Intelligent optimization method and system for turbine aerodynamic shape based on three-dimensional full flow field rapid prediction

Abstract

The application relates to a turbine aerodynamic shape intelligent optimization method and system based on three-dimensional full-flow field rapid prediction, and relates to the field of turbine aerodynamic shape intelligent optimization. The application solves the problems of how to predict the local high gradient area, how to balance the flow field prediction accuracy, the spatial resolution and the model training cost and the like, and the method comprises the following steps: taking a radial rotary surface as a unit sample to construct a turbine three-dimensional flow field data set, and constructing a neural network model of a turbine channel three-dimensional full-flow field; an intelligent optimization framework integrating prediction, evaluation and optimization is constructed, a blade geometric feature is described in a parameterized mode, key geometric parameters are taken as state variables of a reinforcement learning environment, aerodynamic performance indexes output by a rapid flow field prediction model under given geometric and boundary conditions are taken as reward signals, a reinforcement learning intelligent agent autonomously adjusts blade parameters in a continuous design space, and a strategy network is continuously updated to realize adaptive optimization of multi-target performance.

Description

Technical Field

[0001] This invention relates to the field of intelligent optimization technology for turbine aerodynamic shape, specifically to an intelligent optimization method and system for turbine aerodynamic shape based on rapid prediction of three-dimensional full flow field. Background Technology

[0002] The internal flow processes of aero-engines are highly complex, typically involving numerous intricate physical phenomena such as flow separation, transition, heat transfer, mass transfer, vibration, and noise. Inappropriate flow conditions are detrimental to the safe and stable operation of the engine. Therefore, with the increasing demands for refined aero-engine design, higher requirements are being placed on acquiring high-precision flow field data. Although high-fidelity simulations are becoming increasingly common due to improved computing power, for computationally intensive tasks in the early stages of turbine design and optimization, lower-cost and faster flow field evaluation methods are more suitable.

[0003] Existing technologies have incorporated deep learning methods into flow field prediction tasks, using convolutional neural networks (CNNs) to predict flow fields around irregular geometries with an accuracy of 98%. Regarding input, two geometric representation methods were compared: the signed distance function (SDF) and the blade profile binarization, without considering changes in boundary conditions. The output employs a branched network design, starting from a shared geometric information encoder network, with each flow field variable having its own dedicated decoder branch network, all with identical branch network architectures. This approach preserves the potential correlations between flow field variables while avoiding mutual interference. Results show that the SDF representation achieves superior prediction accuracy; compared to traditional CFD methods, the CNN-based approach achieves a speedup of 2-4 orders of magnitude while maintaining low prediction errors.

[0004] Subsequently, existing technologies have further utilized deep neural networks to achieve, for the first time, forward prediction and inverse problem design tasks for airfoil flow fields based on large-scale datasets. To avoid geometric information distortion caused by pixelation methods, normalized airfoil coordinates are directly used as an accurate description of the geometry. In the forward flow field prediction task, the airfoil geometry encoding network compresses the airfoil geometry into a 16-dimensional sequence, combined with additional boundary conditions such as the angle of attack (AOA) and Reynolds number (RE), and the coordinates (x and y) of the flow field sampling points as inputs to the flow field prediction network. This predicts the flow conditions of airfoils with different geometries under different inflow angles of attack and Reynolds numbers, achieving point-by-point mapping of the flow field prediction. The accuracy on the training and test sets reached 99% and 97%, respectively. In the inverse problem design task, the airfoil geometry encoding network is used as the main body, replacing the airfoil geometry with a surface pressure coefficient curve to establish a mapping from expected performance to normalized airfoil coordinates. The relative error of the airfoil on the training set is less than 0.5%, while the relative error on the test set is less than 2%. Many subsequent studies on airfoil flow field prediction have been based on this framework. Currently, two key issues that deep learning still needs to address in engineering design are how to obtain more accurate and generalizable prediction methods, and how to reduce the dependence on high-quality data and reduce the "cold start" cost in the early stages of surrogate model building.

