Combustion chamber key geometric structure optimization system

Abstract

The invention discloses a combustion chamber key geometric structure optimization system. The optimization system specifically comprises a full-process simulation parameterization subsystem and a reduced-order agent model subsystem. The full-process simulation parameterization subsystem can automatically carry out CFD calculation work, and therefore the requirement for large-scale calculation in the field of combustion chamber geometric structure optimization is met. And the reduced-order proxy model subsystem can establish a mapping relation from geometric input to target output in a high-dimensional data set, evaluate weight contribution of each feature, realize construction of a reduced-order proxy model and finally realize structure optimization. When the system is used for designing and optimizing the combustion chamber, for a required geometric structure, the whole calculation process can be automatically carried out only by repairing parameters which are set in advance; the system supports selection of a large number of geometric structures as input variables; according to the system, important geometric parameters can be screened in a targeted mode, meanwhile, the optimal geometric parameters are given according to the target parameters, the butt joint degree is high, and adaptability is high.

Description

Technical Field

[0001] The present invention relates to an optimization system for key geometric structures of a combustion chamber, and its application fields are the geometric structure optimization design and machining tolerance formulation of a tubular combustion chamber of a gas turbine. Background Art

[0002] A gas turbine is a power machine that uses gas as a working medium to drive rotation. It generates power by heating, expanding, and releasing energy of the gas. Its characteristics of high efficiency and low emissions make it an indispensable device in modern industry and are widely used in fields such as aviation, ships, power generation, petrochemical industry, etc.

[0003] The basic structures of modern gas turbine combustion chambers mainly have three types: tubular combustion chambers, annular combustion chambers, and can-annular combustion chambers. The design of the combustion chamber usually needs to consider factors such as combustion stability, combustion efficiency, and cooling technology to achieve the best performance and reliability. In the research and development of gas turbines, the design and improvement of the geometric structure of the combustion chamber have always been an important research direction, aiming to improve the performance of the engine, reduce fuel consumption, and reduce pollutant gas emissions. By optimizing the geometric parameters of the combustion chamber, more uniform fuel and air mixing can be achieved, the combustion efficiency can be improved, the gas flow dynamic characteristics in the combustion chamber can be improved, turbulence and unstable combustion phenomena can be reduced, and the combustion efficiency and stability can be improved. Therefore, the design and optimization of the combustion chamber geometry are of great significance in the research and development of gas turbines and are one of the key factors for improving the overall performance of gas turbines.

[0004] In the field of geometric structure optimization of gas turbine combustion chambers, the methods adopted by current research methods have two defects: low automation and fewer parameters considered.

[0005] The low automation is reflected in that when studying the differences in different geometric structures, corresponding geometric models are usually established separately, and then geometric preprocessing, mesh division, calculation model setting, parameter recording, result analysis, etc. are carried out separately. When it is necessary to modify or add new geometric parameters, it is necessary to re-enter the geometric editing software, modify and improve the previous model, and then re-import the entire calculation process. Due to the limitations of the software parameter setting program, in this case, large-scale CFD calculations cannot be realized. Limited by this, fewer geometric parameters are selected (usually less than 5 design parameters), the geometric configuration is relatively simple, and the parameter combinations are set artificially.

[0006] Meanwhile, there is a lack of sufficient data sources, making it impossible to establish a mapping relationship from geometric input to target output through the CFD dataset. The selection of the final result is only compared among the artificially set parameter groups. This method of result selection can only compare the advantages and disadvantages between the set values and is difficult to find the optimal geometric parameters. Moreover, the internal structure of the combustion chamber is complex, and overall optimization often requires considering the influence of multiple geometric parameters. Although this method can provide an optimization basis for a specific parameter, it often ignores the complexity and mutual influence of the overall combustion chamber design. And it rarely considers the interaction and synergy among multiple design variables, resulting in the optimization result may not be comprehensive enough.

