Boiler combustion digital twin modeling method fusing atmosphere field and flame visual data

Abstract

The application discloses a boiler combustion digital twin modeling method fusing atmosphere field and flame visual data; the method collects near-wall atmosphere field data, flame visual data, DCS operation parameters and available water-cooled wall state parameters of a coal-fired boiler; abnormal data is eliminated, delay compensation, quality evaluation, time-space synchronization, quality weighted atmosphere field reconstruction and flame temperature field inversion are performed on the multi-source data, a unified state vector containing a data quality vector is formed; a combustion flow field basic simulation model is established based on a CFD simulation platform and simulation prior conditions of working conditions are formed; a data-driven fusion model with physical constraints is trained by using historical operation samples, simulation prior conditions and engineering rules; during training, measurement error, physical residual error, parameter smoothing and simulation prior condition deviation are combined; during online prediction, current state, simulation prior conditions and historical state sequences are inputted to obtain digital twin prediction results of a furnace temperature field, a combustion flow field, a component field, a water-cooled wall state and a risk state.

Description

Technical Field

[0001] This invention belongs to the technical fields of boiler combustion monitoring, industrial process digital twin modeling, and multi-source data fusion in thermal power units. Specifically, it involves a boiler combustion digital twin modeling method that integrates atmospheric field and flame visual data. It is applicable to the online modeling, prediction, and correction of combustion flow field, temperature field, atmospheric field, and water-cooled wall risk status of coal-fired boilers under complex operating conditions such as deep peak shaving, rapid load change, and coal quality fluctuation. Background Technology

[0002] With the increasing demand for flexible operation of thermal power units, coal-fired units often operate under non-design conditions such as fluctuating coal quality, rapid and frequent load changes, ultra-low loads, and daily start-up and shutdown for peak shaving. These conditions simultaneously affect the boiler's combustion stability, economy, environmental protection, and operational safety. In particular, the boiler side is both the power source and the combustion control terminal, making operational safety risks more concentrated.

[0003] Current combustion monitoring mainly relies on conventional flame detectors, DCS operating parameters, tail gas indicators, and a small number of wall temperature or thermal parameters. Conventional flame detectors typically use light intensity information at specific frequencies to determine the presence or absence of a flame, but have a small field of view, leading to problems such as oversights, missed detections, and misjudgments. Furthermore, they are difficult to quantify continuous states such as flame center, flame deviation, temperature field, and combustion stability. When combustion regulation is based solely on parameters such as main steam pressure, flow rate, and oxygen content, the feedback signal is lagging and cannot directly reflect changes in the combustion flow field inside the furnace.

[0004] The near-wall atmosphere field can reflect the reducing atmosphere and corrosion risk near the water-cooled wall. Under low-NOx combustion and staged air combustion conditions, the concentration of reducing components such as H2S and CO is prone to increase near the water-cooled wall, which can induce problems such as high-temperature corrosion, coking, ash accumulation, and flame deflection. Existing near-wall atmosphere measurement schemes are mostly used for single-point or multi-point monitoring, soft measurement, or wall-mounted air conditioning, and have not yet been fully integrated with flame visual data, CFD full furnace models, and online digital twin correction.

[0005] Digital twin technology can establish a mapping model of a physical entity in virtual space and dynamically update it using operational data. Existing boiler digital twin solutions mostly focus on mechanistic models, equilibrium equations, or updates to single types of operational data, resulting in insufficient correction of measured flow, temperature, and atmosphere fields, as well as inadequate model real-time performance and adaptability to a wide range of operating conditions. For pulverized coal boiler combustion processes, if only offline CFD simulation is used, the calculation results depend on boundary conditions and reaction models, making it difficult to quickly respond to real-time operating conditions; if only data-driven models are used, there are problems with insufficient physical interpretability and inadequate out-of-condition generalization ability.

[0006] In order to solve the above problems, this invention is hereby proposed. Summary of the Invention

[0007] One object of the present invention is to provide a method for digital twin modeling of boiler combustion that integrates atmospheric field and flame visual data.

[0008] To achieve the above objectives, the present invention adopts the following technical solution: A method for digital twin modeling of boiler combustion that integrates atmospheric field and flame visual data includes the following steps: Collect near-wall atmosphere field data, flame visual data, DCS operating parameters, and available water-cooled wall state parameters of coal-fired boilers; The near-wall atmosphere field data, flame visual data, DCS operating parameters, and water-cooled wall status parameters are respectively subjected to anomaly removal, timestamp verification, delay compensation, missing data completion, standardization, and quality evaluation to form atmosphere channel quality factor, measuring point quality factor, image quality factor, and water-cooled wall status validity identifier. The preprocessed multi-source data are spatiotemporally synchronized according to a unified sampling period to construct a unified state vector that includes near-wall atmosphere field characteristics, flame visual characteristics, DCS operation characteristics, water-cooled wall state characteristics, operating condition labels, and data quality vectors. The near-wall component concentration field is reconstructed by mass weighting based on the relative spatial relationship of the measuring points, the mass factor of the measuring points, and the component concentration. An atmospheric component consistency residual is constructed, which is used as a physical residual and a risk check quantity to participate in model training, online correction, and risk output. A basic simulation model of the boiler combustion flow field was established using a CFD simulation platform, forming simulation priors indexed by operating condition labels. A data-driven fusion model with physical constraints was trained based on historical operating samples, simulation priors, and engineering rules. During training, the objective function consisted of measurement error terms, physical residual terms, parameter smoothing terms, and simulation prior bias terms. During online prediction, the current unified state vector, simulation priors, and historical state sequences consisting of the unified state vectors of the most recent N sampling periods, model outputs, online corrections, and risk indicators are input into the trained data-driven fusion model to obtain the boiler combustion digital twin prediction results. The digital twin prediction results are mapped to measurable quantities through the observation mapping function. Based on the error between the real-time measurement and the measurable quantities, the state estimation, boundary correction coefficient, weight coefficient and simulation prior or surrogate mapping relationship are corrected online at multiple time scales, and the combustion flow field, temperature field, near-wall atmosphere field, water-cooled wall state, comprehensive risk index and risk persistence index are output.

