Coal combustion with ammonia injection prediction method based on simulation data augmentation and multi-fidelity modeling

Abstract

The present application relates to a kind of coal ammonia combustion prediction methods based on simulation data enhancement and multi-fidelity modeling, comprising: the full-process simulation model of coal ammonia combustion process is built, simulation generates the state parameter data covering different working conditions of coal ammonia combustion, obtains the low-fidelity simulation dataset covering wide operating condition;Using low-fidelity simulation dataset, with the loss function of lower bound loss under evidence, train the Bayesian neural network as pre-training model, it is used to output the predicted value of state parameter;A small amount of high-fidelity experimental dataset is obtained by experiment;High-fidelity experimental dataset and state parameter predicted value are input into calibration model simultaneously, to excavate the coupling mapping relationship of simulation and experimental data, introduce the monotonicity regular penalty term of combustion as mechanism constraint to build total loss function and obtain prediction model;The present application solves the technical problems that the existing method is distorted for coal ammonia mixed combustion extreme condition prediction, weak model interpretability, poor small sample generalization.

Description

Technical Field

[0001] This invention relates to the field of intelligent prediction technology for coal combustion with ammonia, and in particular to a method for predicting coal combustion with ammonia based on simulation data augmentation and multi-fidelity modeling. Background Technology

[0002] Co-combustion of coal and ammonia in coal-fired power units, as a low-carbon retrofit technology that can simultaneously reduce carbon and nitrogen oxide emissions, has become a key research direction in the thermal power industry. Co-combustion of coal and ammonia involves complex physicochemical processes such as multiphase flow combustion and nitrogen oxide generation. Its flue gas composition distribution and boiler thermal efficiency are simultaneously affected by multiple factors such as coal quality, ammonia blending ratio, excess air coefficient, and unit load. The evolution of operating conditions is complex, which places extremely high demands on the global adaptability of prediction models.

[0003] Currently, the two main methods for collecting data on coal-ammonia co-combustion both face practical bottlenecks: First, building a full-condition industrial test platform is costly and time-consuming, and the total number of stable and effective measured samples is very limited due to unit operation and scheduling constraints. Second, while process simulation tools can generate a large number of operating condition samples in batches, the simulation process simplifies complex reactions, and the built-in simplified assumptions on physical properties and kinetics can cause systematic deviations between the simulation output and the actual on-site operating data. Modeling based solely on simulation data is difficult to match the actual operating conditions of the unit.

[0004] While current general multi-source data fusion modeling approaches can utilize both simulation and measured samples simultaneously, they suffer from the following limitations in complex coal-fired ammonia-blended reaction scenarios: First, the training method fails to differentiate between the massive simulation samples and the scarce measured samples in terms of data volume and true value accuracy. A uniform parameter update method can easily overwrite the underlying combustion reaction characteristics learned from simulation. Second, the loss function simply fits the numerical error without incorporating the combustion mechanism regularity constraint that CO2, water vapor, and thermal efficiency change monotonically with the ammonia blending ratio. This results in some operating conditions where the predicted trends violate the basic reaction laws, leading to poor physical interpretability of the model. Third, the correction module uses fixed weights to balance linear and nonlinear biases, failing to adaptively capture the dynamic nonlinear reaction residuals caused by ammonia oxidation. This significantly reduces the prediction accuracy for high ammonia blending and low-load boundary conditions. Furthermore, the network model only outputs a single, definitive predicted value, making it impossible to quantify the prediction confidence interval and determine the reliability of results under extreme ammonia blending and deep peak-shaving conditions, resulting in insufficient reliability for engineering applications. Fourth, a hierarchical parameter update strategy adapted to small sample scenarios is not designed. Fine-tuning with a small amount of measured data can easily destroy the wide-range operating condition characteristics learned from full-domain simulation.

[0005] Therefore, there is an urgent need to develop a multi-fidelity prediction modeling method for coal-ammonia co-combustion products and efficiencies that can reduce dependence on measured samples, integrate the advantages of simulation and experimental data, have combustion mechanism constraints, and possess high accuracy, generalization ability, and interpretability. This method would address engineering pain points such as prediction distortion under extreme coal-ammonia co-combustion conditions, weak model interpretability, and poor generalization under small sample conditions, and provide reliable data support for real-time combustion prediction and carbon pollution co-operation condition optimization of power units. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a method for predicting coal-ammonia blended combustion based on simulation data augmentation and multi-fidelity modeling, in order to solve the technical problems of existing methods such as inaccurate predictions under extreme coal-ammonia blended combustion conditions, weak model interpretability, and poor generalization under small sample conditions.

