Nitrogen oxide emission prediction and optimization method for ammonia-hydrogen combustion chamber of gas turbine

Abstract

The invention discloses a method for predicting and optimizing nitrogen oxide emission of an ammonia-hydrogen combustion chamber of a gas turbine, and the method comprises the steps: carrying out the data fusion and preprocessing according to the experimental data and simulation data of the nitrogen oxide emission of the ammonia-hydrogen combustion chamber; constructing a physical information neural network model; training a physical information neural network model according to the training data set; analyzing a chemical reaction path of the nitrogen oxide according to the physical information neural network model; and optimizing the characteristic parameters of the ammonia-hydrogen combustion chamber based on the path contribution degree according to the physical information neural network model. According to the method, the PINN model is established by utilizing experiment and simulation data, the conservation constraint and the reaction path constraint are set, so that the prediction precision of the model on the nitrogen oxide emission is improved, the efficiency and the physical interpretability of the model are improved, the contribution degree of the reaction path is quantified by utilizing the PINN model, the combustion characteristic parameters are quickly searched and optimized, and the efficiency and the physical interpretability of the model are improved. The calculation time is shortened, and the optimization efficiency of the emission prediction of the nitric oxide in the ammonia-hydrogen combustion chamber is improved.

Description

Technical Field

[0001] This application relates to a method for predicting and optimizing nitrogen oxide emissions from the ammonia-hydrogen combustion chamber of a gas turbine, belonging to the field of gas turbine technology. Background Technology

[0002] In the current application context, using ammonia-hydrogen as fuel in gas turbines is one of the important research directions for the clean energy transition, contributing to the achievement of low carbon emissions. The application of ammonia-hydrogen combustors in gas turbines is also an important technological path to achieve carbon neutrality. However, ammonia-hydrogen fuel still poses the problem of nitrogen oxide emissions, therefore, nitrogen oxide emission prediction is crucial when designing ammonia-hydrogen combustors.

[0003] Current CFD simulations for predicting nitrogen oxide emissions often rely on detailed chemical reaction mechanisms, resulting in high computational costs and difficulty in coupling with physical fields such as temperature, velocity, and pressure. Methods based on purely data-driven neural network models (such as convolutional neural networks and recurrent neural networks) often lack physical interpretability and exhibit poor generalization ability with small sample sizes. Traditional univariate experimental methods are also problematic due to the complex chain reactions of NH2 / H / NH radicals in ammonia-hydrogen fuels and the strong coupling between the formation pathways of thermal and fuel-based nitrogen oxides, hindering efficient emission prediction. Summary of the Invention

[0004] To address the technical problems mentioned in the background art, this application proposes a method for predicting and optimizing nitrogen oxide emissions from the ammonia-hydrogen combustion chamber of a gas turbine. The method for predicting and optimizing nitrogen oxide emissions from the ammonia-hydrogen combustion chamber of a gas turbine in this application includes the following steps: Based on the experimental and simulation data of nitrogen oxide emissions from the ammonia-hydrogen combustion chamber, data fusion and preprocessing are performed to obtain a training dataset; Construct a physical information neural network model, wherein the input data of the physical information neural network model includes combustion condition parameters and flow field characteristic parameters; The physical information neural network model is trained based on the training dataset to adjust the model parameters of the physical information neural network model and to verify the prediction error and physical consistency of the physical information neural network model. The chemical reaction pathways of nitrogen oxides are analyzed based on the physical information neural network model to quantify the pathway contributions of thermal nitrogen oxides and fuel nitrogen oxides. Based on the physical information neural network model, the characteristic parameters of the ammonia-hydrogen combustion chamber are optimized according to the path contribution to determine the parameter combination that minimizes nitrogen oxide emissions.

[0005] In some embodiments, the step of performing data fusion and preprocessing based on experimental and simulation data of nitrogen oxide emissions from the ammonia-hydrogen combustion chamber to obtain a training dataset includes: Based on preset chemical analysis software, the simulation data of the ammonia-hydrogen combustion chamber is generated according to the chemical reaction mechanism, wherein the simulation data is enhanced by adding fluctuations of operating parameters within a preset range. Based on the experimental data and the simulation data, spatiotemporal alignment, normalization, noise filtering, and outlier processing are performed, and the data are fused into the training dataset according to a preset weight ratio.

[0006] In some embodiments, the input data of the physical information neural network model may also include the pressure, premixing temperature, hydrogen doping ratio, equivalence ratio, and turbulence intensity of the ammonia-hydrogen combustion chamber; The output data of the physical information neural network model includes predicted values ​​for total nitrogen oxide emissions, predicted values ​​for the mass fraction of thermal nitrogen oxide emissions, and predicted values ​​for the mass fraction of fuel nitrogen oxide emissions.

[0007] In some embodiments, the hidden layer of the physical information neural network model includes a residual network structure and embeds an adaptive activation function, and the loss function of the physical information neural network model embeds residuals of mass conservation, momentum conservation, energy conservation, and chemical reaction path rate residuals of thermal nitrogen oxide concentration and fuel nitrogen oxide concentration.

[0008] In some implementations, the loss function of the physical information neural network model is:

[0009] in, L total Let the loss function be... λ data For the weighting coefficients of the data loss term, L data For data loss items, λ physics These are the weighting coefficients for the physical conservation constraint terms. L physics The physical conservation constraints include the mass conservation residual, the momentum conservation residual, and the energy conservation residual. λ chem These are the weighting coefficients for the chemical reaction pathway constraint term. L chem This is a chemical reaction pathway constraint term.

[0010] In some implementations, training and validating the physical information neural network model based on the training dataset to adjust the model parameters of the physical information neural network model and verify the prediction error and physical consistency of the physical information neural network model includes: The physical information neural network model is trained based on the Adam optimizer and the L-BFGS optimizer, using the training dataset, wherein the initial learning rate is set to 0.001. The physical information neural network model is validated based on the training dataset. The validation process incorporates an early stopping mechanism based on the loss from the training dataset. Validation metrics include a relative error of less than or equal to 5% for predicted thermal nitrogen oxide emissions, an absolute error of less than or equal to 5 ppm for predicted fuel nitrogen oxide emissions, and residuals of mass conservation, momentum conservation, and energy conservation all less than 10%. -3 .

