CNN-PINN-based gas turbine combustion chamber performance prediction method

Abstract

The invention provides a CNN-PINN-based gas turbine combustion chamber performance prediction method, belongs to the technical field of combustion chamber combustion, and aims to solve the problems of high resource consumption and high grid dependence of the conventional CFD calculation at present, and the method comprises the steps: S1, collecting historical operation data of a combustion chamber, generating simulation data through numerical simulation, and carrying out the calculation of the simulation data; performing comparative analysis on the simulation data and the collected experimental operation data to prepare a data set; s2, preprocessing the data set; s3, establishing a CNN-PINN neural network prediction model according to the data characteristics and the target; s4, establishing a loss function containing a data error term and a physical information error term; s5, training and evaluating the established model by using the preprocessed data; and S6, performing performance prediction on target data by using the trained model, and performing reverse normalization on a prediction result to obtain prediction data corresponding to multiple targets.

Description

TECHNICAL FIELD

[0001] The application relates to a CNN-PINN-based gas turbine combustor performance prediction method and belongs to the technical field of combustor combustion. BACKGROUND

[0002] As a key industrial equipment, the gas turbine occupies a core position in the fields of power generation, aviation and ship power, etc. With the driving of the international carbon neutralization target, the performance optimization of the combustor is facing multiple challenges. The overall performance requirements such as low emission, high efficiency and operation stability are increasingly improved. The traditional performance prediction method mainly relies on CFD simulation calculation and experience model. In the traditional CFD simulation calculation, the grid quality is sensitive. The unstructured grid may cause the numerical error to increase. When dealing with complex geometry and physical field problems, the calculation efficiency is low. The NOx prediction relies on the experience formula or CFD simulation. There are problems such as slow speed and large consumption of computing resources. At the same time, the existing methods mainly focus on single parameter optimization. The multi-objective collaborative design of combustion efficiency, pressure loss, outlet temperature distribution and emission is lacked. It is difficult to meet the global design requirements of the combustor.

[0003] At present, deep learning provides a new idea for modeling of the combustion field. However, the pure data-driven model relies on a large amount of labeled data and lacks physical interpretability. In the field of combustion with high experimental cost, the applicability is limited. The physical information neural network (PINN) partially alleviates this problem by embedding the control equation into the loss function. However, the fully connected architecture is difficult to efficiently extract the spatiotemporal multi-scale features of the combustor flow field, and the processing capacity of the high-dimensional physical field is insufficient. SUMMARY

[0004] The application is proposed to solve the problems of large consumption of traditional CFD calculation resources and strong grid dependence, and further proposes a CNN-PINN-based gas turbine combustor performance prediction method.

[0005] The technical scheme adopted by the application to solve the above problems is as follows: Step 1: Collecting combustor historical operation data, sequentially performing numerical simulation and data analysis on the collected combustor historical operation data, and establishing a performance database; Step 2: Preprocessing the data in the performance database; Step 3: Establishing a CNN-PINN neural network model; Step 4: Establishing a loss function containing data error items and physical information error items; Step 5: Using the preprocessed performance database data, combining the established loss function and adopting a phased training strategy to train the CNN-PINN neural network model; Step 6: input the target data into the trained CNN-PINN neural network model to obtain multi-target parameter prediction results, and perform inverse normalization processing on the multi-target parameter prediction results to obtain actual values of the combustion efficiency, pressure loss, outlet temperature distribution, and NOx emission.

[0006] Further, step 1 specifically includes: collecting combustion chamber historical operation data, wherein the combustion chamber historical operation data includes combustion efficiency, pressure loss, outlet temperature distribution, and NOx emission data, and the combustion efficiency, pressure loss, and NOx emission data constitute combustion chamber flow field time series data; establishing a combustion chamber mathematical model through modeling software, and performing numerical simulation of the combustion chamber mathematical model under different operating conditions to obtain simulation data, wherein the simulation data includes simulation data of the combustion efficiency, pressure loss, outlet temperature distribution, and NOx emission; comparing and analyzing the simulation data with the collected combustion chamber historical operation data, and establishing a performance database based on the simulation data after the comparison and analysis.