[0007] The technical solution adopted in this invention is as follows: This invention provides a method for predicting coal combustion with ammonia blending based on simulation data augmentation and multi-fidelity modeling, comprising: A full-process simulation model of the coal-fired ammonia-blended combustion process was constructed. Within the set range of boiler operating parameters, state parameter data covering coal-fired ammonia-blended combustion under different operating conditions were generated through simulation, resulting in a low-fidelity simulation dataset covering a wide range of operating conditions. Based on the low-fidelity simulation dataset, and using the evidence lower bound loss as the loss function, a Bayesian neural network is trained as a pre-trained model, which takes the running parameters as input and outputs the predicted values ​​of the state parameters. Set experimental conditions within the simulation range and obtain a high-fidelity experimental dataset through experiments, which includes measured operating parameters and corresponding state parameters. The high-fidelity experimental dataset and the predicted values ​​of the state parameters are simultaneously input into the calibration model, which is used to fit the linear term in the coupling mapping relationship between simulation data and experimental data, capture the characteristic patterns of the nonlinear term in the coupling mapping relationship, obtain the linear deviation correction amount and the nonlinear deviation correction amount, and output the final predicted value of the state parameters after weighted fusion; at the same time, a combustion monotonicity regularization penalty term is introduced as a mechanism constraint to construct the total loss function, optimize the parameters of the calibration model, and obtain the prediction model; Combustion monotonicity regularization penalty term The calculations include:

[0008] In the formula, The input vector variables of the model represent the variable space of continuous operating conditions; for The sample point in the input represents the first input point. Group working condition sample; N This represents the total number of operating conditions. , , These are the predicted values ​​of CO2 volume fraction, water vapor volume fraction, and boiler efficiency output by the pre-trained model, respectively. The ReLU trimmer is used to generate a non-zero penalty only when the gradient violates the combustion law.

[0009] The preferred technical solution is: The calibration model employs an adaptive mechanism to establish a coupled mapping relationship between low-fidelity simulation data and high-fidelity experimental data, including: The characteristic laws of the linear and nonlinear terms are fitted using linear and nonlinear calibration networks, respectively:

[0010] In the formula, To calibrate the high-fidelity predicted state parameters output by the model; These are the predicted values ​​of the state parameters output by the pre-trained model. , These are the measured operating parameters and corresponding state parameters in the high-fidelity experimental dataset; This is a function that maps simulation data to experimental data. and The linear deviation correction amount output by the linear calibration network and the nonlinear deviation correction amount output by the nonlinear calibration network are respectively calculated. β As a weighting factor; Adaptive dynamic updating of weight gradients includes:

[0011] In the formula, Represents the total loss function Weighting factors The gradient.

[0012] The construction of the total loss function includes:

[0013] In the above formula: , is the monotonicity penalty weight coefficient. Loss of the lower bound of evidence:

[0014] In the above formula, For variational expectation; For high-fidelity experimental datasets The likelihood term represents the likelihood of a given model network weights. Below, the model generates information about... The logarithmic probability; For weight Approximate variational posterior distribution Compared with the preset prior distribution The KL divergence.

[0015] A full-process simulation model of the coal-ammonia blending combustion process was built using the Aspen Plus system. The RYIELD module and Fortran subroutines were used to simulate the non-equilibrium staged pyrolysis of pulverized coal. The RGIBBS module was used to simulate the complete reaction equilibrium state of cracked gas, air and ammonia at high temperature based on the principle of minimizing Gibbs free energy. The composition of the final products after the coal or cracked residue is mixed with ammonia and combusted was solved by combining the conservation of all elements.

[0016] The full-process simulation model couples the mixed combustion process with the heat exchange system to achieve synchronous output of combustion products, boiler efficiency, and steam parameters.

[0017] The construction of the low-fidelity simulation dataset covering a wide range of operating conditions includes: Based on the principle of equal heat ammonia blending, multiple ammonia blending ratios are set within a preset wide operating range, and the corresponding coal feed rate, ammonia volume, excess air coefficient, and simulation values ​​of state parameters obtained from simulation calculations are matched. After data preprocessing, a sample set covering a wide range of operating conditions is formed.