[0009] Preferably, the near-wall atmosphere field data includes one or more concentrations of CO, H2S, SO2, NOx, and O2; the flame visual data includes flame images, flame brightness, flame outline, flame center, flame pulsation, flame deflection, and temperature field obtained from radiation images; and the water-cooled wall state parameters include one or more of the following: water-cooled wall tube temperature, wall temperature change rate, wall surface heat flux density, infrared thermometry results, water-cooled wall tube thickness reduction obtained during maintenance, or wall state parameters estimated from CFD prior and historical data.

[0010] Preferably, the multi-source data preprocessing includes: performing constant flow dilution state verification, filter anti-clogging state judgment, dew point control, range verification, outlier removal, sampling delay estimation, and timestamp verification on near-wall atmosphere measurement data to form atmosphere channel quality factors; performing bad frame removal, image denoising, registration, exposure or saturation verification, background interference suppression, and segmentation confidence evaluation on flame visual data to form image quality factors; performing range verification, step anomaly identification, and interpolation alignment on DCS operating parameters; and performing validity verification, measurement point location matching, and updating hysteresis records on water-cooled wall state parameters; the measurement point quality factor is obtained by weighting multiple atmosphere channel quality factors at the same measurement point and serves as a component of the data quality vector.

[0011] Preferably, when performing temperature field inversion on the flame visual data, the radiation ratio is first constructed based on the radiation intensity of the two temperature measurement bands, and then the apparent temperature of the flame is calculated by combining the emissivity correction coefficient. The inverted temperature field is then corrected by image intensity residual, temperature field smoothing constraint and image quality constraint. The corrected temperature field is used as the flame visual feature or measured / inverted data and input into the data-driven fusion model. The flame deflection index is determined based on the distance between the flame center and the geometric center of the furnace or the target flame center, and the combustion stability index is determined based on the flame brightness variation coefficient, flame area variation coefficient, temperature field fluctuation, flame center position standard deviation, and flame pulsation frequency within a continuous sampling period.

[0012] Preferably, the near-wall component concentration field reconstruction includes: constructing spatial weights based on the spatial distance from the measuring point to the location to be estimated, the distance attenuation coefficient, and the measuring point quality factor; performing weighted interpolation on the concentration of at least one component among CO, H2S, SO2, NOx, and O2 to obtain the near-wall component concentration field; the atmosphere component consistency residual is constructed based on the synergistic change relationship of at least two components among CO, H2S, SO2, and O2 under the same operating conditions, used to identify single component drift or abnormal measuring points, and used as the physical residual term, online correction weight adjustment amount, or risk verification amount; when calculating the reducing atmosphere index, SO2 is used as the sulfur conversion state and component consistency verification term, and the reducing sulfur characterization quantity is jointly determined by H2S, SO2, and operating condition label, and then the reducing atmosphere index is obtained based on CO, the reducing sulfur characterization quantity, and O2.

[0013] Preferably, the CFD simulation platform is the Fluent platform; the simulation prior is generated based on the air volume and damper opening vector, load and thermal state vector, coal mill and burner operation combination, coal quality vector, and CFD model parameters, including one or more of the velocity field, temperature field, component field, particle concentration field, and wall heat flow field.

[0014] Preferably, the physical residual term includes one or more of the following: mass conservation residual, energy conservation residual, component conservation residual, boundary consistency residual, and atmosphere component consistency residual.

[0015] Preferably, the data-driven fusion model employs one of the following: recursive least squares surrogate model, radial basis function surrogate model, Gaussian process surrogate model, or shallow neural network. During modeling, training samples are first constructed using a unified state vector, CFD simulation priors, historical state sequences, and corresponding measured or inverted data. The model parameters are then trained or corrected using an objective function that includes measurement error terms, physical residual terms, parameter smoothing terms, and simulation prior bias terms. During online runtime, the current unified state vector, similar working condition simulation priors, and historical state sequences are input into the trained model for prediction. When on-site data accumulation is insufficient, an initial mapping relationship is formed using the simulation priors and engineering rules.

[0016] Preferably, the multi-timescale online correction includes state estimation and risk index update at a fast scale, boundary correction coefficient and weight coefficient update at a medium scale, and prior library or surrogate mapping relationship update at a slow scale. The risk persistence index is obtained by weighting the comprehensive risk index within the evaluation window according to the attenuation weight, and is used to distinguish between instantaneous risk and persistent risk.

[0017] Another objective of this invention is to provide a digital twin modeling system for boiler combustion.

[0018] To achieve the above objectives, the present invention adopts the following technical solution: A boiler combustion digital twin modeling system includes a multi-source data acquisition module, a data preprocessing module, a flame visual feature extraction module, a near-wall atmosphere field reconstruction module, a CFD simulation prior module, a data-driven fusion modeling module, an online correction module, and a risk output module. The system is used to execute the above-mentioned boiler combustion digital twin modeling method.

[0019] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention unifies the near-wall atmosphere field, flame visual data, DCS operating parameters, and available water-cooled wall state parameters into a state vector, enabling the boiler combustion digital twin model to simultaneously perceive the overall flame state of the furnace and the near-wall reducing atmosphere state, thus improving the problem that traditional monitoring methods can only indirectly reflect the combustion state.

[0020] 2. This invention transforms discrete measurement points and high-dimensional images into computable, fusionable, and online-updable model inputs by reconstructing the atmosphere field, flame deflection index, stability index, and reducing atmosphere index, thereby improving the model's ability to characterize corrosion, coking, flame deflection, and combustion instability.