[0005] For turbines, to address the issues of prediction accuracy and generalization, one approach is to use deep learning surrogate models to improve the Reynolds stress closure terms in the RANS turbulence model. The coupling of CFD with deep learning models can be categorized into two types: the "freeze" strategy and the "interaction" strategy. The "freeze" strategy directly feeds the final anisotropic Reynolds stress field obtained from high-fidelity simulation into the RANS equations for calculation, with the Reynolds stress field remaining constant throughout the iterations. The "interaction" strategy, on the other hand, uses a surrogate model to predict the Reynolds stress field from the current time-averaged flow field in each RANS iteration until the calculation converges.

[0006] Another approach is to introduce physical constraints and mathematical principles to improve the architecture design of direct mapping networks and enhance the interpretability of the model. Existing technologies disclose a graph computation method that designs a dual graph neural network (DGNN) model. By analogy with the finite element method for unstructured meshes, it achieves predictions from blade geometry to flow field and core performance parameters. Wang et al. proposed an attention-enhanced neural operator network based on deep operator networks and the Galerkin attention mechanism.

[0007] In summary, the application of deep learning in the rapid evaluation of flow field and performance of air-cooled turbines still faces several core challenges: the complex vortex system inside the turbine leads to the existence of a large number of high gradient regions, and how to predict these local high gradient regions remains an unsolved problem; the three-dimensional flow field of the turbine has a large data scale and high modeling cost, and how to balance the flow field prediction accuracy and spatial resolution with the model training cost is the second problem that needs to be solved. Summary of the Invention

[0008] To overcome the problems in existing technologies regarding how to predict these local high gradient regions and how to balance flow field prediction accuracy, spatial resolution, and model training costs, this invention proposes a method and system for intelligent optimization of turbine aerodynamic shape based on rapid prediction of the three-dimensional full flow field.

[0009] To solve the above-mentioned technical problems, the present invention is achieved through the following technical solution: Option 1: This invention proposes an intelligent optimization method for turbine aerodynamic shape based on rapid prediction of the three-dimensional full flow field. The method includes the following steps: Step 1: Construct a turbine three-dimensional flow field dataset using radial revolution surfaces as unit samples. The turbine three-dimensional flow field dataset includes a training set, a dataset, and a validation set. Step 2: Construct a neural network model of the three-dimensional full flow field of the turbine channel based on the turbine three-dimensional flow field dataset built in Step 1; Step 3: Based on the neural network model of the three-dimensional full flow field of the turbine channel constructed in Step 2, construct an intelligent optimization framework integrating prediction, evaluation, and optimization. Describe the blade geometry in a parameterized manner, use key geometric parameters as state variables of the reinforcement learning environment, and use the aerodynamic performance index output by the fast flow field prediction model under given geometric and boundary conditions as reward signals. The reinforcement learning agent autonomously adjusts the blade parameters in the continuous design space and continuously updates the policy network to achieve adaptive optimization of multi-objective performance.

[0010] Furthermore, a preferred embodiment is provided, wherein the method for constructing the neural network model of the three-dimensional full flow field of the turbine channel in step 2 is as follows: Based on supervised learning methods, establish a radial section airfoil... p and import / export border conditions Departure to complete blade surface field q Forward flow field prediction model The positive flow field prediction model The input data includes radial section airfoil, inlet angle of attack (AOA), and outlet isentropic Mach number (Ma2is) provided in image form, with corresponding relative radial positions.

[0011] Furthermore, a preferred embodiment is provided, in which the Adam algorithm is used to train a neural network model of the three-dimensional full flow field of the turbine channel.

[0012] Furthermore, a preferred embodiment is provided, in which a composite loss function is used to adjust the predicted and actual values ​​of the neural network model of the three-dimensional full flow field of the turbine channel. The composite loss function includes the mean square error and the pixel gradient error, and the weighting coefficient λ is set to 10, i.e.,

[0013]

[0014]

[0015] Where H and W represent the number of sampling points along the axial and circumferential directions for each radial section, respectively, and the subscripts i and j are the indices of the two directions.

[0016] Furthermore, a preferred embodiment is provided, wherein the integrated intelligent optimization framework for prediction-evaluation-optimization described in step 3 includes an overall performance term R. perf and flow field organization item R flow ;

[0017]

[0018]

[0019] in and The weighting of the two reward items is assigned respectively. and These are the efficiency and total pressure loss coefficients, respectively. The corresponding weights; and These are the scale measures of the large eddy and the scale measures of the separation, respectively. This represents the weight of the flow field organization term.

