Method for measuring three-dimensional NOx distribution of hearth of coal-fired power plant boiler

Abstract

The invention is suitable for the technical field of boiler pollutant monitoring, and provides a coal-fired power plant boiler hearth three-dimensional NOx distribution measurement method which comprises the following steps: selecting parameter data related to coal-fired boiler NOx emission from DCS field operation data and CFD simulation data, and fusing the parameter data and the CFD simulation data to construct a basic modeling data set; performing cleaning and standardization processing on the fused data; dividing the data into a plurality of typical working condition categories, selecting representative training sample points from the fused data in each working condition by adopting a layering importance resampling method based on the consistency of a PDE residual error and a light path, and obtaining a modeling data set after resampling; constructing a three-dimensional NOx distribution reconstruction model based on a multi-model intelligent combination strategy; and verifying the model and outputting a result. According to the method, high-precision and near-real-time reconstruction of the NOx in the hearth can be realized under the conditions of sparse measurement and complex working conditions, and the method has good engineering applicability and popularization value.

Description

Technical Field

[0001] This invention belongs to the field of boiler pollutant monitoring technology, and particularly relates to a three-dimensional NO monitoring system for the furnace of a coal-fired power plant boiler. x Distribution measurement methods. Background Technology

[0002] Existing power plants typically install CEMS (Continuous Emission Monitoring System) devices at the tail end of the boiler to monitor NO. x Online monitoring is conducted, but the results are mostly single-point or limited cross-sectional information from the outlet, which is insufficient to reflect the NO content inside the furnace. x The spatial distribution and spatiotemporal evolution characteristics of ammonia lead to a reliance on experience for ammonia injection and air distribution regulation, resulting in problems such as ammonia escape, local excess air, and uneven combustion. Optical absorption tomography (OFT) can provide multipath integral information under high-temperature, high-radiation environments, but due to factors such as sparse optical paths, ash accumulation disturbances, and unsteady flow fields, direct reconstruction of the three-dimensional field still exhibits ill-posed and ill-posed characteristics. On the other hand, Computational Fluid Dynamics (CFD) can provide physical priors for convection-diffusion-reaction, but it is sensitive to boundaries and parameters, and relying solely on it often leads to predictions that deviate from reality. Therefore, there is an urgent need for a three-dimensional reconstruction method that can integrate sparse measurements, physical priors, and data-driven learning to achieve engineering-usable and robust furnace NOx. x Three-dimensional measurement methods. Summary of the Invention

[0003] The purpose of this invention is to provide a three-dimensional NO content in the furnace of a coal-fired power plant boiler. x The distribution measurement method aims to solve the problems mentioned in the background art above.

[0004] The present invention is implemented as follows: a three-dimensional NO in the furnace of a coal-fired power plant boiler. x The distribution measurement method includes the following steps:

[0005] Selecting data from DCS field operation data and CFD simulation data related to the NO of coal-fired boilers x Emission-related parameter data will be merged to construct a basic modeling dataset;

[0006] The merged data is then cleaned and standardized.

[0007] The data is divided into several typical working condition categories. Within each working condition, a hierarchical importance resampling method based on the consistency of PDE residuals and light path is used to select representative training sample points from the fused data to obtain the resampled modeling dataset.

[0008] Constructing a three-dimensional NO based on a multi-model intelligent combination strategy xDistributed reconstruction model;

[0009] Validate the model and output the results.

[0010] This invention provides a three-dimensional NO content in the furnace of a coal-fired power plant boiler. x The distributed measurement method, by fusing DCS operational data and CFD simulation results, and combining voxelization, wavelet denoising, and graph total variational smoothing, effectively improves the integrity and stability of the data. Based on this, a hierarchical resampling method based on PDE residuals and optical path consistency is proposed, which retains key region features while reducing data size, significantly improving modeling efficiency. Through a combination of algebraic tomography, physical constraint optimization, and a physics-guided deep learning model, the method achieves accurate measurement of furnace NO₂. x The distributed high-precision 3D reconstruction incorporates optical path consistency and PDE constraints in the loss function, enabling the results to possess both data fitting capability and physical consistency. In the verification stage, multi-index evaluation is adopted. Ultimately, the reconstruction results can be intuitively displayed through a 3D visualization system and linked with the ammonia injection and air distribution systems to achieve intelligent optimization and emission control of boiler operation, demonstrating good engineering application value and promising prospects for promotion. Attached Figure Description

[0011] Figure 1 A three-dimensional NO content in the furnace of a coal-fired power plant boiler is provided as an embodiment of the present invention. x Flowchart of the distribution measurement method;