[0011] In some embodiments, the chemical reaction pathways of nitrogen oxides are analyzed based on the physical information neural network model to quantify the pathway contributions of thermal nitrogen oxides and fuel nitrogen oxides, including: Based on the physical information neural network model, the sensitivity of the thermal nitrogen oxides to temperature and the sensitivity of the fuel nitrogen oxides to NH2 and HCN atomic groups are calculated by automatic differentiation. Based on the temperature sensitivity of the thermal nitrogen oxides and the sensitivity of the fuel nitrogen oxides to the NH2 and HCN groups, normalization and spatiotemporal integration calculations are performed to determine the path contribution of the Zeldovich mechanism pathway, the NH2 pathway, and the HCN pathway to nitrogen oxide generation.

[0012] In some embodiments, optimizing the characteristic parameters of the ammonia-hydrogen combustor based on the path contribution degree according to the physical information neural network model to determine the parameter combination that minimizes nitrogen oxide emissions includes: Multiple sets of uniformly distributed parameter samples are generated based on the Latin hypercube design, and the nitrogen oxide emission values ​​at each parameter sample are predicted based on the physical information neural network model. The parameter samples with the expected improvement level within a preset range are selected based on the Bayesian algorithm, and crossover, mutation, and selection operations are performed in combination with the genetic algorithm until the parameter samples meet the preset convergence conditions, so as to determine the parameter combination that minimizes the nitrogen oxide emission value.

[0013] In some implementations, the genetic algorithm includes: The parameter samples with the lowest nitrogen oxide emission values ​​were selected as the initial population. Uniform crossover and Gaussian mutation are performed on the initial population, wherein the standard deviation of the Gaussian mutation is set to 5% of the range of values ​​for the parameter samples; If the expected improvement of the parameter samples is less than 1% during 10 consecutive iterations, or if the number of iterations reaches a preset maximum value, the genetic algorithm is terminated.

[0014] In some embodiments, the method further includes: When the deviation between the experimentally measured total nitrogen oxide emissions and the predicted total nitrogen oxide emissions exceeds a preset threshold, local operating condition data is supplemented and the physical information neural network model is retrained. Adjust the range of hydrogen doping ratio and premixing temperature in the input data of the physical information neural network model to improve the accuracy of the physical information neural network model.

[0015] The beneficial effects of this application are: by using experimental and simulation data from the ammonia-hydrogen combustion chamber to establish a physical information neural network model, and further setting conservation constraints and chemical reaction path constraints, the accuracy of the model's prediction of nitrogen oxide emissions is improved, as well as the overall efficiency and physical interpretability of the model. Moreover, the established physical information neural network model can be used to quantify the path contribution of the nitrogen oxide generation process to guide the optimization of variables and parameter sampling selection strategies, thereby enabling rapid search and optimization of combustion characteristic parameters in the combustion chamber, reducing computation time, and thus improving the optimization efficiency of nitrogen oxide emission prediction in the ammonia-hydrogen combustion chamber. Attached Figure Description

[0016] Figure 1 This is one of the flowcharts illustrating the method for predicting and optimizing nitrogen oxide emissions from the ammonia-hydrogen combustion chamber of a gas turbine in the embodiments of this application; Figure 2 This is the second flowchart illustrating the method for predicting and optimizing nitrogen oxide emissions from the ammonia-hydrogen combustion chamber of a gas turbine in the embodiments of this application. Figure 3 This is the third flowchart illustrating the method for predicting and optimizing nitrogen oxide emissions from the ammonia-hydrogen combustion chamber of a gas turbine in the embodiments of this application. Figure 4 This is the fourth flowchart illustrating the method for predicting and optimizing nitrogen oxide emissions from the ammonia-hydrogen combustion chamber of a gas turbine in the embodiments of this application. Figure 5 This is the fifth flowchart illustrating the method for predicting and optimizing nitrogen oxide emissions from the ammonia-hydrogen combustion chamber of a gas turbine in the embodiments of this application. Figure 6 This is the sixth flowchart illustrating the method for predicting and optimizing nitrogen oxide emissions from the ammonia-hydrogen combustion chamber of a gas turbine in the embodiments of this application. Figure 7 This is the seventh flowchart illustrating the method for predicting and optimizing nitrogen oxide emissions from the ammonia-hydrogen combustion chamber of a gas turbine in the embodiments of this application. Detailed Implementation

[0017] Please see Figure 1 The method for predicting and optimizing nitrogen oxide emissions from the ammonia-hydrogen combustion chamber of a gas turbine according to the embodiments of this application specifically includes the following steps: Step 01: Based on the experimental data and simulation data of nitrogen oxide emissions from the ammonia-hydrogen combustion chamber, perform data fusion and preprocessing to obtain the training dataset.

[0018] Specifically, constructing a Physics-Informed Neural Network (PINN) model requires NOx emission data from an ammonia-hydrogen combustion chamber. Considering the existence of extreme operating conditions during experiments, such as high pressure or high hydrogen doping ratios, which are difficult to meet or reproduce, it is necessary to obtain NOx emission experimental data through gas combustion experiments in the ammonia-hydrogen combustion chamber. Furthermore, it is necessary to utilize existing chemical reaction process simulation software platforms to simulate chemical reaction processes under extreme conditions, thus supplementing the experimental data. After obtaining both the experimental and simulation data of NOx emissions, the two sets of data are fused and preprocessed according to predetermined rules to form a training dataset that can be used to train PINN.