[0021] 3. This invention provides physical priors using the Fluent basic simulation model, achieves rapid mapping using data-driven models, and uses physical residuals and measured errors to jointly constrain parameter solving or online correction, thus balancing the interpretability of the mechanistic model and the real-time performance of the data-driven model.

[0022] 4. This invention uses an online correction mechanism to update model parameters with real-time atmosphere, visual, and operational data, which can reduce model drift caused by changes in coal quality, load, and air distribution, and improve the wide operating condition adaptability of the boiler combustion digital twin model.

[0023] 5. This invention outputs combustion flow field, temperature field, atmosphere field, wall heat flow field, and comprehensive risk index, which can provide quantitative basis for combustion safety evaluation, corrosion / coking early warning, and subsequent air distribution optimization. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 This is the overall flowchart of the boiler combustion digital twin modeling method of the present invention.

[0026] Figure 2 This is a schematic diagram of the multi-source input and unified state vector of near-wall atmosphere field, flame visual data, DCS operating parameters and state parameters of the present invention.

[0027] Figure 3 This is a schematic diagram of the structure of the CFD basic simulation model and the data-driven fusion model of the present invention.

[0028] Figure 4 This is a flowchart of the online correction closed-loop process of the digital twin model of the present invention.

[0029] Figure 5 This is a schematic diagram illustrating the output and risk assessment of the digital twin model of this invention. Detailed Implementation

[0030] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0031] like Figure 1 As shown, the method for digital twin modeling of boiler combustion, which integrates atmospheric field and flame visual data, includes the following steps: Multi-source data acquisition: Acquire near-wall atmosphere field data, flame visual data, DCS operating parameters, and available water-cooled wall state parameters of coal-fired boilers; Preprocessing includes data cleaning, specifically, anomaly removal, timestamp verification, delay compensation, missing data completion, standardization, and quality evaluation are performed on the near-wall atmosphere field data, flame visual data, DCS operating parameters, and water-cooled wall status parameters to form atmosphere channel quality factors, measuring point quality factors, image quality factors, and water-cooled wall status validity indicators. The preprocessed multi-source data are spatiotemporally synchronized according to a unified sampling period to construct a unified state vector that includes near-wall atmosphere field characteristics, flame visual characteristics, DCS operation characteristics, water-cooled wall state characteristics, operating condition labels, and data quality vectors. The near-wall component concentration field is reconstructed by mass weighting based on the relative spatial relationship of the measuring points, the mass factor of the measuring points, and the component concentration. An atmospheric component consistency residual is constructed, which is used as a physical residual and a risk check quantity to participate in model training, online correction, and risk output. A basic simulation model of the boiler combustion flow field was established using a CFD simulation platform, forming simulation priors indexed by operating condition labels. A data-driven fusion model with physical constraints was trained based on historical operating samples, simulation priors, and engineering rules. During training, the objective function consisted of measurement error terms, physical residual terms, parameter smoothing terms, and simulation prior bias terms. During online prediction, the current unified state vector, simulation priors, and historical state sequences consisting of the unified state vectors of the most recent N sampling periods, model outputs, online corrections, and risk indicators are input into the trained data-driven fusion model to obtain the boiler combustion digital twin prediction results. The digital twin prediction results are mapped to measurable quantities through the observation mapping function. Based on the error between the real-time measurement and the measurable quantities, the state estimation, boundary correction coefficient, weight coefficient and simulation prior or surrogate mapping relationship are corrected online at multiple time scales, and the combustion flow field, temperature field, near-wall atmosphere field, water-cooled wall state, comprehensive risk index and risk persistence index are output.

[0032] Specifically, the method for digital twin modeling of boiler combustion, which integrates atmospheric field and flame visual data, includes the following steps: S1: Multi-source data acquisition: Data on the near-wall atmosphere field, flame visual data, DCS operating parameters, and available water-cooled wall condition parameters of the coal-fired boiler's water-cooled walls are collected. Near-wall atmosphere field data is obtained from the near-wall atmosphere measurement channel, including concentrations of one or more of CO, H2S, SO2, NOx, and O2. Flame visual data is obtained from multispectral or visible light imaging channels, including flame images, flame brightness, flame outline, flame center, flame pulsation, flame deflection, and the temperature field retrieved from radiation images. DCS operating parameters are obtained from the unit control system, including one or more of the following: load, total air volume, primary air volume, secondary air volume, supercombustion air volume, wall-attached air volume, coal quantity, pulverizer combination, oxygen quantity, furnace pressure, main steam pressure, and main steam flow rate. Water-cooled wall condition parameters include one or more of the following: water-cooled wall tube temperature, wall temperature change rate, wall heat flux density, infrared thermography results, water-cooled wall tube thickness reduction obtained during maintenance, or wall condition quantities estimated from CFD prior and historical data. The above data are hierarchically numbered according to the original measured values, derived features, estimated field variables, and risk assessment inputs; among them, the estimated field variables refer to the temperature field, velocity field, component concentration field, wall heat flow field, and water-cooled wall state field estimated on the furnace or near-wall spatial grid, and the sampling time, data channel, spatial location of the measuring point, and operating condition label are retained.

[0033] S2: Preprocessing of near-wall atmosphere data: The high-temperature flue gas component data output from the near-wall atmosphere measurement channel undergoes constant-flow dilution, filtering, anti-clogging, dew point control, range verification, outlier removal, sampling delay estimation, and timestamp verification to generate measurement point number, sampling time, component concentration, and data quality identifier. To enable the subsequent reconstruction process to distinguish between valid and disturbed data, for the g-th atmosphere component channel, an atmosphere channel quality factor is generated based on channel validity, sampling delay, signal fluctuation, and the residual relationship with adjacent time periods.

