Turbine blade anti-sinking cooling collaborative optimization method based on proxy model and genetic algorithm

Abstract

The invention discloses a turbine blade anti-sinking cooling collaborative optimization method based on a proxy model and a genetic algorithm, and the method comprises the steps: firstly constructing an initial sample library through Latin hypercube sampling, building an initial neural network proxy model of a deposition rate, a heat exchange coefficient and a flow resistance coefficient, and then driving a multi-target genetic algorithm to carry out global optimization; in the anti-sinking cooling collaborative optimization process, a Pareto non-dominated solution generated in the middle and later periods of genetic algorithm iteration is selectively fed back to a training sample so as to enhance the prediction precision of an agent model in a key area, and a new round of iterative optimization is triggered. And continuously improving the quality of the solutions and the reliability of the model through the closed-loop iteration mechanism, and finally reordering based on comprehensive evaluation of all Pareto solution sets to determine an optimal anti-sinking cooling structure scheme. According to the invention, the efficient and high-precision automatic design of the anti-sinking cooling structure of the blade is realized.

Description

Technical Field

[0001] This invention belongs to the field of design optimization technology for key components of aero-engines, specifically involving a collaborative optimization method for turbine blade anti-sinking cooling based on surrogate model and genetic algorithm. Background Technology

[0002] When aero-engines and ground-based gas turbines operate in environments with severe particulate pollution, particulate matter enters the external flow channels and internal cooling channels of turbine blades along with the main combustion gas and cool air, causing deposits on both the inner and outer surfaces of the blades. These deposits not only clog the combustion gas channels and reduce turbine efficiency but also block cooling film vents and impede internal cooling airflow, significantly increasing the risk of blade overheating and failure. This severely impacts the economy and safety of the engine and gas turbine. Therefore, optimizing the dust-proof and sediment-reducing cooling structure inside the blades to achieve a balance between efficient cooling and deposit control has become a key challenge for improving the performance and reliability of turbomachinery.

[0003] Currently, optimization design for this type of problem often employs simulation-driven methods. However, high-fidelity numerical simulations (such as computational fluid dynamics (CFD)) are computationally expensive and time-consuming, making them difficult to directly apply to optimization processes requiring numerous iterations. Therefore, the industry often uses surrogate model techniques, which involve constructing a computationally cheaper approximate model using a limited number of simulation samples to replace actual simulations for rapid evaluation. This model is then combined with global optimization methods such as genetic algorithms to find the optimal design solution.

[0004] However, the traditional "single-step modeling, single-step optimization" strategy has significant shortcomings when facing the problem of optimizing the layout of dust-proof and sediment-reducing cooling structures. The deposition distribution of particulate matter within the cooling structure exhibits typical characteristics of "point discretization" and "abrupt gradient at the edges," meaning that the deposition amount does not change gradually in space, but rather undergoes a sharp jump from high to zero values ​​in local areas. This characteristic leads to the following problems: 1. The surrogate model has insufficient prediction accuracy in key areas: The initial surrogate model is built based on global sparse samples. Although it can reflect the overall trend, it is difficult to achieve accurate fitting in the key optimization space. 2. Optimization results get stuck in suboptimal solutions: Optimization based on surrogate models with low accuracy in key regions can easily mislead the search direction of genetic algorithms, making it difficult for them to converge to a true, high-quality Pareto front. The results obtained are often not globally optimal and may even fail to meet engineering requirements.

[0005] While some studies have attempted to improve model accuracy by increasing the initial number of samples, this linearly increases the cost of early-stage simulations, and blind sampling cannot guarantee that the new samples will effectively improve the model's predictive ability in key regions. Therefore, how to effectively improve the fitting accuracy of the surrogate model in key regions within a controllable computational cost, and guide the optimization algorithm to obtain a truly effective Pareto optimal solution, has become a critical technical bottleneck that urgently needs to be overcome in this field. Summary of the Invention

[0006] To address the problems existing in the prior art, the present invention aims to provide a collaborative optimization method for turbine blade anti-sinking cooling based on a surrogate model and a genetic algorithm. This method has higher optimization result quality, higher computational efficiency, stronger adaptability to complex problems, and good engineering versatility.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] A collaborative optimization method for turbine blade anti-sinking cooling based on surrogate model and genetic algorithm includes the following steps:

[0009] S1: Sample points are collected in the geometric parameter design space of turbine blade cooling structure by Latin hypercube sampling to construct an initial sample library, and initial neural network surrogate models of deposition amount, heat transfer coefficient and flow resistance coefficient in turbine blade cooling structure are trained based on the initial sample library.

[0010] S2: Using the initial neural network surrogate model as a fitness evaluator, drive the multi-objective genetic algorithm to perform global optimization and obtain a set of Pareto non-dominated solutions;

[0011] S3: In the middle and late stages of the multi-objective genetic algorithm iteration, selectively extract some Pareto non-dominated solutions from the generated solutions;

[0012] S4: Add the extracted Pareto nondominated solutions to the initial sample library and update the sample library;

[0013] S5: Retrain the initial neural network proxy model based on the updated sample library;

[0014] S6: Repeat steps S2 to S5 to perform iterative optimization until the preset termination condition is met.