[0019] For gas combustion experiments in ammonia-hydrogen combustion chambers, experimental techniques such as planar laser-induced fluorescence (PLIF), particle image velocimeter (PIV), chemiluminescence analyzer, Fourier transform infrared spectroscopy (FTIR), high-frequency pressure sensor, and data acquisition system are generally used to acquire data on multiple parameters in the chemical reaction process involved in the gas combustion experiment. The parameters that can be acquired generally include: the concentration of key free radicals for the generation of thermal NOx such as O, H, and OH; the concentration of intermediate products generated in the reaction pathway of fuel NOx such as NH2, HCN, and NCO; velocity field; turbulence intensity; total NOx concentration; thermal NOx concentration; fuel NOx concentration; pressure pulsation parameters; and thermoacoustic oscillation frequency.

[0020] Please continue reading. Figure 2 Step 01 further includes the following steps: Step 011: Based on the preset chemical analysis software, generate simulation data of the ammonia-hydrogen combustion chamber according to the chemical reaction mechanism. The simulation data is enhanced by adding fluctuations of operating parameters within a preset range.

[0021] Specifically, for the simulation of chemical reaction processes under extreme conditions, traditional computational fluid dynamics (CFD) simulation software can generally be used. For example, to ensure the chemical reaction simulation data is as accurate as possible, ANSYS Chemkin software is used to perform the above simulation process. The specific chemical reaction mechanism is the Stagni mechanism of ammonia-hydrogen combustion. The solver is coupled with turbulence models such as the Ke model. The output simulation data includes temperature field, pressure field, velocity field, concentration field of all chemical components, and decomposition of NOx along the chemical reaction path. The above decomposition generally includes the specific gravity contribution parameters of thermal NOx and fuel NOx.

[0022] In addition, for the equivalence ratio and hydrogen doping ratio input parameters in the simulation process, in order to fill the gaps in the experimental data, Gaussian noise within a preset range will be added to the input parameters to make the operating condition difference. For example, the preset range can generally be ±5%. In addition, some extreme equivalence ratios and / or extreme hydrogen doping ratios that are difficult to achieve in gas combustion experiments will be added to fill the gaps in the operating conditions and ensure the integrity of the data.

[0023] Step 012: Based on the experimental data and simulation data, perform spatiotemporal alignment, normalization, noise filtering, and outlier processing, and fuse them into a training dataset according to a preset weight ratio.

[0024] Specifically, to ensure consistency between experimental and simulation data, and thus provide high-quality training and validation datasets for the PINN training process, the experimental and simulation data are first spatiotemporally aligned. For example, this spatiotemporal alignment process also includes performing Min-Max normalization on the input parameters of the simulation process. The specific execution method can refer to the Min-Max normalization execution process in current related technologies, and this application does not impose specific limitations. Secondly, the experimental and simulation data are normalized. Normalization mainly includes performing a logarithmic transformation on the NOx concentration in the output parameters to address dimensional differences. For example, let the NOx concentration in the output parameters be... Y The corresponding concentration value after transformation is Y log Therefore, the above logarithmic transformation can be expressed as the following formula:

[0025] Next, noise filtering is performed on the experimental and simulation data that have undergone spatiotemporal alignment and normalization. For example, Gaussian filtering with a standard deviation of 2 pixels or median filtering with a 3×3 window is used for the data obtained by planar laser-induced fluorescence (PLIF) or particle image velocimeter (PIV), while wavelet filtering is used to remove noise for pressure pulsation parameters.

[0026] In addition to spatiotemporal alignment, normalization, and noise filtering, the above methods also include outlier handling. Specifically, based on the statistical 3σ criterion, data points deviating from the mean by ±3 standard deviations can be deleted. Simultaneously, experimental and simulation data are compared, and data with obvious conflicts are manually removed. Furthermore, any data that may exhibit deviations is also addressed. X The following formula can be used to process it into a normal value:

[0027] in, X norm These are the normal values ​​after processing. X max This corresponds to the maximum value of the experimental data or the maximum value of the simulation data. X min This represents the minimum value of the corresponding experimental data or the minimum value of the corresponding simulation data.

[0028] Finally, after completing data preprocessing steps such as spatiotemporal alignment, normalization, noise filtering, and outlier handling, the preprocessed experimental and simulation data are fused. Different fusion weights are used for the experimental and simulation data during fusion; for example, the fusion weight for the experimental data is set to 0.7, and the fusion weight for the simulation data is set to 0.3, ultimately forming the training dataset for training PINN. Furthermore, considering that both experimental and simulation data contain certain errors, measurement errors can be further labeled on the experimental data, and grid convergence errors can be labeled on the simulation data. This quantifies uncertainties such as errors, facilitating quantitative adjustments during PINN training to ensure the accuracy of PINN.

[0029] Please continue reading. Figure 1 The method for predicting and optimizing nitrogen oxide emissions from the ammonia-hydrogen combustion chamber of a gas turbine in this application also includes: Step 02: Construct a physical information neural network model. The input data for the physical information neural network model includes combustion condition parameters and flow field characteristic parameters.

[0030] Specifically, assuming the training and validation datasets are ready, the next step is to construct a PINN, which can simultaneously predict both thermal and fuel NOx emissions. The input data in its input layer includes combustion parameters and flow field characteristic parameters. The combustion parameters specifically include the pressure in the ammonia-hydrogen combustor. P Premixing temperature T pre Hydrogen doping ratio X H and equivalent ratio Φ The flow field characteristic parameters include turbulence intensity. The data output by its output layer includes the predicted total NOx emission concentration, as well as the predicted emission concentrations of thermal NOx and combustible NOx.

[0031] Furthermore, the PINN model in the above implementation has a backbone network with 8 fully connected layers, each with 256 neurons. It primarily mitigates the vanishing gradient phenomenon by employing residual connections (ResNet), and also embeds an adaptive activation function, the Swish function, into the hidden layers. The specific expression is as follows:

[0032] Where e is the natural constant. β As the core adjustment parameter of the Swish function, in the PINN model of the above implementation, β It is a learnable adjustable parameter, whose main function is to automatically adjust the nonlinear expression capability of the model based on the prediction process of NOx emission concentration.

[0033] Furthermore, the PINN model also embeds mass conservation residuals, momentum conservation residuals, and energy conservation residuals into its loss function, so that the prediction results output by the PINN model satisfy physical laws.