[0018] The state parameters include combustion product composition, product mass flow rate, and boiler efficiency.

[0019] The operating parameters include coal feed rate, ammonia flow rate, input air flow rate, or excess air coefficient.

[0020] Perform monotonicity verification on the prediction curve of the trained prediction model, including: Under the condition of constant coal feed rate, the predicted value output by the model passes the validation when the following conditions are met: The predicted volume fraction of CO2 in the combustion products of coal with ammonia increases monotonically with increasing ammonia content, the predicted volume fraction of H2O increases monotonically with increasing ammonia content, and the predicted boiler efficiency decreases monotonically with increasing ammonia content.

[0021] The high-fidelity experimental dataset contains no more than 10 sample groups, while the low-fidelity simulation dataset contains far more than 10 sample groups.

[0022] The technical solution of the present invention can achieve at least some of the following beneficial effects: This invention first constructs a simulation system model of ammonia-blended coal combustion process. By adjusting boiler operating parameters, a low-fidelity dataset covering a wide range of operating conditions and sufficient data is obtained, but with low accuracy. A Bayesian neural network (BNN) is pre-trained based on the massive low-fidelity samples, embedding Bayesian inference and uncertainty quantification mechanisms. Then, high-fidelity data with high accuracy but scarce data is collected from actual industrial measurements. Based on the high-fidelity samples and the predicted values ​​of the pre-trained model, a composite multi-fidelity Bayesian neural network (MF-BNN) is constructed. This network incorporates hierarchical thawing differential training, a loss function with a built-in switchable combustion monotonicity regularization penalty term, and adaptive dynamic allocation of branch fusion weights. A dual-branch structure for linear steady-state deviation correction and nonlinear response residual correction is built. By fusing the two types of data, multi-fidelity joint modeling is completed, resulting in the target prediction model. This invention can accurately predict the multi-component products and efficiency of coal-fired ammonia-blended combustion under extreme operating conditions where experimental data is sparse or uncovered (requiring only 10 or fewer industrial test samples). The prediction results strictly follow the combustion reaction law, exhibit strong generalization performance and outstanding physical interpretability over a wide operating range, and provide reliable data support for combustion regulation and synergistic carbon emission reduction of coal-fired ammonia-blended units. Attached Figure Description

[0023] Figure 1 This is a flowchart illustrating the method of an embodiment of the present invention.

[0024] Figure 2 This is a schematic diagram of the full-process simulation model of coal-fired ammonia-blended combustion based on Aspen Plus, according to an embodiment of the present invention.

[0025] Figure 3 This is a schematic diagram of the model architecture and training of the prediction model in an embodiment of the present invention.

[0026] Figure 4 This is a comparison chart showing the results of the prediction and comparison models of the present invention for the predicted and actual values ​​of CO2 in flue gas from coal-fired ammonia-blended flue gas. Detailed Implementation

[0027] The specific embodiments of the present invention are described below with reference to the accompanying drawings.

[0028] See Figures 1 to 3 This embodiment of a method for predicting coal combustion with ammonia based on simulation data augmentation and multi-fidelity modeling includes: S1. Construct a full-process simulation model of coal-fired ammonia-blended combustion. Within the set range of boiler operating parameters, simulate and generate state parameter data covering coal-fired ammonia-blended combustion under different operating conditions to obtain a low-fidelity simulation dataset covering a wide range of operating conditions.

[0029] As a specific method, the physical architecture of the full-process simulation model for coal-fired ammonia-blended combustion is as follows: Figure 2As shown, this full-process simulation model is built based on the Aspen Plus system, coupling the three major systems of coal-ammonia co-combustion, flue gas heat exchange, and steam power generation within the same process, without separation or disconnection of assumptions. The PR-BM property calculation method is preferably adopted to adapt to the coal-ammonia-air multi-component system. For unconventional coal components, the HCOAKGEB model and DCOALIGT module are used to calculate enthalpy and density. A complete set of steady-state assumptions is proposed based on the real conditions of industrial boilers to ensure that the simulation conforms to reaction equilibrium and closely approximates engineering reality, guaranteeing the physical rationality of the dataset from the source. Specifically, to fit the steady-state engineering boundary conditions and simulation computational load, the following model assumptions are optimized: The reaction process of coal-ammonia co-combustion in the boiler is divided into two steady-state stages: pyrolysis and combustion, and modeled accordingly. The coal powder particles are regular in shape and have a uniform internal temperature. In coal, H, N, O, and S elements undergo complete gas-phase transformation, while C element is not completely oxidized; There is no pressure loss in the furnace; Ash is an inert component and does not participate in the reaction; The flue gas inside the furnace is fully mixed, and the gas phase reaction quickly reaches equilibrium.