[0034] Among them, v g (t) represents the channel validity, τ g (t) represents the sampling or transmission delay, n g (t) represents the noise or fluctuation level, ρ g (t) represents the residual characteristics relative to adjacent time periods or similar operating conditions, ψ g For the normalized evaluation function. In one possible implementation, ψ g It is obtained by weighting the effective bits, delay penalty, noise penalty, and residual penalty:

[0035] Where, χ g For the valid bits of the channel, ℓ τ 、ℓ n and ℓ ρ As an empirical weight, `clip` indicates that the results are restricted to the range of 0 to 1. This quality factor is incorporated into subsequent spatiotemporal synchronization, concentration field reconstruction, and online correction processes, reducing the impact of low-confidence data on model updates.

[0036] S3: Flame visual feature extraction: Denoising, registration, background interference suppression, region segmentation, and feature extraction are performed on the flame image sequence to obtain the flame center position, outer contour area, brightness distribution, color or spectral features, pulsation features, stability score, two-dimensional temperature field, and flame deflection direction. Specifically, the flame center position is determined by the weighted centroid of the flame region's pixel intensity; the outer contour area is calculated from the effective pixel area of ​​the segmented flame region; the brightness distribution is characterized by the mean, extreme values, and quantiles of the flame region's grayscale or radiance intensity; the color or spectral features are characterized by the intensity ratios of different bands or RGB channels; the pulsation features are obtained from the brightness, area, or center position spectra of the continuous image sequence; and the stability score is determined by a combination of brightness fluctuations, area fluctuations, temperature fluctuations, and flame center drift. For multispectral images, the radiance ratio is first constructed based on the radiance intensity of the two temperature measurement bands.

[0037] In the emissivity correction factor κ εGiven a value, the apparent temperature T of the flame can be calculated using the following formula:

[0038] Where λ1 and λ2 are two temperature measurement bands, I(λ1,t) and I(λ2,t) are the radiation intensities of the corresponding bands, C2 is the second radiation constant, and κ is the second radiation constant. ε This is used to characterize the emissivity ratio of the two bands and for calibration correction. After obtaining the initial temperature value, temperature field smoothing is performed by combining the flame region segmentation results and image quality indicators to constrain the influence of local saturation, smoke occlusion, or boundary missegmentation on the inversion results.

[0039]

[0040]

[0041] Where Imeas is the measured flame image intensity, Ipred(T) is the predicted image intensity calculated from the temperature field T, LT is the temperature field smoothing operator, Mq is the weighting operator formed by image quality or segmentation confidence, and λT and μT are weighting coefficients given based on historical data and engineering experience.

[0042] S4: Construction of Flame Deflection and Combustion Stability Indicators: Calculate the flame center rf(t) based on the flame image grayscale or radiant intensity I(p,t): rf(t) = Σp pI(p,t) / Σp I(p,t) Where p represents a pixel or spatial position on the image or furnace projection plane. Let r0 be the geometric center of the furnace or the center of the target flame, and D be the characteristic length of the furnace. Then the flame deflection index is: Kf(t) = ||rf(t) - r0|| / D To distinguish between short-term image fluctuations and sustained combustion instability, flame brightness, flame area, temperature field, flame center position, and flame pulsation frequency were statistically analyzed over N consecutive sampling periods. First, a stability perturbation Ss(t) was constructed, and then the combustion stability index Ks(t) was obtained. Ss(t)=b1*CV(IN)+b2*CV(AN)+b3*CV(TN)+b4*STD(rf,N) Ks(t) = 1 - σ[Ss(t)] Wherein, CV represents the coefficient of variation, STD represents the standard deviation, IN is the flame brightness sequence over N consecutive sampling periods, AN is the flame area sequence, TN is the flame temperature sequence, rf,N is the flame center position sequence over N consecutive sampling periods, and b1 to b4 are weighting coefficients given by historical operating data and engineering experience and can be corrected online; in the formula, b1*CV(IN), b2*CV(AN), b3*CV(TN), and b4*STD(rf,N) all represent the multiplication of the weighting coefficients with the corresponding fluctuation statistics. The dominant frequency of flame brightness, low-frequency oscillation energy, and continuous drift direction of the flame center can be used as supplementary inputs to Ss(t) to characterize the continuous skew and combustion instability trend.

[0043] S5: Spatiotemporal Synchronization and Unified State Vector Construction: like Figure 2 As shown, the near-wall atmosphere field data, flame visual data, DCS operating parameters, and available water-cooled wall state parameters are time-aligned according to a unified sampling period Δt. For the i-th data source, the timestamp of its k-th sampling point is denoted as ti,k. At a unified time t, the nearest neighbor sampling sequence number is first determined:

[0044]

[0045]

[0046] Where Ki is the set of valid sampling sequence numbers of the i-th data source, et,i(t) is the synchronization time error, and δt is the maximum allowable synchronization error. For high-frequency data such as flame images, the mean, extreme values, fluctuations, and dominant frequency features are extracted in each sampling period; for DCS continuous variables, nearest neighbor hold or linear interpolation is used to map to a unified time; for near-wall atmosphere field rotation data, alignment is performed according to the measurement point switching order, sampling delay, and valid sampling sequence number; for water-cooled wall infrared temperature measurement, maintenance wall thickness, or other intermittent update data, update lag and synchronization error are retained. After standardization and operating condition coding, a unified state vector is formed:

[0047] Wherein, G'(t) is the standardized near-wall atmosphere field characteristic, formed by the concentration of each component, the rate of change of concentration, the quality factor of the measuring point, the statistics of the near-wall reconstructed concentration field, and the residual of the atmosphere component consistency; V'(t) is the standardized flame visual characteristic, formed by the flame center, outline area, brightness distribution, color or spectral characteristics, pulsation characteristics, two-dimensional temperature field, skew index, and stability index; D'(t) is the DCS operation characteristic; W'(t) is the available water-cooled wall state characteristic, formed by the wall temperature, the rate of change of wall temperature, the wall heat flux density, the infrared thermometry results, the wall thickness reduction during maintenance, the estimated wall state quantity, and its update lag; B(t) is the operating condition label composed of coal type, load range, burner commissioning mode, coal mill combination, etc.; Q'(t) is the data quality vector composed of the atmosphere channel quality factor, the measuring point quality factor, the image quality factor, the update lag, and the synchronization error. For the i-th measuring point, the measuring point quality factor qi(t) can be obtained by weighting the quality factors qi,g(t) of each atmosphere component channel at that measuring point:

[0048] Where, μ g q represents the confidence weight of each component channel. Therefore, q g (t), q i Q(t) and Q'(t) form a progressive relationship from channel quality, measurement point quality to the unified state vector quality component.