[0034] The mass conservation residual is used to constrain the physical consistency of the flow field prediction, preventing the creation or disappearance of mass during the model prediction process. The equation for the mass conservation residual is expressed as follows:

[0035] in R mass For the residual of quality conservation, ρ Let ρ be the fluid density and u be the velocity field.

[0036] The momentum conservation residual is used to constrain the velocity and pressure fields to conform to the laws of fluid dynamics. The equation for the momentum conservation residual (i.e., the Navier-Stokes equation) is as follows:

[0037] in R momentum For momentum conservation residuals, ρ Let U be the fluid density, u be the velocity field, and P be the pressure field. μ Let g be the dynamic viscosity of the fluid, and g be the acceleration due to gravity.

[0038] The energy conservation residual is used to constrain the distribution of the temperature field. The equation for the energy conservation residual is expressed as follows:

[0039] in R energy For the energy conservation residual, ρ Let U be the fluid density, u be the velocity field, and T be the temperature field. k The thermal conductivity of the fluid, c p The specific heat at constant pressure of a fluid. The term "chemical reaction source term" is used to represent the rate of consumption of reactants or the rate of formation of products in a chemical reaction. h i Enthalpy is the enthalpy of each component in the fluid.

[0040] It should be noted that the fluid density mentioned above... ρ Dynamic viscosity of fluids μ thermal conductivity of fluids k Specific heat at constant pressure of fluids c p Enthalpy of each component in the fluid h i All quantities are known, and the fluids involved refer to the mixture of ammonia-hydrogen gas and air in actual gas combustion experiments or simulated gas combustion experiments.

[0041] The three sets of residual constraints mentioned above together form the physical conservation constraint term in the PINN model, as shown in the following formula:

[0042] in N coll To calculate the number of configuration points within the domain.

[0043] Next, the PINN model further embeds the chemical reaction path rate residuals of thermal NOx concentration and fuel NOx concentration into its loss function, so that the prediction results output by the PINN model meet the chemical reaction path in the ammonia-hydrogen combustion chamber.

[0044] Specifically, the chemical reaction path rate constraint for thermal NOx concentration is generally based on the Zeldovich mechanism. The main function of this constraint is to ensure that the formation rate of NO atom groups in the PINN prediction process is consistent with the Zeldovich mechanism.

[0045] The Zeldovich mechanism includes the following three key chemical reactions, where the element symbols in the formula represent atoms or groups of atoms: Reaction 1: , Reaction 2: , Reaction 3: , Where k1, k2, and k3 are the Arrhenius rate constants for the corresponding reactions.

[0046] Based on the above mechanism, the residual constraint equation corresponding to the thermal NOx concentration can be obtained as follows:

[0047] in R Zeldovich This represents the constraint residual corresponding to the thermal NOx concentration. t k is the reaction time. -1 is the reverse Arrhenius rate constant for reaction 1.

[0048] The chemical reaction pathway rate constraint for fuel NOx concentration is generally based on the decomposition pathway of NH3 and the conversion pathway of HCN. The main function of this constraint is to ensure that the concentration changes of NH2 and HCN groups in the PINN prediction process conform to the decomposition and consumption rates in the reaction process.

[0049] Fuel nitrogen is generally converted to NOx through intermediates such as NH2 and HCN, and the specific reactions are as follows: Reaction 4 (decomposition of NH3): , Reaction 5 (NH2 oxidation): , Reaction 6 (HNO conversion): , Reaction 7 (HCN oxidation): , Reaction 8 (NCO oxidation): , Where k4~k8 are the Arrhenius rate constants for the corresponding reactions.

[0050] Based on the above reaction mechanism, the chemical reaction path rate constraint for fuel NOx concentration can be divided into two parts: NH2 path constraint and HCN path constraint. The specific residual constraint equations are as follows:

[0051]

[0052] The formula for calculating the Arrhenius rate constant mentioned above is as follows:

[0053] in A i For the reaction i Pre-index factor b i For the reaction i Temperature index E i For the reaction i activation energy, R The gas constant is... T The temperature is the gas temperature.

[0054] The three sets of residual constraints mentioned above together form the chemical reaction path constraint term in the PINN model, as shown in the following formula:

[0055] Furthermore, based on the experimental data, the data loss residual constraints in the PINN model can be obtained, and the specific formula for the corresponding data loss term is as follows:

[0056] in, N data The number of samples for experimental or simulation data. For the first j The measured values ​​of thermal NOx emission concentration for each sample. For the first j The measured values ​​of NOx emission concentration from fuel sources for each sample. For the first j Predicted values ​​of thermal NOx emission concentration for each sample. For the first j Predicted NOx emission concentrations for each sample of fuel.

[0057] Finally, by combining the preset weighting coefficients with the data loss term, the physical conservation constraint term, and the chemical reaction path constraint term, the loss function of the entire PINN model is obtained, as follows:

[0058] in, L total Let be the loss function of the PINN model. λ data For the weighting coefficients of the data loss term, L data For data loss items, λ physics These are the weighting coefficients for the physical conservation constraint terms. L physics The physical conservation constraints include the mass conservation residual, the momentum conservation residual, and the energy conservation residual. λ chem These are the weighting coefficients for the chemical reaction pathway constraint term. L chem This is a chemical reaction pathway constraint term.

[0059] Please continue reading. Figure 1 The method for predicting and optimizing nitrogen oxide emissions from the ammonia-hydrogen combustion chamber of a gas turbine in this application also includes: Step 03: Train and validate the physical information neural network model based on the training dataset and validation dataset to adjust the model parameters of the physical information neural network model and verify the prediction error and physical consistency of the physical information neural network model.