[0007] Therefore, the commonly used research methods at present have a large gap with engineering applications, consume a lot of ineffective time in the design and research process stages, seriously affect the project progress and even lead to the delay of the product production progress. At the same time, their results often have great limitations, resulting in less reference and guiding value for engineering practice. It can be seen that the current research methods cannot meet the requirements of complex geometric optimization. This requires the development of a new combustion chamber key geometric structure optimization system, and at the same time, the development of a dimensionality reduction method, which needs to be able to quickly obtain key sensitive parameters from the high-dimensional variable space, reduce the sampling points, and finally realize the design and optimization of the combustion chamber, and at the same time evaluate different tolerances, which can effectively guide engineering practice. Summary of the Invention

[0008] In view of the above problems commonly existing in the field of combustion chamber geometric structure optimization, the present invention proposes a combustion chamber key geometric structure optimization system. Through the full-process simulation parameterization subsystem of this system, a relatively large number of geometric structures can be supported for selection. Based on this full-process simulation parameterization subsystem, only a large number of different design combinations need to be given, and the whole process from geometric structure change to mesh generation and then to simulation can be automatically completed, and finally the target monitoring values required for research can be automatically output, thus saving a large amount of labor costs required for configuring each set of calculation cases. Finally, the reduced-order surrogate model subsystem systematically analyzes the exported data, establishes a mapping relationship from input to output, and finally finds the optimal structure through a series of dimensionality reduction optimization algorithms to provide a reference for engineering design.

[0009] The present invention is implemented by the following technical solutions:

[0010] A key geometric structure optimization system for a combustion chamber, specifically including a full-process simulation parameterization subsystem and a reduced-order surrogate model subsystem; the full-process simulation parameterization subsystem can automatically complete the full-process CFD calculation work, thus solving the demand for large-scale calculations in the field of combustion chamber geometric structure optimization; the reduced-order surrogate model subsystem can establish a mapping relationship from geometric input to target output in a high-dimensional dataset, and finally find the optimal geometric structure of the combustion chamber to achieve its structural optimization. When using this system to design and optimize the combustion chamber, for the required geometric structure, only the pre-set parameters need to be modified, and the entire calculation process can be automatically carried out; and this system supports selecting a large number of geometric structures as input variables, up to more than 50 groups of parameters; this system can specifically screen important geometric parameters, and at the same time give the optimal geometric parameters according to the target parameters, with high docking degree and strong adaptability.

[0011] In the above technical solution, further, the full-process simulation parameterization subsystem includes a combustion chamber three-dimensional geometric model construction module, a geometric structure parameterization module, a sampling module, and a CFD calculation module; the combustion chamber three-dimensional geometric model construction module is used to construct a combustion chamber three-dimensional geometric model according to the experimental sample; the geometric structure parameterization module is used to parameterize the initial geometric structure of the combustion chamber three-dimensional geometric model to generate initial geometric parameters, and this module can be used as a quick modification path for the combustion chamber three-dimensional geometric model; the sampling module is used to randomly generate a design combination of initial geometric parameters according to the combustion chamber three-dimensional geometric model; the CFD calculation module is used to perform three-dimensional simulation calculations on the design combination of initial geometric parameters.

[0012] The reduced-order surrogate model subsystem includes a DNN model training module, a SHAP analysis and dimensionality reduction module, and an optimization module; the DNN model training module is used to train a DNN model according to the three-dimensional simulation calculation results, with the input of the DNN model being the initial geometric parameters and the output being the target parameters; the SHAP analysis and dimensionality reduction module is used to perform SHAP analysis on the initial geometric parameters according to the target parameters to obtain the weights of each geometric parameter, and screen out the key geometric parameters to achieve dimensionality reduction of the parameters; the optimization module is used to find the optimal value of the target parameters according to the key geometric parameters and output the key geometric parameters corresponding to the optimal value of the target parameters.