[0049] S6: Near-wall atmosphere field reconstruction: Based on the relative spatial relationships of the near-wall measuring points, data quality factors, and the concentrations of each component, the concentration field of the near-wall region of the furnace is reconstructed. For the estimated concentration of component s at spatial location r, spatial weights are first constructed according to distance and data quality, and then the concentrations of the effective measuring points are weighted and summed.

[0050]

[0051]

[0052] Where cs(ri,t) represents the measured or pre-treated concentration of component s at the i-th near-wall measuring point ri, ĉs(r,t) represents the reconstructed concentration of component s at the spatial location r to be estimated, ωi(r,t) represents the spatial interpolation weight of the i-th measuring point at location r; d(r,ri) is the spatial distance, β is the distance attenuation coefficient, εd is the correction amount to avoid zero-distance singularity, and qi(t) is the measuring point quality factor or the corresponding component channel quality factor. In the above formula, ωi(r,t) is obtained by normalizing Ωi(r,t), which considers both measuring point quality and spatial distance, and therefore corresponds to the aforementioned measuring point quality factor qi(t) and the third weight normalization formula. If the measuring point distribution and historical data meet the statistical interpolation conditions, Kriging, Gaussian process regression, or radial basis function interpolation can also be used to achieve reconstruction. To avoid misjudgment of risk caused by single component anomalies, atmosphere consistency residuals are also constructed based on the synergistic change relationship of components such as CO, H2S, SO2, and O2.

[0053] Where Rg(t) is the residual of atmospheric component consistency, Δ represents the change relative to the stable baseline condition, φg is the correlation function formed by combining historical operating data and engineering experience, and B is the condition label. Specifically, Rg(t) is used to determine whether the coordinated changes of H2S, CO, SO2, and O2 conform to the historical patterns under the same load, coal quality, and air distribution method; when Rg(t) exceeds the threshold, the quality factor of the corresponding measuring point or channel can be reduced, online correction weight adjustment can be triggered, and it can be used as a risk verification quantity in the calculation of the comprehensive risk index. In one possible implementation, φg adopts a sliding window regression function or piecewise linear function of the same operating condition, with the change in CO, the change in SO2, and the condition label as input, and outputs an estimated value of the change in H2S. Subsequently, the near-wall reducing atmosphere score Sg(t) is calculated from the reconstructed concentration field, and the reducing atmosphere index Kg(t) is obtained after normalization mapping:

[0054]

[0055]

[0056] Among them, the variables marked with an asterisk are the normalized statistics of the corresponding components in the near-wall region, S red (t) represents the reducing sulfur characterization quantity determined by H2S, SO2 and operating condition labels. SO2 is included in the calculation as a sulfur conversion state and component consistency check item. a1 to a3 are weighting coefficients given by historical operating data, maintenance records and engineering experience and can be corrected online.

[0057] S7: Constructing the basic CFD simulation model: like Figure 3 As shown, based on the furnace structure, burner arrangement, air-coal boundary, coal quality parameters, and operating conditions of the coal-fired boiler, a basic model of the combustion flow field of the entire furnace is established using a CFD simulation platform, with Fluent platform being the preferred choice. This model includes sub-models for turbulence, radiative heat transfer, discrete phase of pulverized coal particles, volatile matter release, coal char reaction, component transport, and boundary condition correction.

[0058] The CFD basic model is calculated offline under typical operating conditions to form a simulation prior database:

[0059] Among them, U f Let Q be the vector of air volume and damper opening. b Let M be the load and thermal state vector, and C be the combination of pulverizer and burner operation. coal Let Θ be the coal quality vector. cfd Y is a parameter of the CFD model. cfd This includes one or more of the following: velocity field, temperature field, component field, particle concentration field, and wall heat flow field. The simulation prior database is indexed by load range, coal quality range, air distribution method, and burner operation combination. During online modeling, snapshots or interpolation results of field variables under similar operating conditions are called as physical priors for the digital twin model.

[0060] S8: Constructing a data-driven fusion model with physical constraints: The unified state vector X(t) and the CFD simulation prior Y are used to... cfd The boiler combustion digital twin prediction results are obtained by inputting the historical state sequence Z(tN:t) and Z(tN:t) into the fusion mapping model:

[0061] Wherein, Fθ is a machine learning, statistical learning, surrogate model, or a combination thereof determined by on-site data conditions, and Z(tN:t) is the historical state sequence of the most recent N sampling periods, specifically including the historical unified state vector X(tN:t-1), the historical digital twin output Ŷ(tN:t-1), the historical online correction error e(tN:t-1), the boundary correction coefficient, the weight coefficient, and the historical risk index. In one possible implementation, Fθ adopts a recursive least squares surrogate model, a radial basis function surrogate model, a Gaussian process surrogate model, or a shallow neural network; in the modeling stage, training samples are first constructed from historical operating data, CFD simulation samples, and engineering rules to perform offline training or initial calibration of the model parameters θ; in the online prediction stage, the trained Fθ is called for rapid inference, and incremental correction is performed based on the real-time error. The digital twin prediction results include at least the furnace temperature field, the near-wall atmosphere field, the combustion flow field, the flame deflection index Kf, the stability index Ks, the reducing atmosphere index Kg, and the risk index. When solving for parameters or performing online correction, the objective function consists of measurement error, physical residuals, parameter smoothing terms, and CFD prior bias:

[0062]

[0063]

[0064]

[0065] Wherein, Ymeas is the measured or inverted data composed of the near-wall atmosphere field, the flame visual temperature field, and the available water-cooled wall state parameters; Ly represents the measurement error term between the predicted result and the measured or inverted data; Lp represents the physical residual term; Lθ represents the parameter smoothing term between the current model parameters and the previous cycle model parameters θprev; Lc represents the prior bias term between the predicted result and the prior Ycfd of the CFD simulation under similar operating conditions; and α, β, γ, and η are weighting coefficients initialized from historical operating data, engineering experience, and expert rules, and can be corrected online. The physical residual is further expressed as:

[0066] Wherein, Rm, Re, Rs, Rb, and Rg represent the mass conservation residual, energy conservation residual, component conservation residual, boundary consistency residual, and atmosphere component consistency residual, respectively. Rm is used to constrain the mass balance of the inlet, outlet, and source terms; Re is used to constrain the energy balance between heat input, radiative heat transfer, convective heat transfer, and wall heat flux; Rs is used to constrain the generation and consumption relationships of components such as O2, CO, H2S, and SO2; Rb is used to constrain the consistency of air volume, coal quantity, wall heat flux, and boundary correction coefficients with the DCS and CFD boundaries; and Rg is used to constrain the coordinated changes of near-wall reducing atmosphere components. These residual terms are used to limit the deviation of the data-driven model from the basic combustion mechanism.

[0067] S9: Online Correction and Model Update: like Figure 4 As shown, the real-time acquired near-wall atmosphere field, flame visual temperature field, and DCS operating parameters enter the online correction module, and are first processed by the observation mapping function M. obs The prediction error is then calculated by mapping the full-field results from the digital twin to measurable quantities.

[0068] The online correction module updates the fusion model parameters, CFD boundary correction coefficients, and risk assessment weights based on the error e(t). In one recursive update method, the parameters are updated according to the following formula:

[0069]

[0070]

[0071] Where Kt is the online gain matrix, Pt-1 is the parameter covariance matrix of the previous period, MobsT is the transpose of the observation mapping function Mobs or the transpose of the linearized observation matrix, Γt is the innovation covariance matrix formed by the observation mapping, parameter covariance, and measurement noise, and R is the measurement noise covariance matrix. Online correction can be performed at three time scales: fast, medium, and slow. The fast scale updates the state estimate and risk index; the medium scale updates the boundary correction coefficients and weight coefficients; and the slow scale updates the prior library or surrogate mapping relationship after accumulating historical data. Recursive least squares, sliding window optimization, filtering correction, or engineering rule correction can all be used to achieve the above online updates.

[0072] S10: Digital Twin Model Output and Application: like Figure 5 As shown, the online-corrected digital twin model outputs the furnace combustion flow field, temperature field, near-wall atmosphere field, wall heat flux field, and risk status. The comprehensive risk is first calculated by combining the reducing atmosphere index, flame deflection index, combustion stability index, wall heat flux, and wall temperature to form a risk score S.r (t), which is then mapped to the comprehensive risk index K. r (t):

[0073]

[0074] Where Kg is the reducing atmosphere index, Kf is the flame deflection index, Ks is the combustion stability index, qw* and Tw* are the normalized wall heat flux density and water-cooled wall temperature, respectively, and c1 to c5 are weighting coefficients given by historical operation records, maintenance records, engineering experience, or expert rules and can be corrected online; when the measured heat flux or wall temperature of the water-cooled wall is unavailable, qw* and Tw* can be estimated from CFD priors, infrared thermography, or historical data. To avoid false alarms caused by transient noise, the risk persistence index is further calculated:

[0075]

[0076]

[0077]

[0078] Where M is the length of the risk persistence assessment window, and λ r K represents the attenuation coefficient. r The model's continuous indicators can be categorized into normal, attention, early warning, and alarm levels; the level thresholds are determined by historical operating data and engineering experience, and are gradually revised as the project's operating data accumulates. The model's output results can be used for combustion safety assessment, corrosion / coking early warning, and subsequent optimization and control interfaces.

[0079] This invention provides a boiler combustion digital twin modeling method that can simultaneously utilize near-wall multi-component atmosphere field, flame visual characteristics, DCS operating parameters, CFD basic simulation results, and online feedback errors. By integrating physical mechanism constraints and data-driven model fusion, it improves the matching degree and real-time performance of online prediction of boiler combustion status, risk assessment, and dynamic model correction. This method can improve the matching degree between the boiler combustion digital twin model and actual operating conditions, as well as its adaptability to a wide range of operating conditions.

[0080] Furthermore, the present invention can solve the following technical problems: Conventional flame detection and DCS parameters are difficult to directly and continuously characterize furnace flame deflection, combustion stability, temperature field and component field changes; near-wall atmosphere field measurement points are discrete, and field data are easily affected by high temperature, high dust, high humidity and strong corrosive environment, making it difficult to model with the overall combustion state of the furnace; traditional CFD simulation is difficult to update quickly under real-time operating conditions, and purely data-driven models lack physical constraints; the drift of model boundary conditions under deep tuning and variable load conditions leads to the problem of decreased matching degree between boiler combustion digital twin model and actual furnace state.

[0081] Example 1: Digital Twin Modeling of Boiler Combustion Based on the Bulian Project This embodiment uses the coal-fired boiler of the Bulian project as the implementation platform. During deep peak shaving and variable load operation, the boiler experiences problems such as flame deflection, enhanced reducing atmosphere near the wall, and increased risk of water-cooled wall corrosion and coking. On-site monitoring of the near-wall atmosphere, visual flame monitoring, and DCS operation data acquisition provide a data foundation for digital twin modeling.