[0012] Figure 2 The experimental results after data processing and sampling are shown in the figure provided in the embodiment of the present invention;

[0013] Figure 3 The model prediction results provided in this embodiment of the invention (a is a scatter plot, b is the distribution of prediction error with predicted value);

[0014] Figure 4 The experimental results error histogram provided for embodiments of the present invention. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0016] A three-dimensional NO in the furnace of a coal-fired power plant boiler x The distribution measurement method specifically includes the following steps:

[0017] Step 1: Data Acquisition. Select data from DCS field operation data and Computational Fluid Dynamics (CFD) simulation data related to the NO of the coal-fired boiler. xEmission-related parameter data are integrated to construct a basic modeling dataset. The DCS side provides nine operating parameters, including load, total air volume, total coal consumption, and primary / secondary air volume. The CFD side provides three-dimensional coordinates, heat flux density field, temperature field, and NOx data under 66 operating conditions. x Thirteen simulation parameter fields, including prior distribution, are fused spatially between the two types of data to obtain the basic modeling dataset, which contains 66 9196324×22 matrices for modeling.

[0018] Step 2: Data preprocessing, cleaning and standardizing the fused 3D data:

[0019] Sparse point cloud data is voxelized into a regular 3D mesh, which facilitates spatial domain processing;

[0020] Subsequently, a three-dimensional wavelet thresholding method was used to remove high-frequency noise, and graph total variational (GTV) smoothing was combined to suppress isolated outliers in the voxel neighborhood, thereby eliminating interference while maintaining furnace NO x Spatial structural characteristics of the distribution;

[0021] Finally, min-max normalization is applied to all features to map parameters from different sources and with different dimensions to a unified scale, ensuring that the input features are stable and reliable in subsequent feature selection and 3D reconstruction modeling.

[0022] Step 3: Working condition classification and modeling data resampling:

[0023] Based on the number of coal mills in operation and the combination of mills in operation, the typical operating conditions selected in the numerical simulation are classified, such as single mill operation, dual mill operation and full load operation, to ensure that the model training can cover the full range of operating conditions.

[0024] Since CFD simulation data points can reach millions, directly using them for training not only consumes computational resources but also leads to model overfitting. Therefore, this invention proposes a hierarchical importance resampling method based on PDE residuals and optical path consistency, specifically:

[0025] First, the furnace space is divided into layered areas such as the burner zone, recirculation zone, near-wall layer, and convection channel.

[0026] Calculate the PDE residual, gradient intensity, optical path consistency error and uncertainty in each region, and integrate them to form the voxel importance weight;

[0027] Then, three-dimensional Poisson disk sampling is performed in each layer according to the weights, while ensuring that there are a sufficient number of sampling points in the neighborhood of each ray path;

[0028] Finally, the sample set is dynamically updated through iterative active learning, and the most important regions are sampled in a more intensive manner, so as to balance global representativeness and key detail fidelity with a limited sample size.

[0029] Finally, 50,000 typical sample data points were selected as modeling data.

[0030] Step 4: 3D NOx Distribution Reconstruction Model (NOx-3DNet). This involves constructing a multi-model intelligent combination framework for reconstructing the 3D NOx distribution. This includes: using an algebraic tomography method (SART / SIRT combined with regularization) as the initial reconstruction to provide a warm start and teacher signal for the deep learning model; embedding the convection-diffusion-reaction partial differential equation (PDE) residuals as physical constraints into the optimization process based on the velocity and temperature fields provided by CFD; and employing a physics-guided 3D convolutional neural network (3DU-Net) as the core learning model, simultaneously introducing optical path integral consistency error and PDE residual regularization into the loss function to balance data-driven expressiveness and physical consistency. This multi-model intelligent combination framework can adaptively select or fuse results under different operating conditions, enabling accurate reconstruction of NOx distribution inside the furnace. x High-precision 3D reconstruction of distribution;

[0031] Step 5: Model Validation and Prediction Results: Validating the model using MAE (Mean Absolute Error), RMSE (Root Mean Square Error), sMAPE (Symmetric Mean Percentage Error), and R². 2 The results are verified using indicators such as the coefficient of determination. Finally, the results are displayed intuitively in a three-dimensional visualization system and linked with the ammonia injection and air distribution control system to generate combustion optimization and emission control recommendations.

[0032] The specific implementation of the present invention will be described in detail below with reference to specific embodiments.

