Search

CN-115114868-B - Turbine loss model construction method based on deep learning and loss weight analysis

CN115114868BCN 115114868 BCN115114868 BCN 115114868BCN-115114868-B

Abstract

The invention discloses a turbine blade loss model construction method based on deep learning and loss weight analysis, which aims to explore how to correct an existing turbine blade loss model, and is characterized in that each loss in the blade loss is split and compared with the prediction size of the existing model, coefficients and terms needing to be corrected and correction terms needing to be added are obtained through analysis, and then a loss prediction model with a correction form is formed. And by comparing turbine profile losses with different profile parameters, the profile parameter variables to be considered are found. And establishing a functional relation between the leaf pattern parameter variable to be considered and the coefficient (or item and correction item to be added) to be corrected by using the artificial neural network model, and bringing the functional relation into a loss prediction model with a correction form, thereby constructing a turbine leaf pattern loss prediction model. The method can accurately predict the turbine blade profile loss with a larger attack angle working range.

Inventors

  • WANG SONGTAO
  • LI HEQUN
  • FANG XINGLONG
  • OUYANG YUQING
  • Fang Kanxian
  • CHEN YINGJIE

Assignees

  • 哈尔滨工业大学

Dates

Publication Date
20260512
Application Date
20220713

Claims (10)