[0020] Furthermore, a preferred implementation is provided, in which normalized absolute error QE is used to predict the training set and the dataset, that is, .

[0021] Furthermore, a preferred embodiment is provided in which the ratio of the training set, validation set, and test set is 4:1:1.25.

[0022] Option 2: A turbine aerodynamic shape intelligent optimization system based on rapid prediction of three-dimensional full flow field. The system is implemented based on the method described in Option 1, and includes: The data acquisition module is used to construct a turbine three-dimensional flow field dataset using radial rotating surfaces as unit samples. The turbine three-dimensional flow field dataset includes a training set, a dataset, and a validation set. The model building module is used to construct a neural network model of the three-dimensional full flow field of the turbine channel based on the turbine three-dimensional flow field dataset constructed by the data acquisition module. The optimization module is used to construct an intelligent optimization framework integrating prediction, evaluation, and optimization based on the neural network model of the three-dimensional full flow field of the turbine channel built by the model building module. It describes the blade geometry in a parameterized way, uses key geometric parameters as state variables of the reinforcement learning environment, and uses the aerodynamic performance index output by the fast flow field prediction model under given geometric and boundary conditions as the reward signal. The reinforcement learning agent autonomously adjusts the blade parameters in the continuous design space and continuously updates the policy network to achieve adaptive optimization of multi-objective performance.

[0023] Option 3: A computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the method described in Option 1.

[0024] Option 4: A computer device, including a memory and a processor, wherein the memory stores a computer program, and when the processor runs the computer program stored in the memory, the processor executes the method described in Option 1.

[0025] The advantages of this invention are: This invention discloses an intelligent optimization method and system for turbine aerodynamic profiles based on rapid prediction of the three-dimensional full flow field. It focuses on the rapid prediction of the complex three-dimensional aerodynamic flow field of a turbine and the intelligent optimization of key blade geometric parameters. By inputting blade geometry parameters and boundary conditions, the rapid prediction model can efficiently solve the three-dimensional flow field of the turbine under specified operating conditions and thereby calculate the overall performance parameters. These performance parameters serve as evaluation indicators for the optimization algorithm, enabling intelligent optimization of the turbine aerodynamic profile geometry and providing a reliable means for the efficient optimization design of similar turbomachinery components and the entire machine.

[0026] This invention is also applicable to applications such as the optimization design of complex aerodynamic profiles for air-cooled turbine blades in aero engines and gas turbines. Attached Figure Description

[0027] Picture 1 This is a schematic diagram of the neural network architecture for full flow field prediction as described in Implementation Method 1.

[0028] Picture 2 The low-temperature precipitate described in Implementation Method 1 Ma 2is A schematic diagram comparing the label values ​​and model prediction values ​​of the pressure distribution curve at a typical cross-section.

[0029] Among them, (a) represents 5% leaf height. Ma 2is =0.4 cross section pressure distribution curve schematic diagram, (b) is 5% blade height, Ma 2is =0.778 cross section pressure distribution curve schematic diagram, (c) is 50% blade height, Ma 2is A schematic diagram of the pressure distribution curve at section 0.4, where (d) represents 50% blade height. Ma 2is =0.778 cross section pressure distribution curve schematic diagram, (e) is 95% blade height, Ma 2is A schematic diagram of the pressure distribution curve at section 0.4, where (f) represents the 95% blade height. Ma 2is =0.778 cross section pressure distribution curve schematic diagram.

[0030] Picture 3 The high-performance version described in Implementation Method 1 Ma 2is A schematic diagram of the pressure distribution curve at a typical cross-section.

[0031] Among them, (a) represents 5% leaf height. Ma 2is =1.0 cross section pressure distribution curve schematic diagram, (b) is 5% blade height, Ma 2is =1.2 Cross-section pressure distribution curve schematic diagram, (c) is 50% blade height, Ma 2is =1.0 cross section pressure distribution curve schematic diagram, (d) is 50% blade height, Ma 2is =1.2 Cross-section pressure distribution curve schematic diagram, (e) is 95% blade height, Ma 2is =1.0 cross section pressure distribution curve schematic diagram, (f) is 95% blade height, Ma 2is =1.2 Schematic diagram of pressure distribution curve at section 1.2.