[0007] Further, step 2 specifically includes: performing time synchronization and space matching processing on the data in the performance database; for the noise of the data after time synchronization and space matching, using sliding window moving average processing, and simultaneously removing the noise under different operating conditions; performing outlier processing on the data after noise removal, wherein the outlier processing includes outlier removal and outlier filling processing; performing target-specific normalization and standardization processing on the data after outlier processing, wherein the target-specific normalization and standardization processing specifically includes: using Min-Max normalization to normalize the combustion efficiency and NOx emission data to the [0, 1] interval, using Z-Score standardization processing to normalize the pressure loss data, and using global normalization processing to normalize the outlet temperature distribution data; dividing the data after target normalization and standardization processing into a training set and a validation set in a ratio of 8:2.

[0008] Further, the CNN-PINN neural network model established in step 3 includes an input layer, a feature extraction layer, a convolution layer, a pooling layer, a full connection layer, and an output layer; the input layer is used to receive preprocessed data; the feature extraction layer is a multi-channel structure, wherein the number of channels corresponds to the dimension of the input data, and is used to extract common features of the combustion chamber flow field and temperature field, establish a shared feature extraction layer, perform multi-modal data fusion on the combustion chamber flow field time series data and outlet spatial temperature distribution data, and perform time series feature extraction on the combustion chamber flow field time series data to convert 1D time series signals into pseudo 2D images; The convolutional layer includes a 1D convolution kernel and a 2D convolution kernel, the 1D convolution kernel is used to process the features of the extracted combustion chamber flow field time series data, and the 2D convolution kernel is used to process the features of the extracted spatial temperature distribution data; The pooling layer is used to reduce the dimensionality of the features of the combustion chamber flow field time series data and the features of the outlet spatial temperature distribution data after convolution processing; The fully connected layer is used to integrate the features of the combustion chamber flow field time series data and the features of the outlet spatial temperature distribution data after dimensionality reduction, and to perform secondary dimensionality reduction on the integrated features, and to map the integrated features after secondary dimensionality reduction to the target output layer.

[0009] Further, the loss function established in step 4 includes data error terms and physical information error terms, and specifically includes: combustion efficiency loss function, pressure loss loss function, outlet temperature distribution loss function, NOx emission loss function; The combustion efficiency loss function includes data error terms and physical information error terms, the physical information error terms of the combustion efficiency loss function include energy balance constraints and unburned hydrocarbon constraints; The pressure loss loss function includes data error terms and physical information error terms, the physical information error terms of the pressure loss loss function include NS equation residuals ; The outlet temperature distribution loss function includes data error terms and physical information error terms, the physical information error terms of the outlet temperature distribution loss function include energy balance constraints and radiation heat transfer constraints ; The NOx emission loss function includes data error terms and physical information error terms, the physical information error terms of the NOx emission loss function include Zeldovich reaction rate residuals , mass conservation residuals and transport equation residuals ; The total loss function is established by weighted sum of the combustion efficiency loss function, the pressure loss loss function, the outlet temperature distribution loss function and the NOx emission loss function , wherein the weights in the weighting process are obtained by one of grid search and adaptive adjustment strategy optimization.

[0010] Further, step 5 specifically includes: Step 5.1: Set the hyperparameters in the training process, and use grid search to optimize the hyperparameters, traverse all hyperparameter combinations, wherein the hyperparameters include but are not limited to learning rate, loss weight and training data batch size; Step 5.2: Calculate the parameter gradient of each layer in the CNN-PINN neural network model by automatic differentiation through the back propagation algorithm, and update the weights according to the set learning rate using the AdamW optimizer; Step 5.3: Adopt a phased training strategy and train the CNN-PINN neural network model according to the training set; Step 5.4: After training is completed, input the test set into the trained CNN-PINN neural network model, use RMES and MAE as evaluation indicators for training evaluation, and optimize the CNN-PINN neural network model according to the evaluation results.