[0015] S7: Perform a final evaluation and sorting of all Pareto nondominated solutions accumulated during the iteration process to determine the globally optimal solution for the turbine blade anti-sinking cooling structure.

[0016] Preferably, the number of sample points in S1 is 8 to 12 times the number of geometric variables to be optimized. This effectively controls the early simulation cost while ensuring that the initial surrogate model has basic global accuracy, and avoids model distortion due to too few samples or resource waste due to too many samples.

[0017] Preferably, the initial neural network proxy model in S1 is one or a combination of the following: artificial neural network, convolutional neural network, generative adversarial neural network. These models exhibit good fitting accuracy and numerical stability when dealing with highly nonlinear and small sample problems, and can give full play to the advantages of deep learning in multi-physics field prediction, providing a reliable prediction basis for subsequent optimization.

[0018] Preferably, the mid-to-late stage of the multi-objective genetic algorithm iteration described in S3 is defined as the 20th to 30th generation. At this stage, the population has fully evolved and tends to converge, and the Pareto solution set produced is of high quality and representative. Selecting points from it to supplement samples can most effectively guide the surrogate model to refine towards the optimal region.

[0019] Preferably, when selectively extracting Pareto non-dominated solutions in S3, solutions that are sparsely distributed in the target space are preferred. This helps to enhance the predictive ability of the surrogate model in regions that have not been fully explored, promotes the complete mining of the Pareto front, and prevents the solution set from falling into local convergence too early.

[0020] Preferably, the strategy for selectively extracting Pareto non-dominated solutions in S3 includes: prioritizing the selection of solutions located at the edge of the Pareto front; secondly, selecting solutions with convex or concave features on the front surface; and finally selecting solutions in blank areas to complete the selection plane. This strategy can systematically ensure that the selected solution set is highly representative in both shape and distribution, thereby efficiently driving the surrogate model to learn the entire front comprehensively and accurately.

[0021] Preferably, when retraining the surrogate model in S5, the surrogate models corresponding to variables whose prediction errors exceed a set threshold are selectively updated; the variables with the set threshold are deposition distribution, deposition amount, heat transfer coefficient, or flow resistance coefficient. This selective update mechanism focuses on prediction difficulties (such as deposition distribution), avoids indiscriminate reconstruction of all surrogate models, and significantly improves the computational efficiency of the model update process.

[0022] Preferably, the number of iterations in S6 is no less than 10 to ensure that the "optimization-feedback" mechanism has enough rounds to improve the accuracy of the surrogate model and the correct search of the genetic algorithm, thereby ensuring that the final solution set converges to a stable and high-quality state.

[0023] Preferably, the final evaluation and ranking in S7 adopts non-dominated sorting and is combined with crowding degree calculation. The non-dominated sorting and crowding degree calculation performed again can filter out the optimal solution layer in all iterations, ensuring the coverage of the final solution set and preventing excellent solutions from being missed.

[0024] This invention addresses the difficulty of solving the Pareto front caused by discrete deposition distribution and abrupt edge gradient changes in the cooling structure layout optimization for aero-engines operating in dusty environments, which requires consideration of heat transfer and cooling, flow loss, and dust prevention and reduction. Compared with existing technologies, the technical solution employed in this invention offers the following advantages:

[0025] 1. Higher quality of optimized results: By constructing a closed-loop mechanism of "optimization-evaluation-learning-re-optimization", the prediction accuracy of the surrogate model in key areas is dynamically improved, making the optimization results closer to the real Pareto frontier, effectively solving the problem of low solution quality caused by the inaccuracy of the model in key areas in traditional methods.

[0026] 2. Higher computational efficiency: By triggering model updates only at critical stages of genetic algorithm convergence (such as the mid-to-late stages), and combining a sparse selection strategy for Pareto non-dominated solutions, the optimization effect is maximized with minimal model reconstruction cost, balancing computational cost and solution accuracy, and avoiding resource waste.

[0027] 3. Enhanced Adaptability to Complex Problems: Traditional surrogate models struggle to balance global trends and local accuracy due to the strong nonlinearity, high gradients, and discreteness of sediment distribution in space, and the complex mapping relationships between sediment distribution and geometric parameters. This invention introduces an adaptive learning mechanism, enabling the surrogate model to self-update in key regions of the cooling structure's geometric parameter design space and gradually approach the true Pareto front, including sediment prediction, thereby significantly improving its ability to solve such complex optimization problems.

[0028] 4. Good engineering versatility: The method has a clear framework, is based on the integration and innovation of mature algorithms, is easy to implement and port to multi-objective optimization design in other engineering fields, and has strong practicality. Attached Figure Description

[0029] Figure 1 This is a flowchart of the method of the present invention.