[0030] As a preferred approach, this embodiment utilizes the RYIELD module (yield reactor) of the Aspen Plus system and a Fortran subroutine to simulate the non-equilibrium staged pyrolysis of pulverized coal; it uses the RGIBBS module (BURN module) based on the principle of minimizing Gibbs free energy to simulate the complete reaction equilibrium state of cracked gas, air, and ammonia at high temperature, and combines the conservation of all elements to solve the final product composition after the combustion of coal or cracked residue with ammonia; the mixed combustion process is coupled with the heat exchange system to achieve the synchronous output of combustion products, boiler efficiency, and steam parameters.

[0031] Specifically, the sensitivity analysis tool built into Aspen Plus is used to set the variation range of operating parameters such as coal feed rate, ammonia flow rate, input air flow rate or excess air coefficient, and at least one of the parameters is systematically changed within the preset range to automatically generate a large number of state parameters covering different operating conditions. In this embodiment, the preferred parameters are the composition of coal-fired ammonia combustion products, product mass flow rate and boiler efficiency data.

[0032] As a preferred method, following the principle of equal-heat ammonia blending, multiple ammonia blending ratios are set within a preset wide operating range, matching the corresponding coal feed rate, ammonia volume, excess air coefficient, and simulation values ​​of state parameters obtained from simulation calculations. After data preprocessing, a low-fidelity simulation dataset covering the wide operating conditions is formed. ,in , These represent multi-dimensional operating parameters, the mass flow rate of various combustion products, and the simulated boiler efficiency values, respectively.

[0033] The ammonia blending ratio is defined as the proportion of ammonia calorific value to the total heat input of the boiler, while keeping the total heat input constant before and after ammonia blending. Within a wide operating range, with a fixed coal feed rate, as the ammonia blending ratio increases, the volume fraction of CO2 in the flue gas gradually decreases, the volume fraction of H2O gradually increases, and the boiler efficiency slightly decreases.

[0034] S2. Based on the low-fidelity simulation dataset, and using the Evidence Lower Bound (ELBO) loss as the loss function, train a Bayesian neural network as a pre-trained model, which takes the running parameters as input and outputs the predicted values ​​of the state parameters.

[0035] Specifically, the Bayesian neural network employs n Dimensional input layer m Dimensional output layer and L A layered hidden layer architecture. The network weights follow a prior probability distribution, and the posterior distribution is approximated through variational inference. When the input vector... At that time, during the forward propagation process, the first l Hidden layers (1≤ l The output of ≤L) is characterized by a weighted probability distribution:

[0036] In the formula: , The first The first hidden layer, the... The output of each hidden layer; for( ) layer and Interlayer weight matrix, Indicates weight Follows a certain probability distribution ; For bias terms; is the ReLU activation function. The predicted probability distribution output by the pre-trained model is:

[0037] In the formula, This represents the predicted probability distribution output by the model. The predicted value output by the model; Given input; For deterministic mapping functions of neural networks; It is an approximate distribution of the variational posterior distribution; For weights.

[0038] Specifically, during the pre-training phase, Bayesian backpropagation is used to iteratively optimize the probability distribution of network parameters, simultaneously outputting predicted values ​​and uncertainty intervals. The loss function during the pre-training phase... Represented as:

[0039] In the formula: For variational expectation; The likelihood term represents the likelihood given the weights. In this case, the model generates observation data. The logarithmic probability; Let KL divergence be the variational posterior and the true posterior. It is an approximate distribution of the variational posterior distribution; Let be the prior distribution of the weights.

[0040] S3. Set the experimental conditions within the simulation range and obtain a high-fidelity experimental dataset through experiments, which includes the measured operating parameters and corresponding state parameters.

[0041] Specifically, the small-scale high-fidelity experimental dataset obtained from the on-site operation of coal-fired ammonia blending is denoted as... ,in The parameters for isothermal ammonia gradient operation include the actual measured values ​​of operating parameters such as coal feed rate, ammonia flow rate, and input air flow rate. This represents the true values ​​of various components of combustion products, mass flow rates, and boiler efficiency measured on-site. The experimental data sample is preferably controlled to be no more than 10 sets.