  1. 1. The turbine loss model construction method based on deep learning and loss weight analysis is characterized by comprising the following steps of: s1, obtaining geometric parameters of a two-dimensional blade profile of a turbine, and carrying out a planar blade grid test on the two-dimensional blade profile of the turbine according to a selected working condition to obtain aerodynamic parameters of an inlet and an outlet of the blade profile of the turbine; step S2, performing CFD calculation on the two-dimensional blade profile of the turbine, and checking CFD calculation results by comparing the CFD calculation with profile pressure distribution, outlet total pressure loss circumferential distribution and outlet total pressure loss obtained by test measurement in the step S1; step S3, extracting the loss size of each component in the leaf type loss and the additional loss size generated by the existence of attack angles of incoming flows by combining the CFD calculation result obtained in the step S2 and the test measurement result obtained in the step S1; S4, splitting each loss component according to the existing zero-attack-angle turbine blade profile loss prediction empirical model, comparing the split loss component with the corresponding partial loss size extracted in the step S3, and analyzing the coefficient to be corrected and the correction item to be added in the zero-attack-angle turbine blade profile loss prediction empirical model; S5, constructing a turbine blade profile loss prediction empirical model and a functional form of a coefficient to be corrected and a correction term to be added in the empirical model capable of predicting additional loss caused by attack angle by adopting a neural network model; And S6, constructing a sample set by using the blade profile parameters obtained in the step S1 and the corresponding blade profile loss obtained in the step S3, and further training a turbine blade profile loss prediction empirical model and an empirical model capable of predicting parasitic loss caused by attack angles by using the sample set to obtain a final result.
  2. 2. The turbine loss model construction method based on deep learning and loss weight analysis according to claim 1, wherein in the step S1, the geometric parameters include an inlet geometric angle β 1 , an outlet geometric angle β 2 , an installation angle γ, a leading edge wedge angle We, a leading edge relative thickness LEd/C, a trailing edge relative thickness TEd/C, a relative maximum thickness tmax/C, a relative pitch t/C, and a leading edge ovality LEb/LEa, the inlet geometric angle β 1 , the outlet geometric angle β 2 , and the installation angle γ are included angles with an axial direction, LEa is an elliptical diameter perpendicular to an incoming flow direction in a leading edge ellipse flowing down with a zero attack angle, LEb is an elliptical diameter identical to the incoming flow direction in the leading edge ellipse flowing down with a zero attack angle, LEd is a blade-shaped leading edge circle diameter, C is a blade-shaped chord length, TEd is a blade-shaped trailing edge circle diameter, tmax is a blade-shaped maximum thickness, and t is a pitch.
  3. 3. The turbine loss model construction method based on deep learning and loss weight analysis according to claim 1, wherein in the step S1, the import-export aerodynamic parameters are import-export aerodynamic parameters under different working conditions, and at least include an import total pressure Inlet static pressure P 1 , outlet static pressure P 2 , inlet mach number Ma 1 , outlet mach number Ma 2 , incoming flow angle of attack i, outlet air flow angle a 2 , total pressure loss Y p of the outlet cross section, profile pressure distribution of the blade profile, and total pressure loss circumferential distribution of the outlet.
  4. 4. The turbine loss model construction method based on deep learning and loss weight analysis according to claim 1, wherein in the step S3, the calculation method of the loss magnitude of each component in the leaf pattern loss is as follows: 1) For the total pressure loss coefficient of the boundary layer The specific calculation mode is as follows: Calculating the energy loss in the boundary layer through the boundary layer energy thickness: ; Wherein, the For the magnitude of the fluid velocity in the boundary layer, For the main flow rate magnitude, To the extent that the fluid density in the boundary layer is high, Is the nominal thickness of the boundary layer; Definition of energy thickness according to boundary layer : ; Wherein, the The energy loss in the boundary layer is as follows: ; according to the ideal kinetic energy of the outlet calculated under the condition that the assumed flow is isentropic flow: ; Wherein, the In order to achieve a mass flow rate, As the total enthalpy of the inlet, In order to be the static enthalpy of the outlet, The specific heat is fixed for the gas pressure, For the total temperature of the inlet, the total temperature of the inlet is equal to the total temperature of the inlet, Is the specific heat ratio of the glass fiber reinforced plastic material, The isentropic Mach number of the outlet; and finally, calculating the energy loss coefficient of the boundary layer as follows: ; and is converted to an additional layer total pressure loss coefficient Y bl by: ; 2) For wake loss By total loss of pressure in the outlet section Subtracting the loss of the boundary layer The preparation method comprises the following steps: ; the calculation method of the parasitic loss delta phi 2 P generated by the existence of the attack angle of the incoming flow is that the energy loss coefficient under the non-zero attack angle is different from the energy loss coefficient under the zero attack angle.
  5. 5. The turbine loss model construction method based on deep learning and loss weight analysis according to claim 1, wherein in the step S4, the existing model of zero attack turbine blade loss prediction experience is AMDCKO model of turbine loss prediction, and the coefficients to be corrected and the correction terms to be added include: 1) Replacing K 2 =(Ma 1 /Ma 2 ) 2 with K 2 =(Ma 1 /Ma 2 ) a by the power of the Mach number correction factor K 2 expression of the leaf type boundary layer loss, wherein a is a term to be corrected, ma 1 is an inlet Mach number, and Ma 2 is an outlet Mach number; 2) The power in the Mach number correction factor K 1 expression of the leaf type boundary layer loss is replaced by K 1 =1-(1.25(Ma 2 -0.2)) b by K 1 =1-(1.25(Ma 2 -0.2)) 1 , wherein b is a term to be corrected; 3) The weight occupied by boundary layer loss is replaced by a coefficient W P ,W P which is a coefficient 2/3 related to the boundary layer loss in the original size and is a undetermined coefficient; 4) The predicted value of wake loss Y ' TET multiplied by the compressible correction term K tet is expressed as: ; wherein c, d, e, f are both items to be determined.