[0060] Specifically, once the PINN model is established, it can be trained using the training dataset. The PINN model training process mainly includes three parts: training, validation, and testing. Correspondingly, the training dataset in the above implementation needs to be divided into three parts: a training part (approximately 70%), a validation part (approximately 15%), and a testing part (approximately 15%). Random sampling is used during the division according to combustion conditions to ensure that each part can fully cover different combustion conditions. The training step uses the training part of the training dataset, primarily to adjust the PINN model parameters to form its predictive function. The validation step uses the validation part of the training dataset, mainly for performing hyperparameter adjustments and selecting specific computational models to control the current prediction error and maintain the physical consistency of the PINN model. The testing step uses the testing part of the training dataset, primarily to perform a final performance evaluation of the PINN model and ensure its generalization ability.

[0061] Please refer to the following for details. Figure 3 Step 03 further includes: Step 031: Based on the Adam optimizer and L-BFGS optimizer, train a physical information neural network model using the training dataset. The initial learning rate is set to 0.001.

[0062] Specifically, based on the above implementation method, in the initial stage of the training step, the Adam optimizer is generally selected to train the PINN model, and the initial learning rate can be set to 0.001, the purpose of which is to enable the PINN model to converge quickly. In the middle and later stages of the training step, the L-BFGS optimizer is selected to improve training accuracy. Furthermore, in each stage of the training step, the loss function of the PINN model... L total The weight coefficients of each residual constraint term will be gradually adjusted according to the order of each segment of the training steps. The specific adjustment method can be implemented according to the actual situation, and this application does not make specific limitations.

[0063] In the initial stage of the training process, the training dataset uses stable combustion conditions with a hydrogen blending ratio of <15% and an equivalence ratio of 0.8-1.2. In the middle stage, the training dataset uses more complex conditions with a hydrogen blending ratio of 15%-25% and an equivalence ratio of 0.6-1.4. In the later stage, the training dataset uses extreme lean combustion conditions with a hydrogen blending ratio >25% and an equivalence ratio less than or equal to 0.5, or extreme rich combustion conditions with a hydrogen blending ratio >25% and an equivalence ratio greater than or equal to 1.6.

[0064] Step 032: Validate the physical information neural network model based on the training dataset. The validation process incorporates an early stopping mechanism based on the loss of the training dataset. Validation metrics include a relative error of less than or equal to 5% for predicted thermal nitrogen oxide emissions, an absolute error of less than or equal to 5 ppm for predicted fuel nitrogen oxide emissions, and residuals for mass conservation, momentum conservation, and energy conservation all less than 10%. -3 .

[0065] Specifically, based on the above implementation method, the verification step generally involves quantitatively calculating multiple verification indicators for the PINN model. These verification indicators generally include the relative error of the predicted thermal NOx emissions, the absolute error of the predicted fuel NOx emissions, the mass conservation residual, the momentum conservation residual, and the energy conservation residual.

[0066] The verification conditions for the relative error of the predicted thermal NOx emissions are as follows:

[0067] in N The number of predicted values ​​for thermal NOx emissions. Y pred, thermalThese are predicted values ​​for thermal NOx emissions. Y exp, thermal This refers to the thermal NOx emissions included in the experimental data.

[0068] The verification conditions for the absolute error of the predicted NOx emissions from fuel sources are as follows:

[0069] Where MAE is the mean absolute error function. Y pred, fuel These are predicted values ​​for NOx emissions from fuel sources. Y exp, fuel This refers to the NOx emissions from fuel sources included in the experimental data.

[0070] The verification conditions for the residuals of mass conservation, momentum conservation, and energy conservation are as follows:

[0071]

[0072]

[0073] Furthermore, during the iterative training of the PINN model, the training error gradually decreases in the training portion of the training dataset, but gradually increases in the validation portion. Therefore, an early stopping mechanism is introduced during the training of the PINN model to monitor the loss level in the validation portion of the training dataset. For example, the early stopping criterion can be set as follows: if the PINN model has been trained continuously and completely 20 times on the training portion of the training dataset, and the loss level in the validation portion has not decreased during these 20 training iterations, then the training process of the PINN model can be stopped, thereby preventing overfitting and ensuring the model's generalization ability.

[0074] Please continue reading. Figure 1 The method for predicting and optimizing nitrogen oxide emissions from the ammonia-hydrogen combustion chamber of a gas turbine in this application also includes: Step 04: Analyze the chemical reaction pathways of nitrogen oxides based on the physical information neural network model to quantify the pathway contributions of thermal nitrogen oxides and fuel nitrogen oxides.

[0075] Specifically, based on the above implementation method, once the PINN model has been trained, the chemical reaction pathways for NOx generation in the ammonia-hydrogen combustion chamber can be analyzed using the trained PINN model. This allows for the quantification of the pathway contributions of thermal NOx and fuel NOx, facilitating the prediction of subsequent NOx emissions.

[0076] Please see Figure 4 Step 04 specifically includes: Step 041: Based on the physical information neural network model, automatically differentiate to calculate the sensitivity of thermal nitrogen oxides to temperature, and the sensitivity of fuel nitrogen oxides to NH2 and HCN atomic groups.

[0077] Specifically, the following example illustrates a method for quantifying the path contributions of thermal NOx and fuel NOx. Based on the PINN model, its automatic differentiation function can be used to quantify the contributions of thermal and fuel NOx to their respective chemical reaction pathways. These reaction pathways include the Zeldovich mechanism, NH3 decomposition pathway, and HCN pathway mentioned in the above embodiments. The key intermediate products of thermal NOx are free radicals such as O, N, and OH groups, while the key intermediate products of fuel NOx are NH2, HCN, and NCO. By calculating the gradient of each intermediate product using the automatic differentiation function, the path contribution can be quantified.

[0078] For example, the sensitivity of each chemical reaction pathway is first calculated. The sensitivity described above describes the reactivity of the corresponding chemical reaction pathway, as follows: Temperature sensitivity of thermal NOx S T→thermal It represents the change in thermal NOx formation caused by a unit temperature change, reflecting the reactivity of the Zeldovich mechanism, as shown in the following formula:

[0079] in T The temperature is the gas temperature.