[0013] Further, the initial geometric parameters are determined based on the combustion chamber three-dimensional geometric model, specifically including the rotation angle of the first-stage swirler vane, the radial inclination distance of the first-stage swirler vane, the circumferential inclination distance of the first-stage swirler vane, the rotation angle of the second-stage swirler vane, the radial inclination distance of the second-stage swirler vane, the circumferential inclination distance of the second-stage swirler vane, the radius of the mixing hole, the radius of the fuel hole, the radial offset distance of the fuel hole, and the circumferential offset distance of the fuel hole.

[0014] Further, the target parameter is one or more of the combustion chamber outlet temperature, combustion chamber combustion efficiency, combustion chamber outlet radial temperature distribution coefficient (RTDF), and combustion chamber outlet temperature distribution coefficient (OTDF). The basis for selecting the target parameter is as follows: Analyze the correlation of each target parameter, and use the Pearson correlation coefficient for quantitative evaluation. If there is an obvious linear positive correlation between two parameters, then select one of them as the final optimization target.

[0015] Further, the key geometric parameters include the mixing hole radius, the rotation angle of the second-stage blade, the circumferential inclination distance of the first-stage blade, the circumferential inclination distance of the second-stage blade, the radial inclination distance of the second-stage blade, and the rotation angle of the first-stage blade.

[0016] Further, the DNN model training module is used to train a DNN model based on the three-dimensional simulation calculation results. The specific method is as follows: Randomly divide the three-dimensional simulation calculation results into K folds based on K-fold cross-validation, and use each fold as a validation set to perform multiple rounds of training on the DNN model; Optimize the hyperparameters of the DNN model hidden layer node combination based on the Grid Search method; Screen the key geometric parameters based on the SHAP analysis dimensionality reduction module, and input the key geometric parameters as new initial geometric parameters into the DNN model for continued training until the model accuracy reaches above the set threshold and the retained parameter dimensions meet the calculation requirements.

[0017] Further, the SHAP analysis dimensionality reduction module is used to perform SHAP analysis on the initial geometric parameters according to the target parameter, so as to obtain the weights of each initial geometric parameter, and screen out the key geometric parameters to achieve dimensionality reduction of the parameters. Specifically:

[0018] Calculate the SHAP values of each initial geometric parameter. The initial geometric parameter with a larger SHAP value has a greater weight. The initial geometric parameter with a larger weight is used as the key geometric parameter. The specific calculation formula for the SHAP value of the initial geometric parameter is:

[0019]

[0020] f x (S) = E[f(x)|x s

[0021] where is the Shapley Value of each feature, n is the number of input geometric parameters, S is a subset of the input feature set (x 1 , x 2 , …, x n ), f(x) is the output of the DNN model, x is the input of the DNN model, f x ​It is mean prediction.

[0022] The beneficial effects of the present invention are as follows:

[0023] The present invention proposes an optimization system for the key geometric structures of a combustion chamber, which can significantly improve the calculation and optimization efficiency. Since the design of the combustion chamber involves multiple geometric parameters, the traditional single-batch optimization method can no longer meet the scientific research needs. Therefore, the present invention establishes a geometric structure parameterization module that supports a large number of input and output parameters to meet the requirements of large-scale CFD calculations. Through geometric structure parameterization, repeated and complex geometric modifications are avoided, significantly shortening the preprocessing time of geometric parameters, simplifying the design process, improving work efficiency, and realizing large-scale parametric calculations. Specifically, geometric parameterization is performed on multiple parameters such as the rotation angle, tilt distance, and fuel hole diameter of the double-stage swirler blades, and a large number of design combinations are generated using Latin hypercube sampling to obtain rich data. To analyze this data, a DNN model is established, and the hyperparameters of the two hidden layer nodes are optimized using the Grid Search method. All the data is divided into K folds (groups) through K-fold cross-validation to reduce the impact of the validation set division on the results. The SHAP analysis and dimensionality reduction module evaluates the contributions of each characteristic parameter through SHAP value analysis and performs parameter dimensionality reduction based on this. The retained key geometric parameters are used to regenerate the design combinations, which are finally input into the DNN model for calculation. These new data will be used to establish the mapping relationship between geometric parameters and target parameters, and finally, the combustion chamber is optimized through a genetic algorithm to obtain the best combination of geometric parameters.