[0042] S4. Simultaneously input the high-fidelity experimental dataset and the predicted values ​​of the state parameters into the calibration model for multi-fidelity calibration training; the calibration model is used to fit the linear term in the coupling mapping relationship between simulation data and experimental data, capture the characteristic patterns of the nonlinear term in the coupling mapping relationship, obtain the linear deviation correction and nonlinear deviation correction, and output the final predicted values ​​of the state parameters after weighted fusion. Simultaneously, a combustion monotonicity regularization penalty term is introduced as a mechanism constraint to construct the total loss function, and the parameters of the calibration model are optimized to obtain the prediction model. Among them, the combustion monotonicity regularization penalty term The calculations include:

[0043] In the formula, The input vector of the model represents the variable space of continuous operating conditions, including operating parameters such as coal feed rate, ammonia flow rate, and air flow rate. for The sample point in the input represents the first input point. Group working condition sample;N This represents the total number of operating conditions. , , These are the predicted values ​​of CO2 volume fraction, water vapor volume fraction, and boiler efficiency output by the pre-trained model, respectively. The ReLU trimmer is used to generate a non-zero penalty only when the gradient violates the combustion law.

[0044] As a preferred approach, the calibration model employs an adaptive mechanism to explore the coupling mapping relationship between low-fidelity simulation data and high-fidelity experimental data, including: (1) Linear calibration networks (N) are used respectively. E1 ) and nonlinear calibration network (N E2 Perform a feature fit between the linear and nonlinear terms:

[0045] In the formula, To calibrate the high-fidelity predicted state parameters output by the model; These are the predicted values ​​of the state parameters output by the pre-trained model. , These are the measured operating parameters and corresponding state parameters in the high-fidelity experimental dataset; This is a function that maps simulation data to experimental data. and The linear deviation correction amount of the linear calibration network output and the nonlinear deviation correction amount of the nonlinear calibration network output are respectively used, with β being a weighting factor, which is adaptively and dynamically updated with the uncertainty of model training and Bayesian output. (2) Adaptive dynamic update of weight gradients, including:

[0046] In the formula, Represents the total loss function Weighting factors The gradient.

[0047] Specifically, the Adam optimizer is used to iteratively update the weights: , The learning rate is used during the multi-fidelity calibration phase; when the residual of the ammonia oxidation nonlinear reaction is large and the model prediction confidence interval is wide, gradient-driven learning is employed. Automatically reduce and increase the weight of nonlinear calibration branches; when steady-state thermodynamic deviation dominates. It automatically increases and enhances the linear deviation correction effect.

[0048] Step S4 in this embodiment essentially involves constructing a composite multi-fidelity Bayesian neural network (MF-BNN) model, including linear and nonlinear calibration networks, based on the pre-trained model. This combines low-fidelity simulation data obtained from Aspen Plus simulations with high-fidelity experimental data to establish a coupling mapping relationship between the low-fidelity simulation domain and the high-fidelity experimental domain, thus constructing a unified multi-fidelity prediction model and improving the model's prediction accuracy. Specifically, the linear and nonlinear calibration networks are used to establish a cross-fidelity coupling mapping relationship. This combines the sufficient but less accurate low-fidelity data obtained from Aspen Plus simulations with the high-accuracy but scarce high-fidelity data obtained from actual operation, respectively correcting the inherent biases of the steady-state thermodynamic system and the nonlinear residuals of the dynamic combustion reaction. This establishes a dedicated feature mapping space for the low-fidelity simulation domain and the high-fidelity experimental domain of coal-fired ammonia-blended combustion, thereby significantly improving the accuracy and physical consistency of component prediction under multiple operating conditions.

[0049] The trained model contains three core neural network components: a pre-trained network N that extracts features based on simulation data. S Used to obtain input parameters Output simulation prediction values Two calibration networks, N with linear structure, are responsible for multi-fidelity domain bias correction. E1 With nonlinear structures N E2 These are used to fit the mapping relationship between simulation data and experimental data, respectively. Linear terms Capturing the mapping relationship between simulation data and experimental data Nonlinear terms The characteristic patterns. The predicted final state parameters of the composite multi-fidelity Bayesian neural network model. By N E1 and N E2 The output is weighted and synthesized.