[0080] Sensitivity of fuel-type NOx to NH2 S NH2→fuel It represents the change in the amount of fuel-type NOx generated due to a unit change in NH2 concentration, reflecting the reactivity of the NH3 decomposition pathway, as shown in the following formula:

[0081] Sensitivity of fuel-type NOx to HCN S HCN→fuel It represents the change in fuel-type NOx formation caused by a unit change in HCN concentration, reflecting the reactivity of the HCN pathway, as shown in the following formula:

[0082] Step 042: Based on the temperature sensitivity of thermal nitrogen oxides and the sensitivity of fuel nitrogen oxides to NH2 and HCN groups, perform normalization and spatiotemporal integration calculations to determine the path contribution of the Zeldovich mechanism pathway, NH3 decomposition pathway, and HCN pathway to nitrogen oxide generation.

[0083] Specifically, based on the above implementation method, normalizing the three sets of sensitivities yields three sets of relative contribution values. For example, using... C Representing the relative path contribution, we can obtain the following three sets of relative path contribution values: For the Zeldovich mechanism, the relative path contribution is: ; For the NH3 decomposition pathway, its relative pathway contribution is: ; For the HCN path, its relative path contribution is: ; in .

[0084] Next, by performing spatiotemporal integration on the net formation rate of key chemical reactions, we can directly quantify the absolute pathway contribution of each chemical reaction pathway to NOx formation.

[0085] Specifically, regarding the Zeldovich mechanism, the NO generation rate... The formula is as follows:

[0086] Similarly, for the NH3 decomposition pathway, the rate of NO formation... The formula is as follows:

[0087] For the HCN pathway, the NO generation rate The formula is as follows:

[0088] By performing spatiotemporal integration on each of the aforementioned NO generation rates, we can obtain the NO generation amount for the corresponding path. The general formula for calculating the NO generation amount is:

[0089] in This represents the NO generation rate corresponding to one of the three reaction pathways mentioned above. Q i This represents the amount of NO generated along the same reaction pathway. i It can be NH2 or HCN t 0 represents the initial time. t f Let Ω be the target time for the spacetime integration, and let Ω be the target integration space.

[0090] Based on the above calculations, the total NO generation... Q total The calculation formula is as follows:

[0091] Finally, based on the NO production in each reaction pathway and the total NO production, the absolute path contribution of each reaction pathway can be calculated. C If ' represents the absolute path contribution, then the following conclusions can be drawn:

[0092]

[0093]

[0094] This achieves the quantification of the relative and absolute path contributions of each chemical reaction pathway to NOx generation.

[0095] Please continue reading. Figure 1 The method for predicting and optimizing nitrogen oxide emissions from the ammonia-hydrogen combustion chamber of a gas turbine in this application also includes: Step 05: Based on the physical information neural network model, optimize the characteristic parameters of the ammonia-hydrogen combustion chamber based on path contribution to determine the parameter combination that minimizes nitrogen oxide emissions.

[0096] Specifically, based on the above implementation method, after obtaining the path contribution of each chemical reaction pathway, the PINN model can be guided to perform sampling based on the above path contribution to avoid blind experimentation. Parameters such as hydrogen doping ratio, equivalence ratio, premixing temperature and combustion chamber pressure are used as optimization variables, and the optimization objective is to minimize the corresponding total NOx emission value.

[0097] Please refer to further information. Figure 5 Step 05 further includes: Step 051: Generate multiple sets of uniformly distributed parameter samples based on the Latin hypercube design, and predict the nitrogen oxide emission values ​​at each parameter sample based on the physical information neural network model.

[0098] Specifically, based on the above implementation method, before performing optimization, multiple sets of parameter samples that can uniformly cover the parameter space are first generated according to the Latin Hypercube (LHS) sampling design to ensure that each parameter in the sample is independently distributed. The parameters included in the parameter samples are the same as the input data types included in the PINN model input layer. That is, each set of parameter samples includes pressure, premixing temperature, hydrogen doping ratio, equivalence ratio and / or turbulence intensity. The number of parameter samples can be adjusted according to the actual situation. For example, 50 sets can be taken.

[0099] Next, using the pre-trained PINN model, NOx emissions are predicted using the above-mentioned multiple sets of parameter samples as input data. After the prediction is completed, 50 sets of NOx emission prediction results will be output.

[0100] Step 052: Select parameter samples with the expected improvement level within the preset range according to the Bayesian algorithm, and perform crossover, mutation and selection operations in combination with the genetic algorithm until the parameter samples meet the preset convergence conditions, so as to determine the parameter combination that minimizes the nitrogen oxide emission value.

[0101] Specifically, based on the above implementation method, the next step is to use the generated 50 sets of NOx emission prediction results as a basis to optimize the corresponding parameter samples using Bayesian algorithm and genetic algorithm. The ultimate goal is to determine the parameter combination that minimizes the NOx emission value from the above parameter samples.

[0102] For example, before adopting the Bayesian algorithm, considering the accuracy of the PINN model, 5 sets are randomly sampled from the above 50 sets of parameter samples and simulated using a traditional CFD software platform to ensure that the prediction accuracy error of the PINN model is lower than a preset value. The above preset value can generally be set to 5%, or it can be adjusted according to the actual situation. This application does not make any specific limitations.

[0103] Next, using the PINN model as a surrogate model, a Bayesian algorithm is employed to select one or more sets of parameter samples with the greatest expected improvement for the optimization objective of minimizing NOx emissions. The PINN model is then used for additional prediction to obtain the corresponding NOx emission target. The range of selected parameter samples can be pre-defined. For a specific optimization algorithm, a genetic algorithm can be used, for example. When the genetic algorithm iterates over the parameter samples and the samples meet the convergence condition, the optimization process is complete. The parameter samples that meet the convergence condition are then determined as the parameter combination that minimizes nitrogen oxide emissions. Please refer to further information. Figure 6 The execution steps of the above genetic algorithm are as follows: Step 001: Select the parameter samples with the smallest nitrogen oxide emission values ​​as the initial population.

