacent time steps. and These are the weighting coefficients.

[0045] Through model training, the joint modeling network ultimately outputs the prediction vector for the current time step. With confidence score This provides state feedback and reliable metrics for subsequent reinforcement learning control decisions.

[0046] S3. Generation of emission control strategies based on deep reinforcement learning This step aims to optimize and generate specific physical control commands based on the prediction results output by the joint modeling network and the state-aware data, while ensuring the safe operation boundaries of the underlying devices; the specific process is as follows: Figure 4 As shown: First, the control feasible region is constructed and the operational boundary is mapped. Using the operating condition variables collected by S1 and the standardized low-NOx combustion state samples, the physical limits and single-step change rate limits of each actuator are calculated, and the control feasible region used to define the safety boundary of the action is output. Second, the control quantity generation based on the proximal policy optimization algorithm is executed. The standardized low-NOx combustion state samples output by S1 and the prediction vector output by S2 are used as state input to the reinforcement learning policy network. Under the strict constraints of the control feasible region, the initial control quantity is output by maximizing the expected cumulative reward. Third, the safety protection mechanism based on confidence gating is executed. The confidence score output by S2 is used to intercept and adjust the intensity reduction of the initial control quantity before execution, and the final execution control quantity is output. Finally, the control command generation and result output are executed. The final execution control quantity, the prediction vector, the confidence score, the control feasible region and the physical limits are merged and encapsulated into a complete closed-loop output result. Based on the complete closed-loop output result, the final execution control quantity is sent to the underlying distributed control system to execute the corresponding physical adjustment action.

[0047] S3.1 Control the construction and operation boundary mapping of the feasible region. This step is used to define the exploration space of the reinforcement learning policy and prevent the output of dangerous actions.

[0048] First, this invention defines the control space. For primary air regulation volume Secondary air distribution adjustment amount Graded wind input adjustment amount Coal supply and distribution adjustment amount and burner tilt adjustment The column vectors are formed as follows. Specifically, the primary air regulation amount controls the primary air fan frequency converter to adjust the momentum of the air-coal mixture; the secondary air distribution regulation amount controls the secondary air damper actuator in the main combustion zone to change the tiered combustion air ratio; the staged air input regulation amount controls the burnout air damper to accurately construct a reducing denitrification atmosphere zone; the coal feed distribution regulation amount controls the coal feeder drive motor to increase or decrease the actual fuel feed amount; and the burner tilt angle regulation amount controls the burner tilt angle drive device to change the flame center height and furnace heat absorption rate.

[0049] Secondly, this invention extracts operating condition variables. Boiler load instructions Combined operation mode with coal mill The initial physical limits of each actuator are determined by linear numerical mapping. and Meanwhile, this invention calculates... Offset of the center of the flame With oxygen content The deviation intensity is determined, and the safety margin of the current operating condition is quantified by comparing it with the distribution of historical stable operating data, thereby calculating the single-step rate of change limit of the control action. Based on this, the present invention constructs the basic action boundary. as follows: ; in, The scaling factor is determined by offline cross-validation.

[0050] S3.2 Control quantity generation based on the near-end policy optimization algorithm. This step is used to generate control variables within the defined control feasible region. The optimal control strategy is solved internally. This invention constructs a policy network based on a proximal policy optimization algorithm, wherein the policy network is based on... and prediction vector As state input. To guide the network in learning the optimal balance between nitrogen reduction and stability maintenance, this invention utilizes The various predictive indicators in the definition of the comprehensive reward function as follows: ; in, and Directly extract the self-predicted vector , This is a risk item calculated based on the furnace thermal deviation characterization value. For risk terms calculated based on burnout state characterization values, the weighting coefficients are... to and reward coefficient All were determined through offline multi-objective balance experiments. This invention maximizes the expected cumulative reward, which is achieved by the policy network... Directly output preliminary control quantity under constraints The preliminary control quantity That is, a set of instantiated feature column vectors within the control space; specifically, the initial control quantity The primary air conditioning quantity controls the primary air fan frequency converter to regulate the momentum of the air-powder mixture; the preliminary control quantity The secondary air distribution adjustment quantity controls the secondary air damper actuator in the main combustion zone to change the tiered combustion air ratio; the preliminary control quantity The staged air input adjustment quantity controls the burnout air damper to precisely construct a reducing denitrification atmosphere range; the preliminary control quantity The coal feeding distribution adjustment quantity controls the coal feeder drive motor to increase or decrease the actual fuel feed amount; the preliminary control quantity The burner tilt angle adjustment control device changes the flame center height and furnace heat absorption rate by adjusting the burner tilt angle. Through the coordinated issuance of physical adjustment commands in the above five dimensions, the control task of suppressing nitrogen oxide generation is completed.