  6. 6. The turbine loss model construction method based on deep learning and loss weight analysis according to claim 5, wherein in the step S4, the empirical model capable of predicting the parasitic loss due to the attack angle is a Benner turbine loss prediction model, and the coefficients to be corrected and the correction terms to be added include: 1) Angle of attack variable Is represented by the expression: becomes as follows ; Wherein LEd is the leading edge circle diameter, we is the leading edge wedge angle, β 1 is the inlet geometry, β 2 is the outlet geometry, l, m, n are coefficients to be determined, A, B is the term to be determined; 2) Will utilize angle of attack variables Calculating the piecewise functions with parasitic losses due to angle of attack to combine them into one function 。
  7. 7. The turbine loss model construction method based on deep learning and loss weight analysis according to claim 6, wherein in the step S5, when constructing the functional form of the coefficient to be corrected and the correction term to be added in the turbine blade profile loss prediction empirical model, a, b, C, d, e, f is considered to be the functions of the blade profile relative thickness tmax/C, the ratio γ/(β 1 +α 2 ) of the blade profile mounting angle to the airflow camber angle under the zero attack angle, and the relative pitch t/C; The 6 functions of F 1 、F 2 、G 1 、G 2 、H 1 、H 2 are constructed by adopting an artificial neural network model, the adopted artificial neural network model is provided with 3 hidden layers, each hidden layer is provided with 400 neurons, and the activation function adopts a model of PReLu functions.
  8. 8. The turbine loss model construction method based on deep learning and loss weight analysis according to claim 7, wherein in the step S5, when constructing a functional form of a coefficient to be corrected and a correction term to be added in an empirical model capable of predicting parasitic loss due to attack angle, A, B is considered to be, In the form of the following function: a. for item A to be determined, A is considered to be a function of the ratio of airfoil mounting angle to geometric camber angle γ/(β 1 +β 2 ), airfoil relative thickness tmax/C, relative pitch t/C, leading edge ovality LEb/LEa, namely: ; constructing a function f by using an artificial neural network model, wherein the artificial neural network model is provided with 8 hidden layers, each hidden layer is provided with 2400 neurons, and an activation function is selected as PReLu functions; b. For item B to be determined, B is considered to be a function of the ratio of airfoil stagger angle to geometric camber angle γ/(β 1 +β 2 ), airfoil relative thickness tmax/C, relative pitch t/C, leading edge ovality LEb/LEa, incoming flow angle of attack i, leading edge diameter to pitch ratio LEd/t, leading edge wedge angle We, airfoil convergence ratio cos (β 1 )/cos(β 2 ), namely: ; When the front edge small circle is elliptical, the front edge diameter LEd has the same value as LEa; Constructing a function g by using an artificial neural network model, wherein the artificial neural network model is provided with 8 hidden layers, each hidden layer is provided with 2000 neurons, and an activation function is PReLu functions; c. For the following Thought to be By variation of angle of attack The 1 st to 8 th powers of (i) are functions of variables, namely: ; Constructing functions using artificial neural network model The adopted artificial neural network model is provided with 7 hidden layers, each hidden layer is provided with 3000 neurons, and an activation function is PReLu functions; Determining the sizes of l, m and n and the functions f and g, In the specific form of (2), the specific method is as follows: L, m, n, f, g, phi are brought into the predictive formula of parasitic loss due to attack angle in modified form, the formula is as follows: ; training the model by taking the geometrical parameters of the blade profile and the aerodynamic parameters of the inlet and outlet as inputs and taking the additional loss generated due to attack angle as the sample set of the output, and finally determining f, g, Specific parameters of (a) are defined.
  9. 9. The turbine loss model construction method based on deep learning and loss weight analysis according to claim 8, wherein in the step S6, the specific method for determining the function form of the 6 functions of F 1 、F 2 、G 1 、G 2 、H 1 、H 2 and the specific size of the coefficient W P to be normalized is as follows: 1) For F 1 、F 2 、W P , it is put into the boundary layer loss prediction equation in modified form, which is shown as follows: ; Wherein Y P,AMDC is the blade profile loss size calculated by utilizing a AMDC turbine loss model, and Re is the Reynolds number; Taking the blade geometry parameters and the inlet and outlet aerodynamic parameters as input and taking the boundary layer loss Y bl as output sample set, randomly selecting 8/9 sample capacity to train the model, and finally determining the specific parameters of F 1 、F 2 and the specific size of W P ; 2) For G 1 、G 2 、H 1 、H 2 , it is put into the wake loss prediction formula in modified form, which is shown below: ; wherein Y TET,AMDCKO is the size of wake loss calculated by using a AMDCKO turbine loss model; the blade geometry parameters and the import and export aerodynamic parameters are taken as input, the wake loss Y m is taken as output sample set, and 8/9 of the sample capacity is randomly selected for training the model, so that the function G 1 、G 2 、H 1 、H 2 is finally determined.
  10. 10. The turbine loss model construction method based on deep learning and loss weight analysis according to claim 1, wherein in the step S6, when training the zero attack angle turbine blade type loss prediction empirical model and the empirical model capable of predicting parasitic loss due to attack angle, a sample set is randomly divided into a training set and a test set according to the ratio of 8:1, the model is trained, the generalization ability of the training result is checked, and HuberLoss loss function evaluation model prediction accuracy is selected.