[0011] Further, the phased training strategy adopted in step 5.3 includes: In the first phase of training, the data error terms of the combustion efficiency loss function, the pressure loss loss function, the outlet temperature distribution loss function and the NOx emission loss function are optimized; In the second phase of training, the weight of the physical information error term of the combustion efficiency loss function, the pressure loss loss function, the outlet temperature distribution loss function and the NOx emission loss function is gradually increased; In the third phase of training, the data error term and the physical information error term of the combustion efficiency loss function, the pressure loss loss function and the outlet temperature distribution loss function are jointly optimized until the target reaches the maximum number of iterations or the target loss function value is lower than the target loss function value, and the training is stopped through the early stopping mechanism, and the training is terminated when the training result is unchanged for 10 consecutive rounds.

[0012] Further, the multi-target parameter prediction results in step 6 are de-normalized, including: The combustion efficiency and NOx emission data in the prediction results are de-normalized by Min-Max, the pressure loss in the prediction results is de-normalized by Z-Score, and the outlet temperature distribution in the prediction results is de-normalized globally.

[0013] The beneficial effects of the present application are: 1. The present application first collects historical operation data and selects OpenFoam for numerical simulation, compares and verifies the two, and generates a gas turbine combustion chamber performance database, establishes a neural network prediction model according to the convolutional neural network and physical information characteristics to predict the combustion chamber performance indicators, and embeds the corresponding physical constraints of each performance indicator into the loss function, which enhances the efficiency, interpretability and generalization ability of the model.

[0014] 2. This invention extracts local and global spatiotemporal features from unstructured grid data (such as flow field velocity and temperature images) by establishing a CNN-PINN neural network model. PINN forces the model prediction to conform to physical laws through physical equations. Furthermore, a multi-branch network architecture is constructed to achieve simultaneous prediction of multiple objective parameters such as combustion efficiency, pressure loss, outlet temperature distribution, and NOx emissions.

[0015] 3. This invention combines feature extraction capabilities with strong constraints from physical equations, which not only improves the efficiency of predicting the performance of gas turbine combustors and effectively reduces the huge computational costs associated with traditional CFD when designing combustors, but also compensates for the lack of data in previous convolutional neural networks by applying physical laws, thereby improving prediction accuracy and reducing the time cost for researchers when designing combustors. It provides data support for combustor design, has interpretability and reliability, and meets the needs of researchers for global combustor design. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating a method for predicting the performance of a gas turbine combustor based on CNN-PINN. Figure 2 This is a network structure diagram for predicting the performance of a gas turbine combustor based on CNN-PINN. Detailed Implementation

[0017] Combination Figure 1 and Figure 2 This implementation method is described as follows: Figure 1 As shown, the steps of the gas turbine combustor performance prediction method based on CNN-PINN described in this embodiment include: S1: Collect historical operating data of the combustion chamber, generate simulation data through numerical simulation, compare and analyze the simulation data with the collected experimental operating data, and create a dataset; This implementation method first collects experimental data on air inlet and outlet velocities, fuel inlet and outlet velocities, inlet pressure, outlet pressure, combustion chamber inlet and outlet pressure difference, combustion temperature, component concentrations, ambient temperature field data, and NOx concentration by deploying temperature sensors, pressure sensors, differential pressure sensors, a flue gas analyzer, a Raman spectrometer, a particle image velocimeter, and a multi-point thermocouple array. Strict adherence to operating procedures and standards is required during data acquisition to ensure accuracy and reliability, and to guarantee the effectiveness of subsequent data comparison and analysis.

[0018] After collecting experimental data, this implementation method imports the combustion chamber model into OpenFoam. OpenFoam is used to generate a mesh, and the mesh is refined in the flame zone, nozzle area, and near the wall to ensure accurate temperature gradient and reaction rate. ReactingFoam (a transient combustion solver) is selected. The turbulence equation is chosen from Reynolds Average Number (RANS) equations. For the turbulence and combustion models, the EDC model suitable for detailed chemical reactions was selected. Key boundary conditions (inlet, outlet, wall) were set, and corresponding solution parameters were configured. The FvmDOM model was also added to account for the influence of radiative heat transfer on the temperature field. For NOx, the main combustion field was calculated first, and under the assumption of a frozen flow field, the NOx transport equation was solved through post-processing. The structure was then simplified to obtain simulation data. This simulation data was compared and analyzed with collected experimental data to create a dataset.