[0024] The key of the present invention lies in the geometric parameterization of the swirler blades of the combustion chamber and the mixing holes of the fuel holes, etc. On the one hand, it realizes the function of full-process simulation calculation, and on the other hand, it solves the demand for large-scale calculations in simulation calculations. At the same time, the present invention can analyze data with the help of the DNN model to meet the more precise point-to-point analysis requirements, that is, the analysis requirements from geometric parameter input to target parameter output. At the same time, for the calculation process of the DNN model, the present invention provides a complete analysis method, including the Grid Search method for model self-optimization and K-fold cross-validation; and a new method for dimensionality reduction and optimization is adopted to effectively realize the optimization of the key geometric structures such as the swirler blades and mixing holes of the combustion chamber, making the system able to consider more parameters, have a high degree of automation, and realize full-process simulation calculation. Description of the Drawings

[0025] Figure 1 It is a schematic diagram of the working process of the optimization system for the key geometric structures of the combustion chamber of the present invention.

[0026] Figure 2 It is the geometric structure of the double-stage swirler of this model.

[0027] Figure 3It is a schematic diagram of fuel holes and mixing holes.

[0028] Figure 4 It is a 6-fold cross-validation result graph under the 22-20 node combination.

[0029] Figure 5 It is the contribution degree of each characteristic parameter and its influence on the increase or decrease of the result.

[0030] Figure 6 It is the Max-R of the dimensionality reduction search process 2 variation graph.

[0031] Figure 7 It is the R contour map of different node combinations before and after dimensionality reduction 2 contour map.

[0032] Figure 8 It is the optimization convergence curve graph of the genetic algorithm. Specific implementation manners

[0033] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0034] Such as Figure 1 A combustion chamber key geometric structure optimization system provided by an embodiment of the present invention. The system includes a full-process simulation parameterization subsystem and a reduced-order surrogate model subsystem; the full-process simulation parameterization subsystem includes a combustion chamber three-dimensional geometric model construction module, a geometric structure parameterization module, a sampling module, and a CFD calculation module; the combustion chamber three-dimensional geometric model construction module is used to construct a combustion chamber three-dimensional geometric model according to the experimental sample; the geometric structure parameterization module is used to parameterize the initial geometric structure of the combustion chamber three-dimensional geometric model to generate initial geometric parameters, and this module can be used as a quick modification path for the combustion chamber three-dimensional geometric model; the sampling module is used to randomly generate a design combination of initial geometric parameters according to the combustion chamber three-dimensional geometric model; the CFD calculation module is used to perform three-dimensional simulation calculations on the design combination of initial geometric parameters;

[0035] The reduced-order surrogate model subsystem includes a DNN model training module, a SHAP analysis and dimensionality reduction module, and an optimization module. The DNN model training module is used to train a DNN model based on the three-dimensional simulation calculation results. The input of the DNN model is the initial geometric parameters, and the output is the target parameters. The SHAP analysis and dimensionality reduction module is used to perform SHAP analysis on the initial geometric parameters according to the target parameters, so as to obtain the weights of each geometric parameter and screen out the key geometric parameters to achieve dimensionality reduction of the parameters. The optimization module is used to find the optimal value of the target parameter according to the key geometric parameters and output the key geometric parameters corresponding to the optimal value of the target parameter.

[0036] The above optimization system is used to optimize the combustor, and the specific process is as follows:

