CN-116467925-B - Turbine cooling blade design space dimension reduction method

CN116467925BCN 116467925 BCN116467925 BCN 116467925BCN-116467925-B

Abstract

The invention discloses a turbine cooling blade design space dimension reduction method, which comprises the steps of describing the geometric shape of a turbine cooling blade through control points of N 1 Bezier curves, obtaining a database with q blade shapes through Latin hypercube sampling of N 1 control points, obtaining a group of orthogonal base modes through eigenvoice dimension reduction to describe the geometric change of the blade shapes, modeling the main base modes according to an initial sample in agent model-based optimization, modeling each design target through an integrated agent model, performing variance analysis on each design target, removing design variables irrelevant to the targets, modeling each design target through the integrated agent model based on the existing design variables after removing the corresponding design variables, and considering that the variables are not improved or reduced if correlation coefficients are not improved, removing the variables, and taking the variables obtained through statistics as final design variables.

Inventors

HU KAIBIN
LI ZHEN
Fang Xuantao
JU YAPING
ZHANG CHUHUA

Assignees

西安交通大学

Dates

Publication Date: 20260512
Application Date: 20220803
Priority Date: 20220715

Claims (3)

1. A method for reducing dimension of a turbine cooling blade design space is characterized by comprising the following steps, Step one, determining all design variables of the turbine cooling blade, including cooling system variables and blade profile related variables of the turbine blade, wherein, The cooling system variables of the turbine blades are determined by the cooling unit type to obtain N 1 variable control cooling systems; For the blade profile related variables, determining by a method capable of describing the blade geometry to obtain N 2 variables to control the change of the blade geometry, and dispersing the blade profile into N 3 points; Thus, the number of design variables, all of which are denoted as D 1 ,D 1 , is denoted as N 4 , where N 4 ＝N 1 +N 2 ; Step two, sampling N 2 variables according to a first sampling method to obtain q different types of blade shapes, and establishing a matrix containing q blade shapes, wherein, Each row of the matrix represents N 3 discrete points of each leaf profile, and different rows of the matrix represent leaf profiles represented by different samples; The first sampling method refers to a sampling method capable of generating uniformly distributed samples in a high-dimensional design space; Step three, obtaining q orthogonal base modes to describe geometric variation of the leaf patterns through unsupervised dimension reduction for the matrix containing q leaf patterns; Selecting N 5 orthogonal base modes from q orthogonal base modes, and enabling generalized energy contained in the N 5 orthogonal base modes to be larger than or equal to all mode energy of a first percentage, wherein the number of design variables with the design variable being D 2 ,D 2 is recorded as N 6 , and N 6 ＝N 1 +N 5 ; Modeling each design target of the design variable D 2 by adopting a proxy model, and recording the model precision of each target, wherein the design target is a design performance index, and the model precision refers to the coincidence degree of the actual performance index value of a sample and the performance index predicted by the model; step six, performing variance analysis on each design target to obtain the main effect and the full effect of the design variable on each design target; Step seven, removing design variables irrelevant to the design objective from the design variables D 2 , wherein whether the design variables can be removed at a certain design objective is judged according to the following standard that the values of the main effect and the total effect in the analysis of variance are respectively smaller than a first quantity and a second quantity, The value range of the first quantity is 1% -3%, if the value is smaller than 1%, the variable is difficult to reduce, and if the value is larger than 3%, the excessive design variable is reduced, and the optimization effect is reduced; The value range of the second quantity is 2% -5%, and is larger than the first quantity, if the value range is smaller than 2%, the variable is difficult to reduce, and if the value range is larger than 5%, the excessive design variable is reduced, and the optimization effect is reduced; step eight, recording the remaining design variables as D 3 , modeling each design target again based on the remaining design variables D 3 , and recording the model precision of each design target; judging whether the precision of the model after the re-modeling is improved compared with the precision of the model in the step five, wherein the precision of the model reflects the precision index of model fitting through the root mean square error of the corresponding design target and the decision coefficient; 1) If the variable is improved, the variable removal is considered to be effective, and then the step five is executed again in a mode that the design variable D 3 of the step eight is assigned to the design variable D 2 of the step five, and the steps five to eight are executed again; 2) If the variable is not improved, the variable is not removed, the design variable D 2 when the step five is executed last time is output, and the dimension reduction is terminated.
2. The turbine cooling blade design space dimension reduction method of claim 1, wherein the unsupervised dimension reduction utilizes an intrinsic orthogonal decomposition method.
3. The turbine cooling blade design space dimension reduction method of claim 2, wherein the design objectives comprise five design objectives defined respectively as follows: an objective function 1, namely the cold air quantity; an objective function 2, total pressure recovery coefficient; the total pressure recovery coefficient C p1 is defined as follows: Wherein, the total pressure of the outlet of P out , the total pressure of the inlet of P in and the static pressure of the inlet of P in are shown; An objective function 3, namely the average temperature of the blade; objective function 4, maximum average temperature value of the blade, defined as follows: wherein: T high represents a temperature of 95% or more of the maximum blade temperature; v tem represents the blade volume corresponding to the region above 95% of the maximum blade temperature; V blade denotes the blade volume; the objective function 5, the maximum average temperature gradient value, is defined as follows: wherein: t-gra high represents a temperature gradient that exceeds 95% of the maximum blade temperature gradient; v gra represents the blade volume corresponding to a region exceeding 95% of the blade maximum temperature gradient; V blade denotes the blade volume.