[0082] Reference Figure 2 First, near-wall atmosphere data, flame visual data, DCS operating parameters, and available water-cooled wall condition parameters are collected. The near-wall atmosphere monitoring unit outputs component concentration and data quality indicators; the flame visual monitoring unit outputs flame images, image quality indicators, and feature data; the DCS system outputs operating parameters such as load, air volume, coal volume, oxygen volume, and furnace pressure; and the water-cooled wall condition monitoring or maintenance data outputs wall temperature, wall heat flux, infrared thermography results, maintenance wall thickness reduction, or estimated wall condition parameters. All of the above data, after being managed with a unified timestamp, enters the data preprocessing module.

[0083] Secondly, the data preprocessing module performs outlier removal, missing value handling, standardization, time synchronization, data quality evaluation, and condition label coding on the multi-source data. For atmosphere measurement data, a quality factor is first generated based on channel validity, sampling delay, and fluctuation level, and then used to reconstruct the near-wall concentration field. For flame image data, denoising, segmentation, and feature extraction are first performed, and then the temperature field is inverted based on dual-band radiance ratio and image quality constraints. After the above processing, near-wall atmosphere features, flame visual features, DCS operation features, and condition labels together form a unified state vector X(t).

[0084] Then, a basic simulation model of the boiler for the Bulian project was established using a CFD simulation platform, preferably the Fluent platform. The model generates prior results for the combustion flow field, temperature field, component field, and wall heat flow field under typical operating conditions based on load, coal quality, air distribution method, and burner operation combination, and forms searchable simulation snapshots. During online operation, the digital twin model calls up similar snapshots or interpolation results based on the current operating condition label, allowing the measured data and the prior physical fields to enter the same fusion model.

[0085] Next, the unified state vector X(t) and the CFD simulation prior Y are used. cfd The fusion model is driven by the historical state sequence Z(tN:t) input data to form a digital twin prediction result. The fusion model simultaneously considers measured error, CFD prior bias, parameter smoothing terms, and physical residuals in its objective function. The physical residuals include mass conservation, energy conservation, component conservation, boundary consistency, and atmosphere composition consistency. In this way, the model can both quickly respond to changes in operating conditions using field data and maintain consistency with the basic combustion mechanism of the boiler.

[0086] Finally, the online correction module compares the real-time measurement results with the digital twin prediction results, and updates the model parameters, boundary correction coefficients, and risk weights based on the prediction error. The fast scale is used to update the state estimate and risk indicators, the medium scale is used to correct the boundaries and weights, and the slow scale is used to accumulate historical data and correct prior libraries or surrogate mapping relationships. The model ultimately outputs temperature field, atmosphere field, combustion flow field, flame deflection index, combustion stability index, reducing atmosphere index, instantaneous comprehensive risk index, and risk persistence index, which are used for safety assessment, early warning, and subsequent optimization and control interfaces.

[0087] Example 2 This embodiment is a modified implementation. 1. Atmosphere measurement components can be selected based on coal type, corrosion risk, and environmental control objectives: CO, H2S, SO2, NO. x One or more of O2 and CO2.

[0088] 2. The flame vision model can be implemented using traditional image processing, machine learning models, deep learning models, or a combination thereof. The specific implementation method is determined based on the accumulation of operational data.

[0089] 3. Near-wall atmosphere field reconstruction can be achieved using distance-weighted interpolation, Kriging, Gaussian process regression, radial basis function interpolation, or fusion with CFD prior fields.

[0090] 4. Online correction can employ one or more of the following methods: recursive least squares, sliding window optimization, filtering correction, or engineering rule correction.

[0091] 5. The risk index threshold can be initialized based on historical operating data, maintenance records, corrosion inspection results, and expert experience, and gradually corrected during on-site operation.

[0092] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection, an electrical connection, or a connection that allows communication between them; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components, unless otherwise explicitly limited. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0093] Obviously, the embodiments described above are merely some embodiments of the present invention, not all embodiments. The accompanying drawings show preferred embodiments of the present invention, but do not limit the patent scope of the present invention. The present invention can be implemented in many different forms; rather, these embodiments are provided to provide a more thorough and complete understanding of the disclosure of the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing specific embodiments, or make equivalent substitutions for some of the technical features. Any equivalent structures made using the content of this specification and drawings, directly or indirectly applied to other related technical fields, are similarly within the patent protection scope of this invention.

Claims

1. A method for digital twin modeling of boiler combustion that integrates atmospheric field and flame visual data, characterized in that, Includes the following steps: Collect near-wall atmosphere field data, flame visual data, DCS operating parameters, and available water-cooled wall state parameters of coal-fired boilers; The near-wall atmosphere field data, flame visual data, DCS operating parameters, and water-cooled wall status parameters are respectively subjected to anomaly removal, timestamp verification, delay compensation, missing data completion, standardization, and quality evaluation to form atmosphere channel quality factor, measuring point quality factor, image quality factor, and water-cooled wall status validity identifier. The preprocessed multi-source data are spatiotemporally synchronized according to a unified sampling period to construct a unified state vector that includes near-wall atmosphere field characteristics, flame visual characteristics, DCS operation characteristics, water-cooled wall state characteristics, operating condition labels, and data quality vectors. The near-wall component concentration field is reconstructed by mass weighting based on the relative spatial relationship of the measuring points, the mass factor of the measuring points, and the component concentration. An atmospheric component consistency residual is constructed, which is used as a physical residual and a risk check quantity to participate in model training, online correction, and risk output. A basic simulation model of the boiler combustion flow field was established using a CFD simulation platform, forming simulation priors indexed by operating condition labels. A data-driven fusion model based on historical operating samples, simulation priors and engineering rules is trained to meet physical constraints. During training, the objective function consists of measurement error term, physical residual term, parameter smoothing term and simulation prior bias term. During online prediction, the current unified state vector, simulation priors, and historical state sequences consisting of the unified state vectors of the most recent N sampling periods, model outputs, online corrections, and risk indicators are input into the trained data-driven fusion model to obtain the boiler combustion digital twin prediction results. The digital twin prediction results are mapped to measurable quantities through the observation mapping function. Based on the error between the real-time measurement and the measurable quantities, the state estimation, boundary correction coefficient, weight coefficient and simulation prior or surrogate mapping relationship are corrected online at multiple time scales, and the combustion flow field, temperature field, near-wall atmosphere field, water-cooled wall state, comprehensive risk index and risk persistence index are output.