[0037] Taking a certain tubular combustor as the research object, a three-dimensional model of the tubular combustor is constructed by using the three-dimensional geometric model construction module of the combustor. On the basis of the three-dimensional model of the tubular combustor, the geometric structure parameterization module realizes the parameterization of the key geometric structures based on the SpaceClaim software. After each change of the geometric model, the new model will inherit the set parameters of the previous time, so as to ensure that all the calculation steps of the subsequent CFD calculation module are based on the modified new structure and the preset calculation parameters, thus avoiding the cumbersome workload caused by repeated modification of the geometric structure. The full-process simulation parameterization subsystem is built based on the Ansys Workbench software. Among them, the numerical values of the geometric structure changes will be used as input parameters. For different design points (structure 1, structure 2, structure 3... structure n), they will be sequentially imported into the CFD calculation module for calculation in the subsequent simulation process. A large number of design points are generated through Latin hypercube sampling for calculation. When the full-process simulation parameterization subsystem completes the calculation and exports the data set, a neural network model (DNN model) is constructed to establish the mapping relationship from input to output, and the built-in Grid search method is used to cycle through and search for the best combination of hyperparameters, and the K-fold cross-validation is used to improve the stability and generalization ability of the model. Based on the model calculation results, the SHAP value analysis method is introduced to improve the interpretability of the model. Specifically, the SHAP analysis and dimensionality reduction module is used to obtain the weight ratio of different parameters to the target parameter, and the importance of different parameters can be evaluated according to the sorting of the weight ratio, and then the parameters can be screened and eliminated to achieve the goal of high-dimensional parameter dimensionality reduction. Finally, the optimization module uses the genetic algorithm to realize the optimization process in the multi-dimensional space and gives the geometric parameters corresponding to the optimal result, so as to realize the geometric optimization of the combustor.

[0038] Next, a specific example of an actual tubular combustor will be used to illustrate the system:

[0039] Such as Figure 2The cyclone structure of this model is shown. Its function inside the combustion chamber is to improve the combustion effect by introducing a swirling airflow. First, a three-dimensional geometric model of the combustion chamber is constructed using the combustion chamber three-dimensional geometric model construction module based on the actual structure, and then the geometric structure of the three-dimensional geometric model of the combustion chamber is parameterized using the geometric structure parameterization module. The primary object of parameterization is the cyclone vane. The parameterization of the first-stage vanes of the two-stage cyclone is introduced in detail, and the second-stage vanes are the same as the first-stage vanes. Figure 2 (a) shows the initial shape of the vane, which is the basic geometry for parameterization. The vane height is 10 mm, and the structure is divided into the leading edge, trailing edge, suction side, and pressure side. Figure 2 (b) shows the rotation of the vane. The rotation angle is 10°. The top surface rotates as a whole around the leading edge point, while the bottom surface remains unchanged, resulting in a change in the vane profile. Figure 2 (c) and Figure 2 (d) respectively show the circumferential and radial inclination of the vane. The inclination degree is 5 mm. The way is that the top surface is inclined as a whole circumferentially or radially, and the bottom surface remains unchanged, also resulting in a change in the vane profile. Figure 2 (e) is the overall structure of the two-stage radial cyclone.

[0040] For the two-stage radial cyclone, each stage of vanes has three parameterization methods, with a total of 6 groups of controllable parameters (P1 - P6). By combining different deformation methods, the influence of a single structural change on the whole and the superposition effect of multiple combined changes can be analyzed, providing support for subsequent optimization design.

[0041] As Figure 3 shown in (a) and (b) therein, the fuel holes and mixing holes are the key components for introducing fuel and air respectively.

[0042] The present invention will also parameterize the structures of the main combustion stage fuel holes and mixing holes, which are defined as the changes in the radius and position of the fuel holes and mixing holes. The reference radius of the mixing hole is 1 mm, and the reference radius of the fuel hole is 10 mm.

[0043] So far, the total number of input geometric parameters of this model is 10 groups, which are: the rotation angle, circumferential inclination distance, and radial inclination distance of the two-stage vanes; as well as the diameter of the main combustion stage fuel hole, the circumferential and radial error offsets of the fuel hole, and the diameter of the mixing hole. The variation ranges of the parameters are shown in Table 1 below.