Description

Turbine cooling blade design space dimension reduction method Technical Field The invention belongs to the technical field of turbine cooling blade design, and particularly relates to a dimension reduction method for a turbine cooling blade design space. Background Because of the complex structure of the cooling blades of modern gas turbine turbines, the number of design variables is a complex, high-dimensional, computationally intensive and black box (expensive and black box) design optimization problem typical of design optimization for turbines. The dimension reduction technology is adopted in the optimization design to reduce the dimension of the design space, and is an effective way for solving the HEB problem. At present, the dimension reduction technology facing the HEB optimization problem mainly comprises the following two types, namely a main mode of capturing a shape based on an unsupervised learning method and a design variable which has influence on a design target is screened out from a series of design variables based on a supervised learning method. The method is only suitable for the dimension reduction of two-dimensional blade design variables, and cannot reduce the dimension of the design variables of the turbine blade cooling system. The latter can overcome the limitation that the former is only suitable for leaf profile dimension reduction in theory, but has more defects in practical use. For example, local sensitivity analysis methods may choose the design variables most relevant to the design objective, but may not be applicable to highly nonlinear problems due to failure to take into account the correlation between design variables, while global sensitivity analysis methods may take into account the correlation between design variables, but may require sufficient numerical modeling to calculate the primary and interaction effects of each variable, resulting in high computational costs. The above information disclosed in the background section is only for enhancement of understanding of the background of the invention and therefore may contain information that does not form the prior art that is already known to a person of ordinary skill in the art. Disclosure of Invention Aiming at the problems existing in the prior art, the invention provides a turbine cooling blade design space dimension reduction method, which is realized by the following technical scheme, comprises the following steps, Step one, determining all design variables of the turbine cooling blade, including cooling system variables and blade profile related variables of the turbine blade, wherein, The cooling system variables of the turbine blades are determined by the cooling unit type to obtain N 1 variable control cooling systems; For the blade profile related variables, determining by a method capable of describing the blade geometry to obtain N 2 variables to control the change of the blade geometry, and dispersing the blade profile into N 3 points; Thus, the number of design variables, all of which are denoted as D 1,D1, is denoted as N 4, where N 4＝N1+N2; Step two, sampling N 2 variables according to a first sampling method to obtain q different types of blade shapes, and establishing a matrix containing q blade shapes, wherein, Each row of the matrix represents N 3 discrete points of each leaf profile, and different rows of the matrix represent leaf profiles represented by different samples; The first sampling method refers to a sampling method capable of generating uniformly distributed samples in a high-dimensional design space; Step three, obtaining q orthogonal base modes to describe geometric variation of the leaf patterns through unsupervised dimension reduction for the matrix containing q leaf patterns; Selecting N 5 orthogonal base modes from q orthogonal base modes, and enabling generalized energy contained in the N 5 orthogonal base modes to be larger than or equal to all mode energy of a first percentage, wherein the number of design variables with the design variable being D 2,D2 is recorded as N 6, and N 6＝N1+N5; Modeling each design target of the design variable D 2 by adopting a proxy model, and recording the model precision of each target, wherein the design target is a design performance index, and the model precision refers to the coincidence degree of the actual performance index value of a sample and the performance index predicted by the model; step six, performing variance analysis on each design target to obtain the main effect and the full effect of the design variable on each design target; Step seven, removing design variables irrelevant to the design objective from the design variables D 2, wherein whether the design variables can be removed at a certain design objective is judged according to the following standard that the values of the main effect and the total effect in the analysis of variance are respectively smaller than a first quantity and a second quantity, The value range o