[0032] Picture 4 As described in Implementation Method 1 Ma 2is A schematic diagram of the predicted flow field at a mean diameter of 0.778.

[0033] Picture 5 As described in Implementation Method 1 Ma 2is A schematic diagram of the predicted flow field at a mean diameter of 1.2.

[0034] Picture 6This is a comparative schematic diagram of the predicted vortex structure using the low Mach number operating condition model described in Implementation Method 1.

[0035] Among them, (a) is Ma 2is =0.4 Schematic diagram of the predicted vortex structure using the working condition model, (b) is Ma 2is =0.778 Schematic diagram of the predicted vortex structure using the working condition model.

[0036] Picture 7 This is a comparative schematic diagram of the predicted vortex structure using the high Mach number operating condition model described in Implementation Method 1.

[0037] Among them, (a) is Ma 2is =1.0 working condition model prediction of vortex structure schematic diagram, (b) is Ma 2is =1.2 Schematic diagram of vortex structure predicted by the working condition model.

[0038] Picture 8 This is a schematic diagram of the reinforcement learning optimization results described in Implementation Method 1.

[0039] Among them, (a) is a schematic diagram of the vortex structure before optimization, (b) is a schematic diagram of the vortex structure after optimization, (c) is a schematic diagram of the blade geometry and pressure distribution before optimization, and (d) is a schematic diagram of the blade geometry and pressure distribution after optimization. Detailed Implementation

[0040] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them.

[0041] Implementation Method 1, see Picture 1 to Picture 8 This embodiment describes a method and system for intelligent optimization of turbine aerodynamic shape based on rapid prediction of the three-dimensional full flow field, specifically including the following steps: 1. Strategy for constructing a 3D turbine flow field dataset around a radially rotating surface To meet the requirements of efficient storage and computation, a uniform structured grid is selected to extract and represent the full flow field data. The flow field characteristics of the rotating surfaces segmented along different radial positions have higher similarity and can best characterize the aerodynamic performance of turbines in engineering practice. Therefore, the radial rotating surfaces are selected as unit samples to construct the turbine three-dimensional flow field dataset.

[0042] 2. Multi-path self-attention network architecture for turbine 3D flow field prediction The goal of this section is to establish a system based on supervised learning methods using radial section airfoils.p and import / export border conditions Departure to complete blade surface field q Forward flow field prediction model The current problem is to find a set of model parameters. θ , making .

[0043] Input data includes radial section airfoil and inlet angle of attack provided in image form. AOA export isentropic Mach number Ma 2is The corresponding relative radial position. During training, data is input in mini-batches with a batch size of 50. The Adam algorithm is used.

[188] The model is trained using an exponential learning rate scheduler with an initial learning rate of 1×10⁻⁶. -4 The decay factor is 0.9, and the learning rate steps to 0.9 times the previous value after every 250 iterations. The loss function chosen is the same composite loss function defined in Equation 4-5, which includes mean squared error and pixel gradient error, and the weight coefficient λ is set to 10.

[0044]

[0045]

[0046]

[0047] Where H and W represent the number of sampling points along the axial and circumferential directions for each radial section, respectively, and the subscripts i and j are the indices of the two directions.

[0048] 3. A reinforcement learning-based method for turbine blade geometry optimization This method relies on an established neural network model capable of rapidly and accurately predicting the three-dimensional full flow field of turbine channels. It uses reinforcement learning algorithms as its core to construct an integrated intelligent optimization framework encompassing prediction, evaluation, and optimization. Blade geometry is described parametrically, with key geometric parameters serving as state variables in the reinforcement learning environment. The aerodynamic performance indicators output by the rapid flow field prediction model under given geometric and boundary conditions are used as reward signals. The reinforcement learning agent autonomously adjusts blade parameters within a continuous design space, continuously updating the policy network to achieve adaptive optimization of multi-objective performance. This method achieves global optimization of complex nonlinear aerodynamic responses without explicitly solving for the gradient of the objective function. The most crucial reward function of this algorithm includes the overall performance term R. perf and flow field organization item R flow .