[0019] S2: Preprocess the dataset; S201: Perform time synchronization and spatial matching processing on multi-source heterogeneous data. S202: Clean up noise and outliers in the dataset: Use a sliding window moving average to process noise, and remove noise separately for different operating conditions to avoid global smoothing from masking local features; use deletion or imputation methods for outliers. For time series, the imputation method can be forward or backward padding. S203: Perform sub-objective normalization and standardization on the data: Data normalization unifies features of different dimensions and ranges to a uniform scale, thereby improving model performance and accelerating model convergence, as detailed below: Min-Max normalization was applied to combustion efficiency and NOx, linearly scaling the data to the [0,1] interval: (1); (2); Z-score normalization of pressure loss: (3); Globally normalize the outlet temperature distribution (normalize the entire temperature field data uniformly): (4); S204: Dataset Partitioning: The dataset is divided into training and validation sets using 5-fold cross-validation, with 80% of the data used for training and 20% for validation each time. This helps balance computational efficiency and evaluation stability, reduces the variance of model evaluation results, and improves the accuracy of evaluation.

[0020] S3: Establish a CNN-PINN neural network prediction model based on data characteristics and objectives; As Figure 2 shown, the CNN-PINN neural network prediction model established by the embodiment includes an input layer, a feature extraction layer, a convolution layer, a pooling layer, a fully connected layer, and an output layer, and the details are as follows: S301: Determine the input layer structure: the structure of the network input layer directly determines how the model accepts and processes data. The common features of the combustion chamber flow field and temperature field are extracted by using CNN to establish a shared feature extraction layer; the input layer is set to a multi-channel structure according to the data dimension, supporting the fusion of combustion chamber flow field time series signals and outlet space temperature distribution multi-modal data, and the time series features of the time series signals need to be extracted to facilitate the conversion of 1D time series signals into pseudo 2D images; S302: Determine the depth of the neural network: the design of the depth of the neural network needs to consider many factors, such as task complexity, model type, data size and diversity, computing resources, etc. First, a shallow basic model is constructed, the depth is gradually increased to observe the performance change, the training is dynamically monitored, and a reasonable depth is selected, and a balance is found between the number of layers and the computing resources; S303: Determine the filter size and number: the size and number of filters are key hyperparameters that determine the performance and efficiency of the model. A 2D convolution kernel is used for spatial temperature distribution, and a 1D convolution kernel is used for time series data. In the kernel size selection, small kernels are preferred, and task adaptation problems are considered, such as using long kernels for time series data; the number of kernels is increased from shallow to deep, and the computing hardware constraints are considered; S304: Determine the pooling layer size and depth: the main functions of the pooling layer include dimension reduction, computational complexity, overfitting prevention, and translation invariance enhancement. The default window size for time series features is 3, and for position-sensitive tasks, the pooling is reduced or step convolution is used; the number of pools gradually decreases from deep to shallow, the dimension of the deep network is gradually reduced, and the shallow network reduces the pooling to preserve details. A balance is achieved between dimension reduction efficiency and feature preservation; S305: Design the fully connected layer: the fully connected layer is used to map high-dimensional features to target values, directly affecting the feature integration ability and generalization performance of the model. For the multi-target output of combustion efficiency, pressure loss, outlet temperature distribution, and NOx emission, reasonable dimension reduction is performed on high-dimensional features, the output layer is simplified for regression tasks, regularization is used to prevent overfitting, and the model is ensured to have high precision and good robustness under limited data.

[0021] The present application extracts local and global spatio-temporal features from unstructured grid data (such as flow field velocity and temperature images) by establishing a CNN-PINN neural network model, and the PINN constrains the model prediction to comply with the physical law, further constructs a multi-branch network architecture, and realizes the synchronous prediction of multi-target parameters such as combustion efficiency, pressure loss, outlet temperature distribution, and NOx emission.