[0044] Table 1 Variation ranges of parameters

[0045]

[0046]

[0047] Considering that the combustion gas at the combustor outlet flows downstream to the turbine blades, it is necessary to effectively evaluate whether the temperature distribution at the combustor outlet is uniform. The Radial Temperature Distribition Factor (RTDF) at the combustor outlet is used as the target parameter:

[0048]

[0049] where T t4max represents the highest temperature at the combustor outlet, T t4 represents the average temperature at the combustor outlet, T t3 represents the average temperature at the combustor inlet, T t4rmax represents the maximum average temperature at different radial positions at the combustor outlet. Taking the outlet temperature as the output parameter through Fluent, the RTDF of each group is calculated indirectly.

[0050] The sampling module obtains any combination within the range of geometric parameter changes through Latin hypercube sampling, and conducts large-scale CFD calculations based on the established CFD calculation module, so as to obtain a large amount of data corresponding to different design combinations.

[0051] In terms of data processing, the DNN model training module trains the DNN model according to the calculation results of the CFD calculation module, and constructs the mapping relationship from input to output through the DNN model, that is, the mapping relationship from the rotation angle, circumferential tilt distance, radial tilt distance of the double-stage swirler blades, the diameter of the primary combustion stage fuel holes, the circumferential and radial error offsets of the fuel holes, and the diameter of the mixing holes and other structures to the RTDF. In order to reduce the uncertain influence brought by the randomness of data division, the K-fold cross-validation method is used to rotate and select multiple subsets as the validation set, so as to minimize the error. At the same time, the GridSearch method is used to find the optimal combination of hyperparameters to improve the performance of the model, and finally the configuration of the two hidden layer node combinations 22 and 20 is selected.

[0052] Under this node combination, the K-fold cross-validation results of the training results of its DNN model are as Figure 4 shown.

[0053] Among them, each subgraph shows the relationship between the predicted results and the actual data in different folds. The abscissa T_NN is the predicted result, and the ordinate T_CFD is the actual CFD calculation result. Through this visualization method, the performance of the DNN model in different folds can be intuitively observed, so as to evaluate the accuracy and stability of the DNN model.

[0054] For the dimensionality reduction of 10 input parameters, the SHAP analysis dimensionality reduction module uses the SHAP analysis method as an additive DNN model interpreter. All features (geometric parameters) are considered to contribute to the final output result, and their "contribution degrees" are further evaluated. The calculation process is as follows: Define an adjoint model g that satisfies

[0055]

[0056] where, x′ j ∈ {0, 1} indicates whether the feature ψ j can be observed. Features that cannot be observed will not affect the explanation (i.e., the j-th term in the sum is 0). M is the number of input features, is the Shapley Value of each feature, and ψ 0 is a constant whose value is equal to the predicted mean of all training samples.

[0057] For a trained model, define f x (S) = E[f(x)|x s , where f x is the mean prediction, f(x) is the output of the DNN model, and S is a subset of the input feature set. For given parameters, the explanation (expectation) E[f(x)|x 1 , x 2 , …, x n of the model is solved as follows:

[0058] If the formula (3) can be obtained, that is, the sample-independent expectation of the model prediction value, which can be approximated by the average of the model prediction values of the training samples. This value is independent of the samples.

[0059]

[0060] If S = {x 1}, the formula (4) can be obtained:

[0061]

[0062] Similarly, when S = {x 1 , x 2 , …, x n}, it can be deduced that:

[0063]

[0064] To extrapolate to the general form, the calculation method of Shapley Value (SHAP value) needs to be applied to deduce the explanation formula:

[0065]

[0066] Among them, n is the total number of input geometric parameters, S is a subset of the input feature set, and f x is the mean prediction.

[0067] Based on this, SHAP analysis is introduced into the constructed DNN model to evaluate the contribution degrees of 10 feature parameters, and the results are as Figure 5 shown. The SHAP value method is used to evaluate the weights and perform dimensionality reduction on the structures such as the rotation angle of the double-stage cyclone blade, the circumferential inclination distance, the radial inclination distance, the diameter of the main combustion stage fuel hole, the circumferential and radial error offsets of the fuel hole, and the diameter of the mixing hole.