Description

Turbine loss model construction method based on deep learning and loss weight analysis Technical Field The invention relates to a turbine blade aerodynamic loss prediction model construction method in an aeroengine/gas turbine, in particular to a turbine blade loss model construction method based on deep learning and loss weight analysis. Background The power system of the aero-engine is developed to the fourth generation engine at present and gradually moves to the fifth generation engine, a wider flight envelope is required, and the high-speed rotorcraft also has wider working range requirements on the aero-engine. The turbine is used as an important rotating component for converting heat energy of working media into machinery, and the loss performance requirement in the working range is correspondingly improved, so that new requirements are put on pneumatic design work of the turbine component. The pneumatic design work of the gas turbine is a process of gradually designing and optimizing from low dimension to high dimension, and the low dimension design result is used as the initial value and the basis of the high dimension design. In one-dimensional design, the loss model plays a very large role, and good inlet and outlet aerodynamic conditions can be selected according to the loss model, so that the loss model has important significance for the design of the whole turbine component. However, with the development of technology, the performance of turbine blades is continuously improved, and the prediction accuracy of the existing turbine loss prediction model is poor, especially the prediction performance of blade profile loss for a wide attack angle range is poor. The prediction relative error is about 30 percent (AMDCKO turbine loss model), and the non-zero attack angle turbine loss prediction model has poorer prediction precision in the negative attack angle range. Therefore, correction is required for the existing loss model. Disclosure of Invention The invention aims to provide a turbine loss model construction method based on deep learning and loss weight analysis, which improves the prediction precision of an original model and simultaneously provides a simple and rapid turbine loss model construction method. The invention aims at realizing the following technical scheme: a turbine loss model construction method based on deep learning and loss weight analysis comprises the following steps: S1, obtaining geometric parameters of a two-dimensional blade profile of a turbine, and carrying out a planar blade profile test on the two-dimensional blade profile of the turbine according to a selected working condition to obtain aerodynamic parameters of an inlet and an outlet of the blade profile of the turbine, wherein: The geometric parameters include an inlet geometric angle beta 1, an outlet geometric angle beta 2, a mounting angle gamma, a leading edge wedge angle We, a leading edge relative thickness LEd/C, a trailing edge relative thickness TEd/C, a relative maximum thickness tmax/C, a relative pitch t/C, and a leading edge ovality LEb/LEa; The inlet geometric angle beta 1, the outlet geometric angle beta 2 and the installation angle gamma are included angles with the axial direction, LEa is an elliptical diameter perpendicular to the incoming flow direction in the leading edge ellipse flowing down with zero attack angle, LEb is an elliptical diameter identical to the incoming flow direction in the leading edge ellipse flowing down with zero attack angle, LEd is a diameter of a blade-shaped leading edge circle (the leading edge circle is a perfect circle), C is a blade-shaped chord length, TEd is a diameter of a blade-shaped trailing edge circle (the trailing edge circle is a perfect circle), tmax is a blade-shaped maximum thickness, and t is a grid distance; The inlet and outlet aerodynamic parameters are inlet and outlet aerodynamic parameters under different working conditions and at least comprise inlet total pressure P *1, inlet static pressure P 1, outlet static pressure P 2, inlet Mach number Ma 1, outlet Mach number Ma 2, incoming flow attack angle i, outlet airflow angle alpha 2, total pressure loss Y p of an outlet section, profile pressure distribution of a blade type and outlet total pressure loss circumferential distribution. Step S2, performing CFD calculation on the two-dimensional blade profile of the turbine, and checking CFD calculation results by comparing the CFD calculation with profile pressure distribution, outlet total pressure loss circumferential distribution and outlet total pressure loss obtained by test measurement in the step S1; step S3, combining the CFD calculation result obtained in the step S2 and the test measurement result obtained in the step S1, and extracting the loss size of each component in the leaf type loss and the additional loss size generated by the existence of attack angle of incoming flow, wherein: Methods for calculating the loss size of each comp