[0022] S4: establishing a loss function including data error term and physical information error term; The loss function established by the embodiment includes a combustion efficiency loss function, a pressure loss loss function, an outlet temperature distribution loss function, and a NOx emission loss function, each of which includes a data error term and a physical information error term, and the details are as follows: (1) Combustion efficiency loss function: Data error term: (5); Physical information constraint term: Energy balance constraint: (6); In formulas (5) and (6), the actual heat release , the theoretical heat release , the heat loss , is the fuel consumption rate, is the fuel heat value, is the fuel mass flow rate, is the radiation emissivity, is the Stefan-Boltzmann constant, is the convective heat transfer coefficient, is the ambient temperature.

[0023] Unburned hydrocarbon constraint: (7); In formula (7), is an empirical coefficient, and the combustion efficiency .

[0024] (2) Pressure loss loss function: Data error term: (8); Physical information constraint term: NS equation residual: (9); wherein is the body force, is the dynamic viscosity.

[0025] (3) Outlet temperature distribution loss function: Data error term: (10); Physical constraint term: Energy equation residual: (11); In formula (11), is the thermal conductivity, is the chemical reaction heat release rate, is the radiation heat loss rate.

[0026] Radiation heat transfer constraint: (12); In formula (12), is the radiation intensity, is the wall normal vector.

[0027] (4) NOx emission loss function: The Zeldovich mechanism is an important reaction mechanism for NOx generation, mainly including two reactions: (13); (14); The reaction rate equation is introduced into the loss function to ensure that the NO generation rate predicted by the neural network is consistent with the mechanism. The reaction rate , is defined by the Arrhenius equation: (15); In formula (16), is the pre-exponential factor, is the activation energy, is the temperature, is the temperature constant (usually 0 or 0.5).

[0028] Data error term: (16); Zeldovich reaction rate residual: (17); Mass conservation residual: (18); In formula (18), , is the chemical reaction source term, is the flow rate field.

[0029] Transport equation residual: (19); In formula (19), is the diffusion coefficient, is the reaction source term.

[0030] The total loss function is established by weighted sum of the combustion efficiency loss function, the pressure loss loss function, the outlet temperature distribution loss function and the NOx emission loss function : (20).

[0031] S5: Use the preprocessed data to train and evaluate the established model; S501: The AdamW optimizer is used to improve the weight decay and prevent overfitting in combination with the model and data characteristics; S502: Set the learning rate, loss weight, batch size and other hyperparameters, and select the grid search hyperparameter optimization method to traverse the combination in a small range; S503: The gradient of each layer parameter is calculated by the back propagation algorithm and automatic differentiation, and the optimizer updates the weight according to the learning rate; S504: The data set in step S1 is input into the prediction model for phased multiple iteration training, the first phase only optimizes the data loss (Ldata) ), the second phase gradually increases the physical loss (Lphy gradually increases ), enhances the physical consistency, and the third phase optimizes all loss terms to balance the multi-objective performance; until the target reaches the maximum iteration number or the target loss function value is lower than the target loss function value, stop training, and add an early stopping mechanism, if the performance does not improve for 10 consecutive rounds, stop training; S505: The trained model is evaluated using the test set, and RMES and MAE are used as evaluation indicators to optimize the model Root mean square error: (21); Mean absolute error: (22); In formulas (20) and (21), N is the number of prediction results, is the i prediction result, is the true value corresponding to the prediction result.

[0032] S6: Use the trained model to predict the performance of the target data, and perform inverse normalization on the prediction results to obtain the corresponding prediction data of the multi-objective.

[0033] The combustion efficiency and NOx emission data in the prediction results are normalized by Min-Max; the pressure loss in the prediction results is normalized by Z-Score; and the outlet temperature distribution in the prediction results is globally normalized.