[0104] Specifically, the following example illustrates how the genetic algorithm is executed: First, among the 50 parameter samples generated by Latin hypercube sampling, the 20 parameter samples with the lowest NOx emission prediction values ​​are selected based on the predicted values ​​output by the PINN model, and used as the initial data population for the algorithm. Each parameter sample includes at least pressure, premixing temperature, hydrogen doping ratio, and equivalence ratio. In addition, turbulence intensity can be added as needed.

[0105] Step 002: Perform uniform crossover and Gaussian mutation on the initial population, where the standard deviation of the Gaussian mutation is set to 5% of the range of parameter sample values.

[0106] Specifically, based on the above implementation method, for the parental data population, two sets of parameter samples are randomly selected and uniform crossover is performed to generate corresponding parameter samples in the offspring data population, thereby realizing the iteration of parameter samples. Specifically, the two sets of parameter samples in the parental data population serve as the parent parameter samples and the mother parameter samples, respectively. These two sets of parameter samples correspond to a set of child parameter samples in the offspring data population. Each parameter included in the child parameter sample originates from either the aforementioned parent or mother parameter sample. The specific source is determined by equal probability randomness, meaning that the probability of originating from the parent parameter sample is equal to the probability of originating from the mother parameter sample. The aforementioned parental data population is the initial population for each new iteration process.

[0107] Next, Gaussian mutation is performed on any parameter sample included in the offspring data population. Specifically, for a certain parameter included in any parameter sample in the offspring data population, a random perturbation following a Gaussian distribution is applied. After applying the perturbation, the mean of the parameter should be kept at its current value, and the standard deviation of the parameter should be kept at ±5% of the parameter's range. This allows the parameter sample values ​​to be diversified, thereby improving the environmental adaptability of the optimization results.

[0108] In addition, for all parameter samples that undergo uniform crossover and all sub-parameter samples generated after uniform crossover, NOx emissions need to be predicted using the PINN model in the above implementation. Then, the half of the parameters with the larger NOx emissions are discarded, and two sets of parameter samples are randomly selected again to undergo uniform crossover to achieve parameter sample iteration.

[0109] Step 003: If the expected improvement of the parameter samples is less than 1% during 10 consecutive iterations, or if the number of iterations reaches the preset maximum value, the genetic algorithm is terminated.

[0110] Specifically, based on the above implementation method, there are generally two termination conditions for genetic algorithms: one is that the parameter samples meet the preset convergence conditions, and the other is that the number of iterations reaches the preset maximum value.

[0111] For example, the above convergence condition is: if the expected improvement of the predicted NOx emission value corresponding to the parameter samples that have not been discarded is less than 1% during 10 consecutive iterations, it means that the parameter samples that meet the requirements have been optimized. At this time, it is considered that the parameter samples meet the preset convergence condition, the genetic algorithm is terminated, and the parameter samples that have not been discarded at this time are determined as the parameter combination that minimizes the NOx emission value.

[0112] Regarding the maximum number of iterations, the upper limit of the number of iterations can generally be set to 50. When the number of iterations reaches this value, the iteration will stop and the genetic algorithm will terminate and exit. The purpose of this termination is to avoid infinite loops. At this time, the genetic algorithm can determine the parameter combination that minimizes NOx emissions from the parameter samples that have not been discarded, or it can not output optimization results and prompt the user that the number of iterations has exceeded so that the user can adjust the algorithm in time.

[0113] Please see Figure 7 The above-mentioned genetic algorithm execution process also includes the following steps: Step 004: When the deviation between the experimentally measured total nitrogen oxide emissions and the predicted total nitrogen oxide emissions exceeds a preset threshold, supplement local operating condition data and retrain the physical information neural network model.

[0114] Specifically, based on the above implementation method, during the actual execution of the genetic algorithm, for each set of parameter samples, if corresponding experimental data exists, the total NOx emission value in the corresponding experimental data can be compared with the predicted total NOx emission value output by the PINN model based on that set of parameter samples. If the deviation between the two exceeds a preset threshold, it indicates that the PINN model has an accuracy problem and needs to be retrained by supplementing local operating condition data to ensure its prediction accuracy. The preset threshold can be set to 8% for example, and can be adjusted according to the actual situation in practical applications; this application does not impose specific limitations.

[0115] Step 005: Adjust the hydrogen doping ratio range and premixing temperature range in the input data of the physical information neural network model to improve the accuracy of the physical information neural network model.

[0116] Specifically, the process of retraining the PINN model by supplementing local operating condition data generally involves increasing the range of hydrogen doping ratio and premixing temperature in the input data of the PINN model. For example, based on the current hydrogen doping ratio, the increased hydrogen doping ratio is within ±2% of the current hydrogen doping ratio. Or, based on the current premixing temperature, the increased premixing temperature is within ±10K of the current premixing temperature. The specific adjustment range and the amount of data added can be adjusted according to the actual situation. This application does not impose specific limitations. Within a certain range, the larger the adjustment range and the more data added, the higher the accuracy of the PINN model after retraining.

[0117] It should be added that, Figure 7 The execution flow shown is for illustrative purposes only. Steps 004 and 005 can be executed simultaneously with steps 001-003, or steps 004-005 can be executed separately after steps 001-003 have been paused. Figure 7 This should not be interpreted as a restriction on the order of execution.

[0118] Thus, the method for predicting and optimizing NOx emissions in a gas turbine ammonia-hydrogen combustors provided in this application can establish a PINN model using experimental and simulation data from the ammonia-hydrogen combustors, and further set conservation constraints and chemical reaction path constraints, thereby improving the accuracy of the model's NOx emission predictions and enhancing the overall efficiency and physical interpretability of the model. Moreover, the established PINN model can be used to quantify the path contribution of the NOx generation process to guide variable optimization and parameter sampling selection strategies, thereby enabling rapid search and optimization of combustion characteristic parameters in the combustor, reducing computation time, and thus improving the optimization efficiency of NOx emission prediction in ammonia-hydrogen combustors.