[0033] like Figure 1 As shown, this is an embodiment of the present invention providing a three-dimensional NO content in the furnace of a coal-fired power plant boiler. x The flowchart of the distributed measurement method, taking a 1000MW ultra-supercritical lignite boiler as an example, is a single-furnace, balanced draft, solid ash discharge, π-type, compactly enclosed once-through boiler, including the following steps:

[0034] Step 1, Data Acquisition:

[0035] First, real-time furnace data is acquired from the historical data monitoring system (DCS) and online optical monitoring system of the coal-fired power plant boiler. This data includes important parameters such as furnace temperature, oxygen content, and NOx concentration. In addition, CFD (Computational Fluid Dynamics) simulations provide the system with information about the internal flow, temperature field, and NOx concentration within the furnace. xPrior information about the distribution was used. The CFD simulation employed unstructured mesh generation and a pressure-velocity coupled solver. The turbulence model was the RNG k-ε model, and the combustion model used a component transport model. NO... x The generative model combines thermal, fast, and fuel-based mechanisms.

[0036] Step 2, Data Preprocessing:

[0037] After fusing DCS operational data with CFD simulation results, the point cloud data is first mapped to a unified 3D mesh through voxelization for subsequent spatial analysis and modeling. The following processing flow is used for the voxelized 3D field:

[0038] (1) Three-dimensional wavelet threshold denoising: Multi-scale decomposition of the voxel field is performed to remove high-frequency noise components and retain the main structural information;

[0039] (2) Graph Total Variation (GTV) Smoothing: Perform graph-constrained smoothing within the neighborhood of the voxel mesh to suppress isolated outliers while preserving boundary and local structural features;

[0040] (3) min-max normalization: Maps data features of different dimensions and sources to a unified numerical range to ensure that the input data are comparable and stable in the subsequent feature selection and model training process;

[0041] Let the original point cloud data be ,in For spatial coordinates, For the corresponding CFD or DCS parameter values, N represents the total number of point clouds, which is 9,196,324 points. First, the point cloud is mapped to a regular 3D mesh through voxelization. :

[0042] ;

[0043] in, For the index (voxel number) of the 3D mesh. This represents the set of points falling into a voxel. for The number of midpoints The average value of all points within the voxel lattice is used to obtain a regular three-dimensional field;

[0044] Then, V is subjected to three-dimensional wavelet decomposition and a soft threshold is applied. As shown below:

[0045] ;

[0046] in, The approximate (low-frequency) coefficients after wavelet decomposition represent the large-scale structure. Indicated in scale Down, direction The detail factor includes high-frequency noise information. For the threshold, For symbolic functions, This indicates a soft thresholding operation, which suppresses high-frequency coefficients below the threshold to 0;

[0047] Based on this, graphical total variation (GTV) smoothing is introduced, resulting in the optimized field shown below:

[0048] ;

[0049] in, The voxel field variables to be optimized. This is the initial field after wavelet denoising. To ensure fidelity, the smoothed result closely approximates the original data. The balancing coefficient adjusts the weights for fidelity and smoothing. Let be the set of adjacent edges of voxels, representing the connection relationships between adjacent voxels. This represents the edge weight between voxels i and j, typically set to 1 or related to similarity. This represents the difference between adjacent voxels, penalizing local mutations.

[0050] Finally, the smoothed voxel field is normalized using the min-max method, as shown below:

[0051] ;

[0052] in, These are the smoothed voxel values; and Let these be the maximum and minimum values ​​of this feature across all primes. The result is normalized, and the numerical range is mapped to [0,1].

[0053] Step 3: Working condition classification and modeling data resampling:

[0054] The dataset is divided according to the boiler's operating conditions, and typical operating conditions are categorized based on the number and combination of coal mills in operation; due to the fusion of CFD simulation and actual measurement, the three-dimensional NO... x Fields typically contain millions of voxels; directly using all of them for training would not only consume enormous computational resources but also easily lead to model overfitting. Therefore, this invention introduces a hierarchical importance resampling method based on PDE residuals and optical path consistency within each type of working condition to preserve key feature regions with a limited number of samples, such as... Figure 2 The figure shown is a graph of the experimental results after data processing and sampling.

[0055] The specific steps are as follows:

[0056] (1) Spatial layering: According to the furnace structure and flow field characteristics, the three-dimensional computational domain is divided into typical spatial layers such as burner zone, recirculation zone, near-wall layer and convection channel to ensure that different physical regions are covered;

[0057] (2) Indicator calculation and weight allocation: Four types of indicators are calculated on each voxel:

[0058] The PDE residual norm R(x) characterizes the deviation from the consistency of the convection-diffusion-reaction equation;

[0059] Gradient and curvature intensity G(x) are used to identify boundaries and transition regions where NOx distribution changes drastically;

[0060] The optical path consistency error E(x) is obtained by back-projecting the optical path integral error, reflecting the model-measurement deviation.