[0050] In terms of the training mechanism, this embodiment of the model addresses the inherent differences between simulation datasets and industrial test datasets in terms of sample size and ground truth accuracy by setting a two-stage tiered unfreezing training process: The simulation consists of a pre-training phase for combustion simulation and a multi-fidelity calibration phase incorporating combustion monotonicity regularization penalties. Both phases use a composite evidence lower bound (ELBO) with superimposed combustion monotonicity regularization penalties as the loss function. The penalty coefficient for the pre-training phase is... Using only basic ELBO, multiple-fidelity calibration stages are taken Introduction Based on Bayesian variational inference, the posterior distribution of network weights is updated iteratively, taking into account both global feature learning and combustion physics constraints, and suppressing network overfitting caused by a small number of experimental samples.

[0051] Specifically, the composite evidence lower bound (ELBO) employing superimposed combustion monotonicity regularization penalty terms is the mathematical expression of the loss function, i.e., the total loss function:

[0052] In the above formula: Monotonicity penalty weighting coefficient; lower bound loss of evidence during multi-fidelity calibration stage. Represented as:

[0053] in, For variational expectation; For high-fidelity experimental datasets The likelihood term represents the likelihood of a given model network weights. Below, the model generates information about... The logarithmic probability; For weight Approximate variational posterior distribution Compared with the preset prior distribution The KL divergence.

[0054] As a preferred method, the monotonicity of the prediction curve of the trained prediction model is checked to verify the consistency between the model's prediction patterns and the combustion mechanism, including: Under the condition of constant coal feed rate, the predicted value output by the model passes the validation when the following conditions are met: The predicted volume fraction of CO2 in the combustion products of coal with ammonia increases monotonically with increasing ammonia content, the predicted volume fraction of H2O increases monotonically with increasing ammonia content, and the predicted boiler efficiency decreases monotonically with increasing ammonia content.

[0055] To verify the practical application effect of the method in this embodiment, typical thermal coal was selected, and its industrial and elemental analysis properties were measured. The Aspen Plus coal-ammonia co-combustion full-process simulation model built in this embodiment was used, and a large-scale low-fidelity simulation dataset was generated through sensitivity analysis. Based on this dataset, a Bayesian neural network was pre-trained. Then, using a small number of high-fidelity experimental samples, a composite multi-fidelity Bayesian neural network was used to perform linear and nonlinear collaborative multi-fidelity bias calibration to obtain the final prediction model. The coefficient of determination (R²) was used. 2 The root mean square error (RMSE) and the root mean square error (RMSE) are used as evaluation metrics to verify the accuracy of target prediction.

[0056] Specifically, eight sets of industrial test samples were selected for model calibration, with all training features derived from massive low-fidelity data generated by Aspen Plus simulation; the remaining industrial test samples were used as a test set for performance verification. A comparative experiment was conducted between the composite multifidelity Bayesian neural network (MF-BNN) of this embodiment, which features layered thawing, a loss function with a built-in switchable combustion monotonicity regularization penalty term, and branch fusion weights based on Bayesian uncertainty adaptive updates, and traditional DNN, SVR, and XGBoost. The predicted and actual values ​​of CO2 in flue gas for each model were compared as follows: Figure 4 As shown. By Figure 4 As can be seen, the prediction results of the MF-BNN model in this embodiment are highly consistent with the actual values, and the fit is significantly better than that of the SVR, DNN, and XGBoost models. The coefficient of determination R for CO2 prediction is particularly high. 2 It can reach 0.98, and the root mean square error (RMSE) is only 0.087, with prediction accuracy far exceeding that of traditional machine learning models.