[0034] In summary, the present application combines the feature extraction capability with the strong constraint of the physical equation, not only improves the efficiency of the gas turbine combustor performance prediction, but also effectively reduces the huge problem of using traditional CFD calculation resources when designing the combustor, greatly reduces the calculation cost, and makes up for the defects of insufficient data of the previous convolutional neural network through physical laws, improves the prediction accuracy, and reduces the time cost of researchers when designing the combustor, provides data support for designing the combustor, has explainability and reliability, and meets the global design demand of researchers on the combustor.

[0035] The above merely describes preferred embodiments of the present application, and is not intended to limit the present application in any form. Although the present application has been disclosed with preferred embodiments as above, it is not intended to limit the present application, and any person skilled in the art can make some changes or modifications to the above disclosed technical contents to obtain equivalent embodiments with equivalent changes, without departing from the technical solution of the present application. Any simple modification, equivalent replacement and improvement of the above embodiments, as long as it does not depart from the technical solution of the present application, and is within the spirit and principle of the present application, shall be within the protection scope of the present application.

Claims

1. A CNN-PINN based gas turbine combustor performance prediction method, characterized in that, The application relates to a method for predicting multi-target parameters of a combustion chamber, and belongs to the technical field of combustion chamber performance prediction. The method comprises the following steps: Step 1: collecting combustion chamber historical operation data, and sequentially performing numerical simulation and data analysis on the collected combustion chamber historical operation data to establish a performance database; Step 2: preprocessing data in the performance database; Step 3: establishing a CNN-PINN neural network model; Step 4: establishing a loss function comprising data error items and physical information error items; Step 5: training the CNN-PINN neural network model by using the preprocessed performance database data, combining the established loss function and adopting a phased training strategy; 2. The CNN-PINN based gas turbine combustor performance prediction method of claim 1, wherein, Step 6: inputting target data into the trained CNN-PINN neural network model to obtain multi-target parameter prediction results, and performing anti-normalization processing on the multi-target parameter prediction results to obtain actual values of combustion efficiency, pressure loss, outlet temperature distribution and NOx emission. Step 1 specifically comprises: collecting combustion chamber historical operation data, wherein the combustion chamber historical operation data comprises combustion efficiency, pressure loss, outlet temperature distribution and NOx emission data, and the combustion efficiency, pressure loss and NOx emission data constitute combustion chamber flow field time series data; establishing a combustion chamber mathematical model by using modeling software, and performing numerical simulation on the combustion chamber mathematical model under different working conditions to obtain simulation data, wherein the simulation data comprises simulation data of combustion efficiency, pressure loss, outlet temperature distribution and NOx emission; 3. The CNN-PINN based gas turbine combustor performance prediction method of claim 1, wherein, comparing and analyzing the simulation data with the collected combustion chamber historical operation data, and establishing a performance database based on the simulation data after the comparison and analysis. Step 2 specifically comprises: performing time synchronization and space matching processing on the data in the performance database; for the noise of the data after the time synchronization and space matching, performing sliding window moving average processing, and simultaneously removing the noise under different working conditions; performing abnormal value processing on the data after the noise removal, wherein the abnormal value processing comprises abnormal value removal and abnormal value filling processing; performing target-specific normalization and standardization processing on the data after the abnormal value processing, wherein the target-specific normalization and standardization processing specifically comprises: adopting Min-Max normalization to [0, 1] interval for the combustion efficiency and NOx emission data, adopting Z-Score standardization processing for the pressure loss data, and adopting global normalization processing for the outlet temperature distribution data; 4. The CNN-PINN based gas turbine combustor performance prediction method of claim 1, wherein, dividing the data after the target normalization and standardization processing into a training set and a verification set according to a ratio of 8:

2. The CNN-PINN neural network model established in step 3 comprises an input layer, a feature extraction layer, a convolution layer, a pooling layer, a full connection layer and an output layer; the input layer is used for receiving the preprocessed data; the feature extraction layer is a multi-channel structure, wherein the number of channels corresponds to the dimension of the input data, is used for extracting common features of combustion chamber flow fields and temperature fields, establishes a shared feature extraction layer, performs multi-modal data fusion on the combustion chamber flow field time series data and the outlet space temperature distribution data, and extracts time sequence features for the combustion chamber flow field time series data to convert 1D time sequence signals into pseudo 2D images; The convolutional layer includes a 1D convolution kernel for processing the extracted combustion chamber flow field time series data features and a 2D convolution kernel for processing the extracted spatial temperature distribution data features; The pooling layer is used to reduce the dimensionality of the features of the combustion chamber flow field time series data and the outlet spatial temperature distribution data after convolution processing; The fully connected layer is used to integrate the reduced dimensionality of the combustion chamber flow field time series data and the outlet spatial temperature distribution data, and to perform secondary dimensionality reduction on the integrated features, and to map the secondary dimensionality reduction of the integrated features to the target output layer.

5. The CNN-PINN based gas turbine combustor performance prediction method of claim 1, wherein, The loss function established in step 4 includes data error terms and physical information error terms, and specifically includes: Combustion efficiency loss function, pressure loss loss function, outlet temperature distribution loss function, NOx emission loss function; The combustion efficiency loss function includes a data error term and a physical information error term, the physical information error term of the combustion efficiency loss function including an energy balance constraint and an unburned hydrocarbon constraint; The pressure loss loss function includes a data error term and a physical information error term, the physical information error term of the pressure loss loss function including an NS equation residual ; The outlet temperature profile loss function includes a data error term and a physical information error term, the physical information error term of the outlet temperature profile loss function including an energy balance constraint and a radiative heat transfer constraint ; The NOx emissions loss function includes a data error term and a physical information error term, the physical information error term of the NOx emissions loss function including a Zeldovich reaction rate residual , a mass conservation residual , and a transport equation residual ; A total loss function is established by a weighted sum of a combustion efficiency loss function, a pressure loss loss function, an outlet temperature distribution loss function, and a NOx emission loss function wherein the weights in the weighting process are obtained by one of a grid search and an adaptive adjustment strategy optimization.

6. The CNN-PINN based gas turbine combustor performance prediction method of claim 1, wherein, Step 5 specifically includes: Step 5.1: Set the hyperparameters in the training process, and use grid search to optimize the hyperparameters, traverse all hyperparameter combinations, wherein the hyperparameters include but are not limited to learning rate, loss weight and training data batch size; Step 5.2: Calculate the parameter gradient of each layer in the CNN-PINN neural network model through automatic differentiation through the backpropagation algorithm, and update the weights according to the set learning rate using the AdamW optimizer; Step 5.3: Adopt a phased training strategy, and train the CNN-PINN neural network model according to the training set; Step 5.4: After training, input the test set into the trained CNN-PINN neural network model, use RMES and MAE as evaluation indicators for training evaluation, and adjust the CNN-PINN neural network model according to the evaluation results.

7. The CNN-PINN based gas turbine combustor performance prediction method of claim 6, wherein, The phased training strategy adopted in step 5.3 includes: In the first phase of training, optimize the data error terms of the combustion efficiency loss function, the pressure loss loss function, the outlet temperature distribution loss function and the NOx emission loss function; In the second phase of training, gradually increase the weight of the physical information error terms of the combustion efficiency loss function, the pressure loss loss function, the outlet temperature distribution loss function and the NOx emission loss function; In the third phase of training, jointly optimize the data error terms and the physical information error terms of the combustion efficiency loss function, the pressure loss loss function, the outlet temperature distribution loss function, until the target reaches the maximum number of iterations or the target loss function value is lower than the target loss function value, and the training is stopped through the early stopping mechanism, and the training is terminated when the training result is unchanged for 10 consecutive rounds.

8. The CNN-PINN based gas turbine combustor performance prediction method of claim 1, wherein, The multi-target parameter prediction results in step 6 are normalized, including: Min-Max denormalization is used for combustion efficiency and NOx emission data in the prediction results; Z-Score anti-standardization is used for pressure loss in the prediction results; Global denormalization is used for outlet temperature distribution in the prediction results.

Images