[0068] From Figure 5 it can be seen that the feature parameters with relatively large contributions to the average temperature at the outlet section of the final combustion chamber are (in descending order): the radius of the P7 mixing hole, the rotation angle of the P4 second-stage blade, the circumferential inclination distance of the P3 first-stage blade, the circumferential inclination distance of the P6 second-stage blade, the radial inclination distance of the P5 second-stage blade, and the rotation angle of the P1 first-stage blade.

[0069] So far, the evaluation of the weights of the internal geometric structure of the combustion chamber has been realized, and then dimensionality reduction processing is completed according to the requirements. As Figure 6 shown, starting from 10 parameters, the red line represents the change of max-R 2 , and the bar chart represents the total contribution of the remaining parameters to SHAP after each elimination. The degree of decrease in the bar chart reflects the importance of each parameter to the overall data set. In order to achieve effective dimensionality reduction of key parameters, a balance must be achieved between optimizing the R 2 value and the total SHAP. Finally, according to the selection process, a configuration containing a total of 6 parameters is selected as the further basic optimization. At this time, R 2 has been improved relative to the level before dimensionality reduction, and the SHAP value (0.935) remains at a relatively high level.

[0070] Select the 6 groups of geometric parameters with the highest weights, focus on optimizing them, and then bring the supplementary data of the 6 key geometric parameters into the DNN model for training. The final dimensionality reduction results are as Figure 7 shown. The picture as Figure 7 shown is a contour plot used to show the distribution of various DNN metrics (R 2 mean, training loss mean, training loss standard deviation) under different parameter combinations, where the R 2 mean is the mean of 6-fold cross-validation. Its horizontal and vertical coordinates respectively represent the number of nodes in the two-layer neural network. From Figure 7 (a) and (b), it can be seen that through the validation set R 2Reflecting the generalization ability of the DNN model, the validation set R of different node combinations 2 The data accuracy after optimization (b) is significantly improved compared with that before optimization (a). The 6-fold R 2 The mean has been raised to 0.94, and the contour gradient changes more smoothly, proving that the error fluctuation range is small and the credibility is higher. Figures (c), (d), (e), and (f) show the mean and standard deviation of the validation loss before and after optimization, respectively. It can be seen that the DNN training has been significantly improved after dimensionality reduction. At this time, the two hidden layer nodes are combined into a 22-20 configuration, and the optimal position is indicated by an asterisk in the figure. In subsequent optimization, the optimal position will be taken as 6 compromise R 2 The optimization is based on the training result of the highest fold, and the Max-R 2 The prediction accuracy has reached 0.97, and the results have good credibility.

[0071] The optimization module brings the data calculated after dimensionality reduction into the genetic algorithm for optimization, and performs 600 iterative calculations. The final optimization result is as follows Figure 8 shown.

[0072] Therefore, after optimization, the best design parameters in the search space are: P7 = 8.4, P4 = -13.5, P3 = 8.9, P6 = -0.8, P5 = -3.4, P1 = 12.2. At this time, the temperature radial distribution coefficient RTDF is reduced from 1.21 to 0.36, with an acceptable error of 10% from the predicted value, and the reduction of the target is 70%.

[0073] It can be seen from the above practical application examples that the key geometric structure optimization system for combustion chambers proposed in the present invention can solve the selection and large-scale calculation problems of a large number of design parameters in the field of combustion chambers more quickly, saving a lot of labor costs compared to general methods. In addition, the system can effectively analyze and calculate the obtained data set, realize data dimensionality reduction and optimization, and thus provide ideal geometric parameters in a targeted manner.