[0061] Uncertainty U(x) is a variance estimate obtained from a deep model or ensemble method;

[0062] After normalization, these indicators are merged into a comprehensive importance weight w(x);

[0063] (3) Weighted sampling: In each spatial layer, based on w(x) as the probability, the three-dimensional Poisson disk sampling method is used to select sample points to avoid excessive sample aggregation. In order to ensure coupling with the measurement, no less than a preset number of sample points are forcibly selected in the neighborhood of each optical path to ensure that the measurement constraints are fully utilized.

[0064] (4) Iterative active learning: The initial selected samples are used to train the model. Then, the weights are updated according to the new residual and uncertainty distribution. Samples are added and low-weight samples are replaced in a certain proportion of the voxel regions with the highest weights. After 2-3 iterations, the sample set gradually concentrates on the most sensitive and important regions for prediction.

[0065] (5) Sample size control: The final sample set is approximately 50,000 points, which significantly reduces the amount of data compared to full-domain sampling, but still maintains the representativeness of the entire furnace structure characteristics, measurement path and key gradient region;

[0066] The mathematical formulas for the above steps are shown below:

[0067] ;

[0068] in For the entire set of voxels; This is a set of sampling points (50,000 points in size); , , , These are the normalized values ​​of PDE residual, gradient intensity, optical path consistency error, and uncertainty, respectively. , , , These are the weighting coefficients for each indicator; This is an indicator function; it takes the value 1 if voxel x is sampled, and 0 otherwise. The Poisson disk constraint term ensures that the sampling points maintain at least [a certain distance]. Maintain social distancing to avoid excessive gatherings; To balance the constraint strength.

[0069] Step 4, 3D NO x Distributed reconstruction model:

[0070] After feature selection and resampling, a three-dimensional NO is constructed. x Distributed Reconstruction Model (NO) x -3DNet), employing a multi-model intelligent combination strategy:

[0071] Algebraic tomography baseline model: Based on the SART / SIRT method combined with Tikhonov or TV regularization, a preliminary estimate of the NOx distribution in the furnace is obtained, providing a hot start for subsequent models;

[0072] Physically constrained optimization model: Utilizing the velocity and temperature fields provided by CFD, the residuals of the convection-diffusion-reaction equations are introduced as constraints in variational optimization to enhance the physical consistency of the results;

[0073] Physics-guided deep learning model: Employing a 3D U-Net as the core model, the input is fused multi-channel features, and the output is NO. x Three-dimensional distribution;

[0074] During training, the loss function is composed of voxel-level error, optical path integral consistency error, PDE residual term, and spatial TV regularization. The optical path consistency error is achieved through differentiable ray integration to ensure that the prediction results match the measured optical path data. If SART reconstruction results are provided, distillation loss is introduced to enhance the stability and convergence speed of the model.

[0075] The mathematical expression for the loss function is as follows:

[0076] ;

[0077] in, For voxel-level error, For optical path consistency error, For PDE residual constraints, For total variation regularization, For distillation loss, , , , , These are the weighting coefficients for each item, adjusted based on the performance on the validation set, used to balance the contributions of different constraints.

[0078] Step 5: Model Validation and Prediction Results:

[0079] Model prediction results are as follows Figure 3 As shown, where 'a' is a scatter plot, it displays the predicted NO x NO to reality x The relationship between the two is that most sample points are closely distributed near the diagonal, indicating that the prediction results are highly consistent with the measured values. The overall trend is linear, the low concentration area has a better fitting effect, and the high concentration area has a slight underestimation, but no systematic bias is found, indicating that the model can effectively capture the overall characteristics of NOx distribution inside the furnace.

[0080] To further analyze the prediction performance, b shows the distribution of prediction error with respect to the predicted value. As can be seen from the figure, the error for the vast majority of sample points is concentrated within ±25 mg / m². 3 Within the range, and with the error distribution being basically symmetrical, it indicates that the model maintains good stability in different concentration ranges. In the high concentration range, the error increases slightly, but overall it is still within the acceptable range for engineering.

[0081] In addition, the prediction error histogram is as follows: Figure 4 As shown, the overall distribution of errors is illustrated: the prediction errors for the vast majority of samples are concentrated between -25 and +25 mg / m³. 3 Within the given interval, the peak value is located near zero, and the overall distribution exhibits approximately normal characteristics. This indicates that the model has no significant bias overall, and the prediction results demonstrate good centralization and stability. Only a small number of samples show prediction errors exceeding ±50 mg / m². 3 However, the frequency of these errors is extremely low, indicating that extreme errors have a limited impact on overall performance, and the model can still maintain high reliability under complex working conditions.

