Turbine loss model construction method based on deep learning and loss weight analysis
By combining deep learning and loss weight analysis with CFD calculations and artificial neural network models, the coefficients of the existing turbine blade loss prediction model are corrected and correction terms are added. This solves the problem of insufficient prediction accuracy of the existing model in a wide angle of attack range and achieves high-precision prediction of turbine blade loss.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- HARBIN INST OF TECH
- Filing Date
- 2022-07-13
- Publication Date
- 2026-05-12
AI Technical Summary
Existing turbine blade loss prediction models have poor prediction accuracy over a wide angle of attack range, especially for turbine loss prediction models with non-zero angle of attack, which have poor accuracy over a negative angle of attack range. The relative prediction error of existing models is around 30%.
A new turbine blade loss prediction model is constructed by using a deep learning-based method and loss weight analysis. This method obtains the geometric parameters and inlet/outlet aerodynamic parameters of the two-dimensional turbine blade profile, combines CFD calculations and experimental measurement results, decomposes the blade profile loss components, and uses an artificial neural network model to construct the relationship, correct the coefficients of the existing model and add correction terms.
It improves the accuracy of turbine blade loss prediction, with the relative error controlled within 5%, and can accurately predict turbine blade losses within a large angle of attack operating range.
Smart Images
Figure CN115114868B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a method for constructing an aerodynamic loss prediction model for turbine blades in aero-engines / gas turbines, specifically a method for constructing a turbine blade loss model based on deep learning and loss weight analysis. Background Technology
[0002] Aero engines, as a power system, have currently reached the fourth generation of engines and are gradually developing towards the fifth generation, requiring a wider flight envelope. High-speed rotorcraft also place wider operating range requirements on aero engines. As a crucial rotating component responsible for converting the thermal energy of the working fluid into mechanical energy, the turbine also faces increased performance requirements within its operating range, thus posing new challenges to the aerodynamic design of turbine components.
[0003] The aerodynamic design of gas turbines is a step-by-step process of design and optimization from low-dimensional to high-dimensional dimensions, with the results of the low-dimensional design serving as the initial values and foundation for the high-dimensional design. In one-dimensional design, the loss model plays a crucial role, enabling the selection of optimal inlet and outlet aerodynamic conditions, which is significant for the design of the entire turbine component. However, with technological advancements and continuous improvements in turbine blade performance, existing turbine loss prediction models exhibit poor prediction accuracy, particularly for blade losses across a wide angle-of-attack range. The relative prediction error is around 30% (AMDCKO turbine loss model); the prediction accuracy of turbine loss prediction models for non-zero angle-of-attack ranges is also poor. Therefore, it is necessary to revise existing loss models. Summary of the Invention
[0004] The purpose of this invention is to provide a turbine loss model construction method based on deep learning and loss weight analysis. While improving the prediction accuracy of the original model, it provides a simple and fast method for constructing a turbine loss model, which can accurately predict the turbine blade loss with a large angle of attack operating range.
[0005] The objective of this invention is achieved through the following technical solution:
[0006] A method for constructing a turbine loss model based on deep learning and loss weight analysis includes the following steps:
[0007] Step S1: Obtain the geometric parameters of the turbine two-dimensional airfoil, and conduct a planar blade cascade test on the turbine two-dimensional airfoil according to the selected operating conditions to obtain the inlet and outlet aerodynamic parameters of the turbine airfoil, wherein:
[0008] The geometric parameters include inlet geometric angle β1, outlet geometric angle β2, mounting angle γ, leading edge wedge angle We, leading edge relative thickness LEd / C, trailing edge relative thickness TEd / C, relative maximum thickness tmax / C, relative grid pitch t / C, and leading edge ellipticity LEb / LEa;
[0009] The inlet geometry angle β1, outlet geometry angle β2, and installation angle γ are all angles with the axial direction. LEa is the diameter of the ellipse perpendicular to the incoming flow direction in the leading edge ellipse under zero angle of attack. LEb is the diameter of the ellipse in the same direction as the incoming flow in the leading edge ellipse under zero angle of attack. LEd is the diameter of the blade leading edge circle (the leading edge circle is a perfect circle). C is the blade chord length. TEd is the diameter of the blade trailing edge circle (the trailing edge circle is a perfect circle). tmax is the maximum thickness of the blade and t is the grid pitch.
[0010] The inlet and outlet pneumatic parameters are the inlet and outlet pneumatic parameters under different operating conditions, and at least include the total inlet pressure P. * 1. Inlet static pressure P1, outlet static pressure P2, inlet Mach number Ma1, outlet Mach number Ma2, incoming flow angle of attack i, outlet flow angle α2, total pressure loss at the outlet section Y p The surface pressure distribution of the blade and the circumferential distribution of the total pressure loss at the outlet.
[0011] Step S2: Perform CFD calculations on the two-dimensional turbine blade profile. By comparing the surface pressure distribution, circumferential distribution of total outlet pressure loss, and magnitude of total outlet pressure loss obtained from the CFD calculations and the experimental measurements in Step S1, the CFD calculation results are verified.
[0012] Step S3: Combining the CFD calculation results obtained in Step S2 and the experimental measurement results obtained in Step S1, extract the magnitude of each component loss in the blade shape loss and the magnitude of the additional loss caused by the angle of attack of the incoming flow, where:
[0013] The methods for calculating the magnitude of each component loss in the airfoil loss include, but are not limited to:
[0014] 1) Regarding the total pressure loss coefficient Y of the boundary layer bl The specific calculation method is as follows:
[0015] The energy loss in the boundary layer is calculated based on the boundary layer energy thickness.