[0049]

[0050]

[0051]

[0052] in and The weighting of the two reward items is assigned respectively. and These are the efficiency and total pressure loss coefficients, respectively. The corresponding weights; and These are the scale measures of the large eddy and the scale measures of the separation, respectively. This represents the weight of the flow field organization term.

[0053] After 4000 rounds of training, the model's mixed loss residual converged to 2.94 × 10⁻⁶. -3 Table 1 lists the mean squared error, normalized error, and gradient error indices of the model on the training and test sets under different input conditions. Since the velocity component's value range includes 0, the conventionally defined relative error cannot be used; therefore, a normalized absolute error QE is introduced:

[0054] Table 1 Statistical evaluation of the current model's performance

[0055] The results show that although the training set only used 16% of the complete dataset, the model performed consistently well across both datasets, with an average QE of approximately 0.22% and an average gradient loss (GL) below 3 × 10⁻⁶. -4 This indicates that it can achieve accurate prediction of high-quality flow fields from boundary conditions.

[0056] Picture 2 and Picture 3 The pressure distribution curves for three typical spanwise sections under four working conditions are presented. The wall pressure distribution is dimensionless according to the following formula. Overall, the model predictions and CFD reference values ​​maintain a high degree of fit, with their curves largely overlapping. Comparisons are made for different... Ma 2is The accuracy of the predictions did not appear to show a significant correlation with the operating conditions, regardless of high... Ma 2is Still low Ma 2is Both the operating conditions and the predicted values ​​maintained a high level of fitting accuracy.

[0057]

[0058] Picture 4 and Picture 5The predicted contour plots and relative error distributions of six basic flow field physical quantities at the mid-diameter section are shown for four different operating conditions. Each subplot displays the CFD reference value, model prediction value, and normalized error value from left to right; the first two are combined for easier comparison. The values ​​are arranged from top to bottom according to density. ρ static pressure P , static temperature T and the three velocity components in the rectangular coordinate system u , v , w The arrangement is such that the velocity in the Cartesian coordinate system is converted from the velocity in the cylindrical coordinate system. Z The axis is a rotation axis.

[0059]

[0060]

[0061]

[0062]

[0063] Where the subscripts r, c, and z represent radial, circumferential, and axial velocities, respectively. The azimuth angle of the sampling point.

[0064] Picture 3 The results show that the predicted values ​​and CFD values ​​fit very well, with most normalization errors less than 1%. Although when Ma 2is When the coefficient of variation is 1.2, the model prediction error is high, which is related to the existence of complex shock waves and expansion wave systems in the flow field at high Mach numbers. Similar situations also commonly occur in the wake region under various operating conditions. Due to the use of uniform grid sampling, the wake region, which has a more complex flow, and the inlet section, which has a relatively gentle flow, use the same grid size, resulting in an increase in the gradient between adjacent samples in the wake region. However, this effect can be mitigated as much as possible through gradient loss constraints.

[0065] In engineering practice, the RANS method is often used to obtain the time-averaged characteristics of large-scale vortex structures in turbine channels and to analyze the loss situation. Picture 6 and 7 The images show the model's reconstruction of typical vortex fields within the turbine under four different operating conditions. The left column represents CFD reference values, and the right column represents model predictions. Vortices are visualized using the Q-criterion, with the identification threshold set to Q = 1 × 10⁻⁶. 4 s -2 The results show that, under most operating conditions, the model predictions accurately capture and reconstruct the typical vortex structure inside the guide vane.

[0066] Table 2 summarizes the time consumption of each stage of model training and prediction. The cost of model training is close to 50% of the data accumulation stage of CFD computation. However, after the model is trained, performing a full-scale prediction on a case with an equivalent grid of approximately 4.2 million grid cells takes only 9 seconds, which is 0.5% of the time required to perform numerical solutions on the same grid size. This means that the trained model has higher time efficiency in the optimization process compared to traditional CFD methods.

[0067] Table 2 Time Costs of Model Training and Data Accumulation

[0068] After using reinforcement learning, the turbine blade profile and the three-dimensional stacking method were optimized. By applying bending and torsion in the end region, the low-energy fluid near the wall was drawn to the middle of the flow channel, and the total pressure loss coefficient was reduced from 0.49 to 0.47.