[0051] S3.3 Confidence-based gating security protection mechanism. This step is used to control the initial control quantity. Pre-execution interception is performed. This invention uses offline training set statistics to determine high-confidence thresholds. Set to 0.85, low confidence threshold Set to 0.60. When At that time, the present invention directly issues preliminary control quantities. ;when At that time, the present invention introduces a scaling factor. Calculate the final execution control quantity This achieves a conservative reduction in the adjustment intensity; when When the prediction fails, the present invention forcibly blocks the strategy network output and switches to a conservative control template built based on expert experience. The conservative control template forcibly sets the control command to a safe air-fuel ratio specified by the expert, increases the total air volume to address potential carbon monoxide accumulation and furnace slagging risks, and simultaneously locks the fuel distribution ratio to ensure the system's bottom-line safety during online operation.

[0052] S3.4 Control command generation and result output. Finally, this invention will generate the final execution control quantity after gating interception. The current status information is merged and sent to the operator's terminal. The complete closed-loop output result of the system at the current moment is recorded as follows: : ; The complete closed-loop output result is then presented. The data is parsed and transmitted to the operator's terminal; the final execution control quantity, representing the adjustment action, is synchronously rendered and displayed on the operator's terminal interface. Prediction vector as the basis for adjustment The control feasible domain as the action safety boundary and action confidence score ; Receive manual review and release instructions from operators based on the above-mentioned rendered display data, and execute the final control quantity according to the manual review and release instructions. The corresponding physical adjustment actions are then executed by sending them down to the underlying distributed control system.

[0053] S4. Experimental Analysis To comprehensively verify the effectiveness, physical interpretability, and industrial operation safety of the deep learning-based low-NOx combustion process modeling and emission control method proposed in this invention, this embodiment constructs a simulation experimental platform that includes multi-source heterogeneous data fusion, joint modeling network training, and deep reinforcement learning strategy evaluation. The experimental dataset is constructed based on historical sensor data from typical industrial combustion devices, and the aforementioned multi-source heterogeneous data standardization processing logic is executed to extract 20,000 standardized low-NOx combustion state samples covering different load ranges, varying coal quality conditions, and equipment physical boundary constraints. The experimental verification process focuses on evaluating the quantitative performance of this invention in three core dimensions: multi-objective state collaborative prediction accuracy, combustion mechanism constraint boundary compliance, and safety control and defense performance based on prediction confidence gating.

[0054] Figure 5 The results of comparative tests on the errors of different algorithms in predicting multi-objective states of low-NOx combustion are presented. The horizontal axis represents different prediction targets, and the vertical axis represents the mean absolute percentage error. Using the basic deep learning soft sensor model, the mean absolute percentage errors for NOx and carbon monoxide predictions are 6.5% and 9.2%, respectively. Using the joint modeling network of this invention, the prediction errors for four indicators—NOx emission concentration, carbon monoxide emission concentration, burnout state characterization value, and furnace thermal deviation characterization value—are 2.4%, 3.8%, 3.1%, and 3.5%, respectively. These objective data demonstrate that the joint modeling network can extract the collaborative mapping relationship of multi-source heterogeneous data, reducing the prediction bias of single-indicator modeling.

[0055] Figure 6 This paper presents a comparison of emission boundary distributions before and after the introduction of combustion mechanism constraints. The horizontal axis represents the predicted NOx emission concentration, and the vertical axis represents the predicted carbon monoxide emission concentration. Without the introduction of mechanism constraints, when NOx values ​​are below 200 mg / m³, the corresponding predicted carbon monoxide values ​​significantly exceed the safe operating upper boundary of 100 mg / m³. After introducing combustion mechanism consistency constraints, all model output samples are limited to a range of 10 to 85 mg / m³. These data distribution characteristics indicate that mechanism consistency constraints correct the prediction bias of the data-driven model under boundary conditions, ensuring the safety and compliance of the control solution results within the feasible region.