Claims

1. A combustion chamber key geometric structure optimization system, characterized in that: Specifically, it includes the full-process simulation parameterization subsystem and the reduced-order proxy model subsystem; The full-process simulation parameterization subsystem includes a combustion chamber three-dimensional geometric model construction module, a geometric structure parameterization module, a sampling module and a CFD calculation module; the combustion chamber three-dimensional geometric model construction module is used to construct a combustion chamber three-dimensional geometric model according to an experimental prototype; the geometric structure parameterization module is used to parameterize the initial geometric structure of the combustion chamber three-dimensional geometric model to generate initial geometric parameters; The sampling module is used to randomly generate a design combination of initial geometric parameters according to the three-dimensional geometric model of the combustion chamber; the CFD calculation module is used to perform three-dimensional simulation calculation on the design combination of the initial geometric parameters; The reduced-order proxy model subsystem includes a DNN model training module, a SHAP analysis dimension reduction module and an optimization module; The DNN model training module is used to train a DNN model according to the three-dimensional simulation calculation result, wherein the input of the DNN model is an initial geometric parameter and the output is a target parameter; The SHAP analysis and dimensionality reduction module is used to perform SHAP analysis on the initial geometric parameters according to the target parameters, thereby obtaining the weights of the initial geometric parameters, and screening out the key geometric parameters to achieve parameter dimensionality reduction; The optimization module is used to find the optimal value of the target parameter according to the key geometric parameter, and output the key geometric parameter corresponding to the optimal value of the target parameter.

2. The combustion chamber key geometry optimization system according to claim 1, characterized in that: The initial geometric parameters are determined based on the three-dimensional geometric model of the combustion chamber, and specifically include a first-stage swirler blade rotation angle, a first-stage swirler blade radial inclination distance, a first-stage swirler blade circumferential inclination distance, a second-stage swirler blade rotation angle, a second-stage swirler blade radial inclination distance, a second-stage swirler blade circumferential inclination distance, a mixing hole radius, a fuel hole radius, a fuel hole radial offset distance, and a fuel hole circumferential offset distance.

3. The combustion chamber key geometry optimization system according to claim 1, characterized in that: The target parameter is one or more of the combustion chamber outlet temperature, the combustion chamber combustion efficiency, the combustion chamber outlet radial temperature distribution coefficient and the combustion chamber outlet temperature distribution coefficient.

4. The combustion chamber key geometry optimization system according to claim 1, characterized in that: The key geometric parameters include the mixing hole radius, the second-stage blade rotation angle, the first-stage blade circumferential inclination distance, the second-stage blade circumferential inclination distance, the second-stage blade radial inclination distance, and the first-stage blade rotation angle.

5. The combustion chamber key geometry optimization system according to claim 1, characterized in that: The DNN model training module is used to train the DNN model according to the three-dimensional simulation calculation results, and the specific method is: based on K-fold cross-validation, the three-dimensional simulation calculation results are randomly divided into K folds, and each fold is used as a validation set to perform multiple rounds of training on the DNN model; based on the Grid Search method, the hyperparameters of the hidden layer node combination of the DNN model are optimized; Based on the SHAP analysis dimension reduction module, key geometric parameters are screened and the key geometric parameters are input into the DNN model as new initial geometric parameters to continue training until the model accuracy reaches above the set threshold and the retained parameter dimensions meet the calculation requirements.

6. The combustion chamber key geometry optimization system according to claim 1, characterized in that: The SHAP analysis and dimensionality reduction module is used to perform SHAP analysis on the initial geometric parameters according to the target parameters, thereby obtaining the weights of the initial geometric parameters and screening out the key geometric parameters to achieve parameter dimensionality reduction, specifically: The SHAP value of each initial geometric parameter is calculated. The initial geometric parameter with a larger SHAP value has a larger weight, and the initial geometric parameter with a larger weight is taken as the key geometric parameter. The calculation formula of the SHAP value of the initial geometric parameter is as follows: f x (S)=E[f(x)|x s ] in, is the Shapley Value of each feature, n is the number of input geometric parameters, S is the initial geometric parameter set (x1, x2, ..., x n ), f(x) is the output of the DNN model, and x is the input of the DNN model; f x is the mean forecast.

7. The combustion chamber key geometry optimization system according to claim 1, characterized in that: The optimization module is used to find the optimal value of the target parameter according to the key geometric parameters, and specifically uses a genetic algorithm to optimize the target parameter.

Images