[0069] Example 1: Data Preparation In practical applications, if targeted training of the network is required, sufficient experimental data should be prepared, and data extraction and preprocessing should be performed. 256, 128, and 128 sampling points should be set along the axial, circumferential, and radial directions, respectively. Each sampling point stores basic flow field physical quantities, including density. ρ static pressure P , static temperature T axial velocity u z Circumferential velocity u c and radial velocity u r Generally, the static temperature and static pressure of the flow field inside the turbine will not exceed the total temperature and total pressure of the incoming flow. To avoid non-physical solutions, the 0-1 normalization method is selected to calculate the maximum and minimum values ​​of each physical quantity in the flow field, scale the data to the range of [0, 1], and at the same time, the output data is forcibly truncated to satisfy the extreme value constraints.

[0070]

[0071] Typically, 3D full flow field data is large in scale, and complete training requires high computational costs. To quickly obtain a reasonable model, a smaller base dataset is extracted from it. It is generally recommended to randomly extract 32 radial sections for training under each working condition, which is equivalent to 1 / 4 of the complete data. The base dataset is divided into training, validation, and test sets in a ratio of 4:1:1.25.

[0072] 2. Model Training The network and its associated code are all written in Python and built on the open-source deep learning framework PaddlePaddle. When training the network on a computer platform equipped with specific hardware, it is generally recommended to use the Adam optimizer with default parameters.

[0073] 3. Performance Verification After model completion, the model's performance on the dataset is evaluated using mean squared error, relative error, peak signal-to-noise ratio (PSNR), and structured similarity criterion. Statistical metrics are compiled into a results table. Full flow field contour plots of the predicted and actual values ​​for the characteristic blade height section are plotted, along with violin plots of the PNR and structured similarity criterion distributions. Subsequently, pressure coefficient and isentropic Mach number curves for the characteristic blade height section are plotted. An appropriate vortex identification method is selected to solve for the vortex structure within the turbine passage. Turbine efficiency, total pressure loss coefficient, vortex intensity scale index, and flow field separation index are calculated based on the built-in code.

[0074] 4. Aerodynamic optimization Based on reinforcement learning, and according to aerodynamic boundary conditions and target aerodynamic parameter requirements, the turbine aerodynamic configuration geometry parameters are optimized to maximize turbine efficiency and minimize total pressure loss coefficient, vortex intensity and scale, and flow separation. Iteration stops when the reward function remains unchanged for 10 consecutive iterations or exceeds the maximum number of iterations.

[0075] In summary, this implementation method trains a multi-scale, multi-path self-attention network based on a turbine three-dimensional flow field database to achieve rapid prediction of the turbine three-dimensional flow field. Based on the three-dimensional flow field characteristics and overall performance parameters, reinforcement learning is used to intelligently optimize the turbine blade geometry parameters and the control parameters of the three-dimensional stacking scheme.

[0076] Those skilled in the art will understand that the above description is merely a preferred embodiment of the present invention, and the features described in the various embodiments and / or claims of this disclosure can be combined or combined in various ways, even if such combinations or combinations are not explicitly described in this disclosure. This is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0077] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention. Clearly, those skilled in the art can make various alterations and modifications to the invention without departing from its spirit and scope. Thus, if these modifications and modifications of the invention fall within the scope of the claims and their equivalents, the invention is also intended to include these modifications and modifications.

Claims

1. A method for intelligent optimization of turbine aerodynamic shape based on rapid prediction of three-dimensional full flow field, characterized in that, The method includes the following steps: Step 1: Construct a turbine three-dimensional flow field dataset using radial revolution surfaces as unit samples. The turbine three-dimensional flow field dataset includes a training set, a dataset, and a validation set. Step 2: Construct a neural network model of the three-dimensional full flow field of the turbine channel based on the turbine three-dimensional flow field dataset built in Step 1; Step 3: Based on the neural network model of the three-dimensional full flow field of the turbine channel constructed in Step 2, construct an intelligent optimization framework integrating prediction, evaluation, and optimization. The blade geometry is described in a parameterized manner, and the key geometric parameters are used as state variables of the reinforcement learning environment. The aerodynamic performance index output by the fast flow field prediction model under given geometric and boundary conditions is used as the reward signal. The reinforcement learning agent autonomously adjusts the blade parameters in the continuous design space and continuously updates the policy network to achieve adaptive optimization of multi-objective performance. The fast prediction model is used to solve the three-dimensional flow field of the turbine under specified operating conditions and to deduce the overall performance parameters accordingly. The integrated intelligent optimization framework for prediction, evaluation, and optimization described in step 3 includes the overall performance term R. perf and flow field organization item R flow ; in and The weighting of the two reward items is assigned respectively. and These are the efficiency and total pressure loss coefficients, respectively. The corresponding weights; and These are the scale measures of the large eddy and the scale measures of the separation, respectively. This represents the weight of the flow field organization term.