[0057] It will be understood by those skilled in the art that the above description is merely a preferred embodiment of the present invention and is not intended to limit 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 embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting coal combustion with ammonia blending based on simulation data augmentation and multi-fidelity modeling, characterized in that, include: A full-process simulation model of the coal-fired ammonia-blended combustion process was constructed. Within the set range of boiler operating parameters, state parameter data covering coal-fired ammonia-blended combustion under different operating conditions were generated through simulation, resulting in a low-fidelity simulation dataset covering a wide range of operating conditions. Based on the low-fidelity simulation dataset, and using the evidence lower bound loss as the loss function, a Bayesian neural network is trained as a pre-trained model, which takes the running parameters as input and outputs the predicted values ​​of the state parameters. Set experimental conditions within the simulation range and obtain a high-fidelity experimental dataset through experiments, which includes measured operating parameters and corresponding state parameters. The high-fidelity experimental dataset and the predicted values ​​of the state parameters are simultaneously input into the calibration model, which is used to fit the linear term in the coupling mapping relationship between simulation data and experimental data, capture the characteristic patterns of the nonlinear term in the coupling mapping relationship, obtain the linear deviation correction amount and the nonlinear deviation correction amount, and output the final predicted value of the state parameters after weighted fusion; at the same time, a combustion monotonicity regularization penalty term is introduced as a mechanism constraint to construct the total loss function, optimize the parameters of the calibration model, and obtain the prediction model; Combustion monotonicity regularization penalty term The calculations include: ， In the formula, The input vector variables of the model represent the variable space of continuous operating conditions; for The sample point in the input represents the first input point. Group working condition sample; N This represents the total number of operating conditions. , , These are the predicted values ​​of CO2 volume fraction, water vapor volume fraction, and boiler efficiency output by the pre-trained model, respectively. The ReLU trimmer is used to generate a non-zero penalty only when the gradient violates the combustion law.

2. The method according to claim 1, characterized in that, The calibration model employs an adaptive mechanism to establish a coupled mapping relationship between low-fidelity simulation data and high-fidelity experimental data, including: The characteristic laws of the linear and nonlinear terms are fitted using linear and nonlinear calibration networks, respectively: ， In the formula, To calibrate the high-fidelity predicted state parameters output by the model; These are the predicted values ​​of the state parameters output by the pre-trained model. , These are the measured operating parameters and corresponding state parameters in the high-fidelity experimental dataset; This is a function that maps simulation data to experimental data. and The linear deviation correction amount output by the linear calibration network and the nonlinear deviation correction amount output by the nonlinear calibration network are respectively calculated. β As a weighting factor; Adaptive dynamic updating of weight gradients includes: ， In the formula, Represents the total loss function Weighting factors The gradient.

3. The method according to claim 1, characterized in that, The construction of the total loss function includes: ， In the above formula: , is the monotonicity penalty weight coefficient. Loss of the lower bound of evidence: ， In the above formula, For variational expectation; For high-fidelity experimental datasets The likelihood term represents the likelihood of a given model network weights. Below, the model generates information about... The logarithmic probability; For weight Approximate variational posterior distribution Compared with the preset prior distribution The KL divergence.

4. The method according to claim 1, characterized in that, A full-process simulation model of the coal-ammonia blending combustion process was built using the Aspen Plus system. The RYIELD module and Fortran subroutines were used to simulate the non-equilibrium staged pyrolysis of pulverized coal. The RGIBBS module was used to simulate the complete reaction equilibrium state of cracked gas, air and ammonia at high temperature based on the principle of minimizing Gibbs free energy. The composition of the final products after the coal or cracked residue is mixed with ammonia and combusted was solved by combining the conservation of all elements.

5. The method according to claim 4, characterized in that, The full-process simulation model couples the mixed combustion process with the heat exchange system to achieve synchronous output of combustion products, boiler efficiency, and steam parameters.

6. The method according to claim 1, characterized in that, The construction of the low-fidelity simulation dataset covering a wide range of operating conditions includes: Based on the principle of equal heat ammonia blending, multiple ammonia blending ratios are set within a preset wide operating range, and the corresponding coal feed rate, ammonia volume, excess air coefficient, and simulation values ​​of state parameters obtained from simulation calculations are matched. After data preprocessing, a sample set covering a wide range of operating conditions is formed.

7. The method according to claim 1, characterized in that, The state parameters include combustion product composition, product mass flow rate, and boiler efficiency.

8. The method according to claim 1, characterized in that, The operating parameters include coal feed rate, ammonia flow rate, input air flow rate, or excess air coefficient.

9. The method according to claim 1, characterized in that, Perform monotonicity verification on the prediction curve of the trained prediction model, including: Under the condition of constant coal feed rate, the predicted value output by the model passes the validation when the following conditions are met: The predicted volume fraction of CO2 in the combustion products of coal with ammonia increases monotonically with increasing ammonia content, the predicted volume fraction of H2O increases monotonically with increasing ammonia content, and the predicted boiler efficiency decreases monotonically with increasing ammonia content.

10. The method according to claim 1, characterized in that, The high-fidelity experimental dataset contains no more than 10 sample groups, while the low-fidelity simulation dataset contains far more than 10 sample groups.

Images