[0056] Figure 7 The correlation distribution between prediction confidence and the safety degradation range of the control strategy is shown. The horizontal axis represents the prediction confidence of the joint modeling network output, and the vertical axis represents the absolute error of emission prediction. When the prediction confidence is greater than or equal to 0.85, the absolute error is distributed in the range of 5 mg / Nm³ and below, and the system executes the direct output strategy of the control quantity. When the prediction confidence is between 0.60 and 0.85, the absolute error value shows an upward trend, and the system triggers a scaling mechanism to reduce the control quantity. When the prediction confidence is below 0.60, the absolute error distribution diverges, and the system pauses the output of the deep reinforcement learning network and switches to a conservative control template. The above response distribution data shows that the confidence gating mechanism effectively intercepts risky control commands under low confidence conditions, which complies with the safety operation specifications of industrial control systems.

[0057] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0058] While the above description illustrates specific embodiments of the present invention, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for modeling and controlling low-NOx combustion processes based on deep learning, characterized in that, Includes the following processes: S1 collects flame image data, process variables and mechanism operation data, flue gas emission data, fuel data and operating condition data during the low-NOx combustion process, and preprocesses them to obtain flame image features, process variable features, flue gas and emission features, fuel features and operating condition context features. S2, construct a low-NOx combustion joint modeling network based on a deep learning and mechanism constraint coupling architecture, use a multimodal coding combustion state feature extraction network to map the feature data obtained in S1 into unified combustion scenario features, and output prediction vectors and confidence scores of multiple core combustion state indicators through a multi-task learning-based emission and combustion state joint prediction module. S3, based on the physical limits of each actuator and the single-step change rate limit, obtains the controllable feasible region; the features obtained in S1 and the prediction vector output in S2 are used together as the state input to the reinforcement learning policy network, and under the constraint of the controllable feasible region, the initial control quantity is output by maximizing the expected cumulative reward. A confidence-based gating security protection mechanism is implemented, which uses the confidence score output by S2 to intercept and reduce the intensity of the initial control quantity before execution, and outputs the final execution control quantity.

2. The method for modeling and controlling low-NOx combustion processes based on deep learning as described in claim 1, characterized in that: The specific process for constructing the flame image features is as follows: First, acquire furnace flame image data of the flame area. And preprocessed flame image samples were obtained. ; Secondly, construct derived visual feature vectors. For preprocessed flame image samples Extracting flame brightness distribution Main combustion zone area ratio Flame center offset undulation of the flame edge and color distribution statistics They are then uniformly concatenated and encapsulated into derived visual feature vectors. ; Subsequently, a short temporal visual sample sequence was constructed. First, for the preprocessed flame image samples from multiple time points... Constructing a time sliding window from the continuous sequence formed The time sliding window The length is set to , the current moment and its predecessor Preprocessed flame image samples at each time point A short-temporal visual sample sequence is obtained by concatenating multidimensional tensors in chronological order. ; Finally, the short temporal visual sample sequence and its corresponding derived visual feature vectors Flame image features are obtained through normalization and feature concatenation operations. .

3. The method for modeling and controlling low-NOx combustion processes based on deep learning as described in claim 1, characterized in that: The process of constructing the process variable characteristics is as follows: The primary air volume is read in real time from the distributed control system, field instrumentation system, and actuator feedback unit. Secondary air volume Total air volume air supply temperature Temperature of each layer of the furnace Flue temperature Oxygen measurement data Perform resampling, anomaly removal, and uniform time step slicing, with a fixed time window. Organizing the data into time series segments yields a preprocessed sequence of process variables. ,right Perform normalization and feature concatenation operations to obtain process variable features. .

4. The method for modeling and controlling low-NOx combustion processes based on deep learning as described in claim 1, characterized in that: The process for constructing the flue gas and emission characteristics, fuel characteristics, and operating condition context characteristics is as follows: Nitrogen oxide emission concentrations are obtained from online flue gas monitoring devices and on-site analysis instruments. carbon monoxide concentration Oxygen content Simultaneously read the carbon content of fly ash The above four data points are combined to form the characteristics of flue gas and emissions. ; Collect coal quality variables Including moisture and fuel nitrogen content ,right Fuel features are obtained by performing nonlinear numerical mapping and multidimensional vector concatenation. ; Collect operating condition variables Including boiler load instructions Combined operation mode of coal mill ,right Perform nonlinear numerical mapping and multidimensional vector concatenation to obtain working condition context features. .

5. The method for modeling and controlling low-NOx combustion processes based on deep learning as described in claim 1, characterized in that: The combustion state feature extraction network utilizing multimodal coding maps the feature data obtained in S1 to unified combustion scene features, specifically as follows: A flame image feature extraction network is constructed to extract and output image branch latent features based on flame image features; a process variable temporal coding network is constructed to extract and output process branch latent features based on process variable features; an emission state coding network is constructed to extract and output emission branch latent features based on flue gas and emission features; an operating condition semantic and component attribute coding network is constructed by concatenating fuel features and operating condition context features to extract and output operating condition branch latent features; then, the image branch latent features, process branch latent features, emission branch latent features, and operating condition branch latent features are concatenated along the channel dimension and fused with attention weighting to output unified combustion scene features.