[0082] After model training, the results are comprehensively evaluated using independent validation sets and experimental data. Validation metrics include: MAE (Mean Absolute Error), RMSE (Root Mean Square Error), sMAPE (Symmetric Mean Percentage Error), and R². 2 The evaluation indexes, such as the coefficient of determination, are used to verify the results and provide a reliable confidence level for subsequent ammonia injection and air distribution control. The formulas for the evaluation indexes are shown below:

[0083] ;

[0084] ;

[0085] ;

[0086] ;

[0087] Where N is the number of samples; This represents the true value (True) of the i-th sample. Pred is the predicted value of the i-th sample. The mean of the true values;

[0088] The prediction model NO provided in the embodiments of the present invention x - A comparison of prediction accuracy between 3DNet and ELM, DNN, and MLP is shown in Table 1:

[0089] Table 1

[0090]

[0091] It can be seen that the prediction accuracy of the embodiments of the present invention has a better advantage than other models.

[0092] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A three-dimensional NO content in the furnace of a coal-fired power plant boiler x The distribution measurement method is characterized by, Includes the following steps: Selecting data from DCS field operation data and CFD simulation data related to the NO of coal-fired boilers x Emission-related parameter data will be merged to construct a basic modeling dataset; The merged data is then cleaned and standardized. The data is divided into several typical working condition categories. Within each working condition, a hierarchical importance resampling method based on the consistency of PDE residuals and light path is used to select representative training sample points from the fused data to obtain the resampled modeling dataset. Constructing a three-dimensional NO based on a multi-model intelligent combination strategy x Distributed reconstruction model; Validate the model and output the results.

2. The three-dimensional NO in the furnace of a coal-fired power plant boiler according to claim 1 x The distribution measurement method is characterized by, The selection of NO from DCS field operation data and CFD simulation data for coal-fired boilers x The steps for fusing emission-related parameter data and other data to construct a basic modeling dataset are as follows: Obtain furnace operating parameters from the DCS system; Obtain flow field, temperature field, and NO from CFD simulation. x Prior information about the distribution The two types of data are spatially fused to obtain the basic modeling dataset.

3. The three-dimensional NO in the furnace of a coal-fired power plant boiler according to claim 1 x The distribution measurement method is characterized by, The steps for cleaning and standardizing the fused data are as follows: Convert point cloud data into a unified mesh; High-frequency noise was removed using a three-dimensional wavelet thresholding method. Smooth outliers in the spatial neighborhood using graph total variation; All features are normalized using min-max normalization.

4. The three-dimensional NO in the furnace of a coal-fired power plant boiler according to claim 1 x The distribution measurement method is characterized by, The steps of dividing the data into several typical working condition categories, and selecting representative training sample points from the fused data using a hierarchical importance resampling method based on the consistency of PDE residuals and ray paths within each working condition to obtain the resampled modeling dataset are as follows: Based on the number and combination of coal mills in operation, the data is divided into several typical operating condition categories; The furnace space is divided into different layered regions. The PDE residual, gradient / curvature intensity, optical path consistency error and uncertainty are calculated in each region and then fused to form the voxel importance weight. Within each layer, three-dimensional Poisson disk sampling is performed according to weights, while ensuring that each ray path has a sufficient number of sampling points in its neighborhood. The sample set is dynamically updated through iterative active learning, and the most important regions are sampled more densely to achieve adaptive optimization of the sample set, resulting in a resampled modeling dataset.

5. The three-dimensional NO in the furnace of a coal-fired power plant boiler according to claim 1 x The distribution measurement method is characterized by, The three-dimensional NO constructed based on the multi-model intelligent combination strategy x The steps of the distributed reconstruction model are as follows: Preliminary reconstruction results are obtained using algebraic tomography, providing a warm start and teacher signal for the deep learning model; Based on the velocity and temperature fields provided by CFD, the residuals of the convection-diffusion-reaction partial differential equations are embedded as physical constraints in the optimization process; A physics-guided 3D convolutional neural network is used to learn from the fused data. The optical path integral consistency error and PDE residual regularization are introduced into the loss function to balance the expressive power of data-driven learning with physical consistency.

6. The three-dimensional NO in the furnace of a coal-fired power plant boiler according to claim 1 x The distribution measurement method is characterized by, The steps for validating the model and outputting the results are as follows: Validate the model's prediction results; The results are linked with the ammonia injection and air distribution control systems to generate combustion optimization and emission control recommendations.

Images