2. The intelligent optimization method for turbine aerodynamic shape based on rapid prediction of three-dimensional full flow field as described in claim 1, characterized in that, The method for constructing the neural network model of the three-dimensional full flow field of the turbine channel in step 2 is as follows: Based on supervised learning methods, establish a radial section airfoil... p and import / export border conditions Departure to complete blade surface field q Forward flow field prediction model The positive flow field prediction model The input data includes radial section airfoil, inlet angle of attack (AOA), and outlet isentropic Mach number (Ma2is) provided in image form, with corresponding relative radial positions.

3. The intelligent optimization method for turbine aerodynamic shape based on rapid prediction of three-dimensional full flow field as described in claim 1, characterized in that, The Adam algorithm was used to train a neural network model of the three-dimensional full flow field in the turbine channel.

4. The intelligent optimization method for turbine aerodynamic shape based on rapid prediction of three-dimensional full flow field as described in claim 3, characterized in that, A composite loss function is used to adjust the predicted and actual values ​​of the neural network model for the three-dimensional full flow field of the turbine channel. This composite loss function includes the mean squared error and the pixel gradient error, with the weighting coefficient λ set to 10. Where H and W represent the number of sampling points along the axial and circumferential directions for each radial section, respectively, and the subscripts i and j are the indices of the two directions.

5. The intelligent optimization method for turbine aerodynamic shape based on rapid prediction of three-dimensional full flow field according to claim 1, characterized in that, The normalized absolute error QE is used to predict the training set and the dataset, i.e. 。 6. The intelligent optimization method for turbine aerodynamic shape based on rapid prediction of three-dimensional full flow field as described in claim 1, characterized in that, The ratio of the training set, validation set, and test set is 4:1:1.

25.

7. A turbine aerodynamic shape intelligent optimization system based on rapid prediction of three-dimensional full flow field, characterized in that, The system is implemented based on the method of claim 1, and the system includes: The data acquisition module is used to construct a turbine three-dimensional flow field dataset using radial rotating surfaces as unit samples. The turbine three-dimensional flow field dataset includes a training set, a dataset, and a validation set. The model building module is used to construct a neural network model of the three-dimensional full flow field of the turbine channel based on the turbine three-dimensional flow field dataset constructed by the data acquisition module. The optimization module is used to build an intelligent optimization framework integrating prediction, evaluation, and optimization based on the neural network model of the three-dimensional full flow field of the turbine channel constructed by the model building module. It describes the blade geometry in a parameterized way, uses key geometric parameters as state variables of the reinforcement learning environment, and uses the aerodynamic performance index output by the fast flow field prediction model under given geometric and boundary conditions as reward signals. The reinforcement learning agent autonomously adjusts the blade parameters in the continuous design space and continuously updates the policy network to achieve adaptive optimization of multi-objective performance. The fast prediction model is used to solve the three-dimensional flow field of the turbine under specified operating conditions and to deduce the overall performance parameters accordingly. The intelligent optimization framework integrating prediction, evaluation, and optimization described in the optimization module includes the overall performance term R. perf and flow field organization item R flow ; in and The weighting of the two reward items is assigned respectively. and These are the efficiency and total pressure loss coefficients, respectively. The corresponding weights; and These are the scale measures of the large eddy and the scale measures of the separation, respectively. This represents the weight of the flow field organization term.

8. A computer storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method described in any one of claims 1-6.

9. A computer device, characterized in that, include: A memory, a processor, and a computer program stored in the memory and executable on the processor, the processor executing the program to implement the method of any one of claims 1-6.

Images