2. The boiler combustion digital twin modeling method according to claim 1, characterized in that, The near-wall atmosphere field data includes one or more concentrations of CO, H2S, SO2, NOx, and O2. The flame visual data includes flame images, flame brightness, flame outline, flame center, flame pulsation, flame deflection, and temperature field obtained from radiation images. The water-cooled wall state parameters include one or more of the following: water-cooled wall tube temperature, wall temperature change rate, wall heat flux density, infrared thermometry results, water-cooled wall tube thickness reduction obtained during maintenance, or wall state parameters estimated from CFD prior and historical data.

3. The boiler combustion digital twin modeling method according to claim 1, characterized in that, The multi-source data preprocessing includes: performing constant flow dilution state verification, filter anti-clogging state judgment, dew point control, range verification, outlier removal, sampling delay estimation, and timestamp verification on near-wall atmosphere measurement data to form atmosphere channel quality factors; performing bad frame removal, image denoising, registration, exposure or saturation verification, background interference suppression, and segmentation confidence evaluation on flame visual data to form image quality factors; performing range verification, step anomaly identification, and interpolation alignment on DCS operating parameters; and performing validity verification, measurement point location matching, and updating hysteresis records on water-cooled wall state parameters. The measurement point quality factor is obtained by weighting multiple atmosphere channel quality factors at the same measurement point and serves as a component of the data quality vector.

4. The boiler combustion digital twin modeling method according to claim 2, characterized in that, When performing temperature field inversion on the flame visual data, the radiation ratio is first constructed based on the radiation intensity of the two temperature measurement bands, and then the apparent temperature of the flame is calculated by combining the emissivity correction coefficient. The inverted temperature field is then corrected by image intensity residual, temperature field smoothing constraint and image quality constraint. The corrected temperature field is used as the flame visual feature or measured / inverted data and input into the data-driven fusion model. The flame deflection index is determined based on the distance between the flame center and the geometric center of the furnace or the target flame center, and the combustion stability index is determined based on the flame brightness variation coefficient, flame area variation coefficient, temperature field fluctuation, flame center position standard deviation, and flame pulsation frequency within a continuous sampling period.

5. The boiler combustion digital twin modeling method according to claim 1, characterized in that, The near-wall component concentration field reconstruction includes: constructing spatial weights based on the spatial distance from the measuring point to the location to be estimated, the distance attenuation coefficient, and the measuring point quality factor; performing weighted interpolation on the concentration of at least one component among CO, H2S, SO2, NOx, and O2 to obtain the near-wall component concentration field; the atmosphere component consistency residual is constructed based on the synergistic change relationship of at least two components among CO, H2S, SO2, and O2 under the same operating conditions, used to identify single component drift or abnormal measuring points, and used as the physical residual term, online correction weight adjustment amount, or risk verification amount; when calculating the reducing atmosphere index, SO2 is used as the sulfur conversion state and component consistency verification term, and the reducing sulfur characterization quantity is jointly determined by H2S, SO2, and operating condition label, and then the reducing atmosphere index is obtained based on CO, the reducing sulfur characterization quantity, and O2.

6. The boiler combustion digital twin modeling method according to claim 1, characterized in that, The CFD simulation platform is the Fluent platform; the simulation prior is generated based on the air volume and damper opening vector, load and thermal state vector, coal mill and burner operation combination, coal quality vector, and CFD model parameters, including one or more of the velocity field, temperature field, component field, particle concentration field, and wall heat flow field.

7. The boiler combustion digital twin modeling method according to claim 1, characterized in that, The physical residual term includes one or more of the following: mass conservation residual, energy conservation residual, component conservation residual, boundary consistency residual, and atmosphere component consistency residual.

8. The boiler combustion digital twin modeling method according to claim 1, characterized in that, The data-driven fusion model employs one of the following: recursive least squares surrogate model, radial basis function surrogate model, Gaussian process surrogate model, or shallow neural network. During modeling, training samples are first constructed using a unified state vector, CFD simulation priors, historical state sequences, and corresponding measured or inverted data. The model parameters are then trained or corrected using an objective function that includes measurement error terms, physical residual terms, parameter smoothing terms, and simulation prior bias terms. During online runtime, the current unified state vector, similar operating condition simulation priors, and historical state sequences are input into the trained model for prediction. When there is insufficient on-site data, an initial mapping relationship is formed based on the simulation priors and engineering rules.

9. The boiler combustion digital twin modeling method according to claim 1, characterized in that, The multi-timescale online correction includes state estimation and risk index update at a fast scale, boundary correction coefficient and weight coefficient update at a medium scale, and prior library or surrogate mapping relationship update at a slow scale. The risk persistence index is obtained by weighting the comprehensive risk index according to the decay weight within the evaluation window, and is used to distinguish between instantaneous risk and persistent risk.

10. A digital twin modeling system for boiler combustion, characterized in that, The system includes a multi-source data acquisition module, a data preprocessing module, a flame visual feature extraction module, a near-wall atmosphere field reconstruction module, a CFD simulation prior module, a data-driven fusion modeling module, an online correction module, and a risk output module. The system is used to execute the boiler combustion digital twin modeling method according to any one of claims 1 to 9.

Images