[0119] The above description is merely a preferred embodiment of this application and is not intended to limit this application in any way. Although this application has disclosed the preferred embodiment as above, it is not intended to limit this application. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the technical solution of this application. Any simple modifications, equivalent substitutions, and improvements made to the above embodiments without departing from the technical solution of this application, based on the technical essence of this application and within the spirit and principles of this application, shall still fall within the protection scope of the technical solution of this application.

Claims

1. A method of NOx emission prediction and optimization for ammonia-hydrogen combustion in a gas turbine combustor, characterized by, The method comprises: According to the experimental data and simulation data of the ammonia-hydrogen combustion chamber, data fusion and preprocessing are performed to obtain a training data set; A physical information neural network model is constructed, wherein the input data of the physical information neural network model comprises combustion operating condition parameters and flow field characteristic parameters; The physical information neural network model is trained according to the training data set, so as to adjust the model parameters of the physical information neural network model and verify the prediction error and physical consistency of the physical information neural network model; According to the physical information neural network model, the chemical reaction path of nitrogen oxides is analyzed to quantify the path contribution degree of thermal nitrogen oxides and fuel-type nitrogen oxides; According to the physical information neural network model, the characteristic parameters of the ammonia-hydrogen combustion chamber are optimized based on the path contribution degree to determine the parameter combination that minimizes the nitrogen oxide emission value.

2. The method of claim 1, wherein, According to the experimental data and simulation data of the ammonia-hydrogen combustion chamber, data fusion and preprocessing are performed to obtain a training data set, comprising: Based on a preset chemical analysis software, the simulation data of the ammonia-hydrogen combustion chamber is generated according to the chemical reaction mechanism, wherein the simulation data is data enhanced by adding operating condition parameter fluctuations in a preset range; According to the experimental data and the simulation data, time-space alignment, normalization, noise filtering and outlier processing are performed, and the training data set is fused according to a preset weight ratio.

3. The method of claim 1, wherein, The input data of the physical information neural network model comprises the pressure, premixed temperature, hydrogen blending ratio, equivalence ratio and turbulent intensity of the ammonia-hydrogen combustion chamber; The output data of the physical information neural network model comprises the predicted value of the total emission concentration of nitrogen oxides, the predicted value of the emission concentration of thermal nitrogen oxides, and the predicted value of the emission concentration of fuel-type nitrogen oxides.

4. The method of claim 3, wherein, The hidden layer of the physical information neural network model comprises a residual network structure and embeds an adaptive activation function, and the loss function of the physical information neural network model embeds a mass conservation residual, a momentum conservation residual, an energy conservation residual, a chemical reaction path rate residual of thermal nitrogen oxide concentration and fuel-type nitrogen oxide concentration.

5. The method of any one of claim 4, characterized in that, The loss function of the physical information neural network model is: wherein, L total is the loss function, The training of the physical information neural network model according to the training data set, the adjustment of the model parameters of the physical information neural network model, and the verification of the prediction error and physical consistency of the physical information neural network model, comprise: data is a data loss term weight coefficient, L data is a data loss term, Based on the Adam optimizer and the L-BFGS optimizer, the physical information neural network model is trained according to the training data set, wherein the initial learning rate is set to 0.001; physics is a physical conservation constraint term weight coefficient, L physics is a physical conservation constraint term, the mass conservation residual, the momentum conservation residual, the energy conservation residual are included in the physical conservation constraint term, According to the physical information neural network model, the chemical reaction path of nitrogen oxides is analyzed to quantify the path contribution degree of the thermal nitrogen oxides and the fuel-type nitrogen oxides, comprising: chem is a chemical reaction path constraint term weight coefficient, L chem is a chemical reaction path constraint term.

6. The method of claim 1, wherein, According to the physical information neural network model, the sensitivity of the thermal nitrogen oxides to temperature and the sensitivity of the fuel-type nitrogen oxides to NH2 groups and HCN groups are calculated by automatic differentiation; ​ According to the training data set, the physical information neural network model is verified, wherein a validation process based on the loss of the training data set introduces an early stopping mechanism, and a validation index of the validation process includes that a relative error of a thermal type nitrogen oxide emission prediction value is less than or equal to 5%, an absolute error of a fuel type nitrogen oxide emission prediction value is less than or equal to 5 ppm, mass conservation residual error, momentum conservation residual error and energy conservation residual error are all less than 10 -3 .

7. The method of claim 1, wherein, ​ ​ According to the sensitivity of the thermal-type nitrogen oxides to temperature and the sensitivity of the fuel-type nitrogen oxides to NH2 groups and HCN groups, normalization and space-time integral calculation are performed to determine path contributions of the Zeldovich mechanism path, the NH2 path and the HCN path to the amount of nitrogen oxides generated.

8. The method of claim 1, wherein, The method further comprises: According to a Latin hypercube design, a plurality of groups of uniformly distributed parameter samples are generated, and nitrogen oxide emission values at each of the parameter samples are predicted according to the physical information neural network model; According to a Bayesian algorithm, the parameter samples with an expected improvement degree within a preset range are selected, and crossover, mutation and selection operations are performed in combination with a genetic algorithm until the parameter samples satisfy a preset convergence condition, so as to determine a parameter combination that minimizes the nitrogen oxide emission value.

9. The method of claim 8, wherein, The genetic algorithm comprises: A plurality of groups of parameter samples with the minimum nitrogen oxide emission values are selected as an initial population; Uniform crossover and Gaussian mutation operations are performed on the initial population, wherein a standard deviation of the Gaussian mutation operation is set to 5% of a value range of the parameter samples; In a case where an expected improvement degree of the parameter samples is less than 1% in a continuous iteration process of 10 generations or the number of iterations reaches a preset maximum value, the genetic algorithm is terminated.

10. The method of any one of claim 9, characterized in that, The method further comprises: In a case where a deviation between an experimentally determined total nitrogen oxide emission value and a predicted value of the total amount of nitrogen oxides exceeds a preset threshold, local operating condition data are supplemented and the physical information neural network model is retrained; A hydrogen mixing ratio range and a premixing temperature range in input data of the physical information neural network model are adjusted to improve accuracy of the physical information neural network model.

Images