[0030] Figure 2 The image shows a comparison of the deposition distribution in the leading edge vortex cavity of the blade before and after the application of this invention (the left side shows the prediction results of the generative adversarial neural network, and the right side shows the CFD calculation results). Detailed Implementation

[0031] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0032] like Figure 1 As shown, the execution flow of the optimization method of this invention is a closed-loop process including initialization and iterative iteration: First, an initial sample library is constructed and an initial neural network surrogate model is trained through Latin hypercube sampling; then, the surrogate model is used to drive the genetic algorithm for optimization; when the optimization enters the middle and late stages, the current Pareto non-dominated solutions are selectively extracted to supplement the sample library, and the surrogate model is updated accordingly; this "optimization-feedback-model update" cycle is repeated until the termination condition is met; finally, all Pareto solutions obtained in each optimization are uniformly ranked and evaluated, and the globally optimal solution is output. Specifically, the following steps are included:

[0033] S1: Sample points are collected in the geometric parameter design space of turbine blade cooling structure by Latin hypercube sampling to construct an initial sample library, and initial neural network surrogate models of deposition amount, heat transfer coefficient and flow resistance coefficient in turbine blade cooling structure are trained based on the initial sample library.

[0034] S2: Using the initial neural network surrogate model as a fitness evaluator, drive the multi-objective genetic algorithm to perform global optimization and obtain a set of Pareto non-dominated solutions;

[0035] S3: In the middle and late stages of the multi-objective genetic algorithm iteration, selectively extract some Pareto non-dominated solutions from the generated solutions;

[0036] S4: Add the extracted Pareto nondominated solutions to the initial sample library and update the sample library;

[0037] S5: Retrain the initial neural network proxy model based on the updated sample library;

[0038] S6: Repeat steps S2 to S5 to perform iterative optimization until the preset termination condition is met.

[0039] S7: Perform a final evaluation and sorting of all Pareto nondominated solutions accumulated during the iteration process to determine the globally optimal solution for the turbine blade anti-sinking cooling structure.

[0040] Figure 2This paper compares the prediction results of deposition distribution within the leading-edge vortex cavity of turbine blades before and after applying the method of this invention. Based on the high-precision CFD simulation results on the right, it can be seen that without the application of this invention, the surrogate model exhibits severe distortion in its prediction of key depositional features: for example, strip-shaped deposits in the Pareto1 scheme are incorrectly predicted as diffuse deposit blocks, while in the Pareto3 and Pareto4 schemes, the model almost completely loses its ability to predict depositional distribution. After 12 iterations of optimization using this invention, the predictive ability of the surrogate model is significantly improved, achieving basically accurate predictions for different morphologies of depositional features (such as strips, dots, and blocks), and even effectively identifying depositional blank areas in Pareto2, Pareto3, and Pareto5.

Claims

1. A collaborative optimization method for turbine blade anti-sinking cooling based on surrogate model and genetic algorithm, characterized in that, Includes the following steps: S1: Sample points are collected in the geometric parameter design space of turbine blade cooling structure by Latin hypercube sampling to construct an initial sample library, and initial neural network surrogate models of deposition amount, heat transfer coefficient and flow resistance coefficient in turbine blade cooling structure are trained based on the initial sample library. S2: Using the initial neural network surrogate model as a fitness evaluator, drive the multi-objective genetic algorithm to perform global optimization and obtain a set of Pareto non-dominated solutions; S3: In the middle and late stages of the multi-objective genetic algorithm iteration, selectively extract some Pareto non-dominated solutions from the generated solutions; S4: Add the extracted Pareto nondominated solutions to the initial sample library and update the sample library; S5: Retrain the initial neural network proxy model based on the updated sample library; S6: Repeat steps S2 to S5 to perform iterative optimization until the preset termination condition is met. S7: Perform a final evaluation and sorting of all Pareto nondominated solutions accumulated during the iteration process to determine the globally optimal solution for the turbine blade anti-sinking cooling structure.

2. The method according to claim 1, characterized in that, The number of sample points mentioned in S1 is 8 to 12 times the number of geometric variables to be optimized.

3. The method according to claim 1, characterized in that, The initial neural network proxy model described in S1 is one of the following or a combination thereof: artificial neural network, convolutional neural network, generative adversarial neural network.

4. The method according to claim 1, characterized in that, The mid-to-late stage of the multi-objective genetic algorithm iteration described in S3 is defined as generation 20 to generation 30.

5. The method according to claim 1, characterized in that, When selectively extracting Pareto nondominated solutions in S3, solutions that are sparsely distributed in the target space are preferred.

6. The method according to claim 5, characterized in that, The strategy for selectively extracting Pareto non-dominated solutions in S3 includes: prioritizing solutions located at the edge of the Pareto front; secondly, selecting solutions with convex or concave features on the front surface; and finally, selecting solutions in blank areas to complete the selected plane.

7. The method according to claim 1, characterized in that, When retraining the surrogate model in S5, the surrogate model corresponding to the variable whose prediction error exceeds a set threshold is selectively updated; the variable with the set threshold is the deposition distribution, deposition amount, heat transfer coefficient or flow resistance coefficient.

8. The method according to claim 1, characterized in that, The iterative optimization described in S6 shall be performed no less than 10 times.

9. The method according to claim 1, characterized in that, The final evaluation and ranking described in S7 adopts a non-dominated ranking and is combined with crowding calculation.

Images