6. The method for modeling and controlling low-NOx combustion processes based on deep learning as described in claim 5, characterized in that: The flame image feature extraction network consists of four two-dimensional convolutional modules. The first convolutional layer has a 3×3 kernel size, 1 input channel, 32 output channels, a stride of 1, and padding of 1. The second convolutional layer has a 3×3 kernel size, 32 input channels, 64 output channels, a stride of 1, and padding of 1. The third convolutional layer has a 3×3 kernel size, 64 input channels, 128 output channels, a stride of 1, and padding of 1. The fourth convolutional layer has a 3×3 kernel size, 128 input channels, 256 output channels, a stride of 1, and padding of 1. Each convolutional module is followed by a batch normalization layer, a ReLU activation layer, and a 2×2 max pooling layer, which ultimately outputs an image feature map. Subsequently, a global average pooling layer and a fully connected layer with an output dimension of 128 are used to... Mapped to image branch latent features .

7. The method for modeling and controlling low-NOx combustion processes based on deep learning as described in claim 1, characterized in that: The joint emission and combustion state prediction module inputs the unified combustion scenario features into a shared feature backbone network and multiple prediction heads, and simultaneously outputs a prediction vector containing predicted values ​​for nitrogen oxide emission concentration, carbon monoxide emission concentration, burnout state characterization, furnace thermal deviation characterization, and combustion stability margin characterization. Subsequently, it performs prediction confidence assessment and model confidence interval generation, quantifies the uncertainty of the model in the forward inference process through a Monte Carlo random deactivation mechanism, and outputs a confidence score corresponding to the prediction vector.

8. The method for modeling and controlling low-NOx combustion processes based on deep learning as described in claim 1, characterized in that: The specific method for obtaining the control feasible region in S3 is as follows: Define control space For primary air regulation volume Secondary air distribution adjustment amount Graded wind input adjustment amount Coal supply and distribution adjustment amount and burner tilt adjustment The column vectors formed by the primary air regulation amount control the primary air fan frequency converter to adjust the momentum of the air-coal mixture; the secondary air distribution regulation amount control the secondary air damper actuator in the main combustion zone to change the tiered combustion air ratio; the staged air input regulation amount control the burnout air damper to accurately construct the reducing denitrification atmosphere zone; the coal feeding distribution regulation amount control the coal feeder drive motor to increase or decrease the actual fuel feed amount; the burner tilt angle regulation amount control the burner tilt angle drive device to change the flame center height and furnace heat absorption rate. By extracting operating condition variables Boiler load instructions Combined operation mode with coal mill The initial physical limits of each actuator are determined by linear numerical mapping. and Simultaneously, calculate the flame center offset. With oxygen content The deviation intensity is determined, and the safety margin of the current operating condition is quantified by comparing it with the distribution of historical stable operating data. The single-step rate of change limit of the control action is then calculated. Constructing basic action boundaries as follows: ； in, The scaling factor is determined by offline cross-validation.

9. The method for modeling and controlling low-NOx combustion processes based on deep learning as described in claim 8, characterized in that: The reinforcement learning policy network in S3 is used to predict vectors. The various predictive indicators in the definition of the comprehensive reward function as follows: ； in, This represents the predicted concentration of nitrogen oxide emissions. This represents the predicted carbon monoxide emission concentration, directly extracted from the prediction vector. ; This is a risk item calculated based on the furnace thermal deviation characterization value. For risk terms calculated based on burnout state characterization values, the weighting coefficients are... to and reward coefficient All were determined by offline multi-objective equilibrium experiments; Score the confidence level; By maximizing the expected cumulative reward, the policy network... Directly output preliminary control quantity under constraints The preliminary control quantity That is, a set of instantiated feature column vectors within the control space.

10. The method for modeling and controlling low-NOx combustion processes based on deep learning as described in claim 9, characterized in that: The confidence-based gating security protection mechanism implemented in S3 is as follows: High confidence threshold based on offline training set statistics Set to 0.85, low confidence threshold Set to 0.60; when the confidence score At that time, the preliminary control quantity was issued directly. ;when At that time, a scaling factor is introduced. Calculate the final execution control quantity ;when If the prediction fails, the policy network output is forcibly blocked, and the system switches to a conservative control template built based on expert experience.

Images