[0016]
[0017] Where u is the magnitude of the fluid velocity within the boundary layer, u ∞ ρ represents the mainstream velocity, ρ represents the fluid density within the boundary layer, and δ represents the nominal thickness of the boundary layer.
[0018] According to the definition of boundary layer energy thickness δ e :
[0019]
[0020] Where, ρ ∞ This represents the mainstream density. Therefore, the energy loss in the boundary layer is:
[0021]
[0022] The ideal kinetic energy at the outlet is calculated based on the assumption that the flow is isentropic:
[0023]
[0024] in, For mass flow rate, h 01 For the total enthalpy of imports, h s For the outlet static enthalpy, C P For the specific heat of a gas at constant pressure, T 01 The total inlet temperature is γ, the specific heat ratio is Ma. 2s The isentropic Mach number for the export.
[0025] The final calculated boundary layer energy loss coefficient is:
[0026]
[0027] The total pressure loss coefficient Y of the boundary layer is converted to the following formula. bl :
[0028]
[0029] 2) Regarding the wake loss Y m Because no shock wave is generated in the studied operating condition, the total loss Y is measured experimentally. p Subtract boundary layer loss Y bl To obtain, that is:
[0030] Y m =Y P -Y bl .
[0031] The additional loss ΔΦ caused by the angle of attack of the incoming flow 2 P The calculation methods include, but are not limited to, the difference between the energy loss coefficient at a non-zero angle of attack and the energy loss coefficient at a zero angle of attack.
[0032] Step S4: Decompose each loss component of the existing zero angle-of-attack turbine blade loss prediction empirical model and compare it with the corresponding loss magnitude extracted in Step S3 to analyze the coefficients that need to be corrected and the correction terms that need to be added in the zero angle-of-attack turbine blade loss prediction empirical model; For the empirical model that can predict the additional loss caused by the angle of attack (variable angle-of-attack model), compare it with the additional loss magnitude extracted in Step S3 to analyze the coefficients that need to be corrected and the correction terms that need to be added in the variable angle-of-attack model.
[0033] The existing empirical model for predicting zero angle-of-attack turbine blade loss is the AMDCKO turbine loss prediction model. The coefficients that need to be corrected and the correction terms that need to be added include:
[0034] 1) The exponent in the expression for the Mach number correction factor K2 for the leaf-shaped boundary layer loss is K2 = (Ma1 / Ma2). 2 Replace with K2 = (Ma1 / Ma2) a , where a is the term to be corrected;
[0035] 2) The exponent in the expression for the Mach number correction factor K1 for the leaf-shaped boundary layer loss is K1 = 1 - (1.25(Ma2 - 0.2)). 1 Replace with K1 = 1 - (1.25(Ma2 - 0.2)) b , where b is the term to be corrected;
[0036] 3) The weight of boundary layer loss is determined by replacing the coefficient 2 / 3, which is only related to the magnitude of boundary layer loss in the original formula, with a constant coefficient W. P W P These are coefficients to be determined;
[0037] 4) Predicted value Y' of wake loss TET Multiply by the compressible correction term K tet Its expression is:
[0038]
[0039] Among them, c, d, e, and f are all items to be determined.
[0040] The empirical model capable of predicting additional losses due to angle of attack is the Benner turbine loss prediction model. The coefficients that need to be corrected and the correction terms that need to be added include:
[0041] 1) Expression for the angle of attack variable χ:
[0042]
[0043] Become
[0044]
[0045] Where LEd is the diameter of the leading edge circle, We is the leading edge wedge angle, β1 is the inlet geometric angle, β2 is the outlet geometric angle, l, m, and n are undetermined coefficients, and A and B are terms to be determined.
[0046] 2) Combine the piecewise functions that use the angle of attack variable χ to calculate the additional loss caused by the angle of attack into a single function ΔΦ. 2 P =F(χ).
[0047] Step S5: Construct the empirical model for predicting turbine blade losses using a neural network model, and define the functional forms of the coefficients that need to be corrected and the correction terms that need to be added in the empirical model that can predict additional losses caused by angle of attack. Where:
[0048] When constructing the functional forms of the coefficients that need to be corrected and the correction terms that need to be added in the empirical model for predicting turbine blade losses, it is assumed that a, b, c, d, e, and f are all functions of the relative blade thickness tmax / C, the ratio of the blade installation angle to the airflow turning angle under zero angle of attack condition γ / (β1+α2), and the relative pitch t / C.
[0049]
[0050] Six functions, F1, F2, G1, G2, H1, and H2, were constructed using an artificial neural network model. The artificial neural network model used has three hidden layers, each with 400 neurons, and the activation function is the PReLu function model.
[0051] When constructing the functional forms of the coefficients that need to be corrected and the correction terms that need to be added in the empirical model that can predict the additional losses caused by the angle of attack, we assume that A, B, and F(χ) have the following functional forms:
[0052] a. For the term A to be determined, it is assumed that A is a function of the ratio of the airfoil installation angle to the geometric turning angle γ / (β1+β2), the relative airfoil thickness tmax / C, the relative pitch t / C, and the leading edge ellipticity LEb / LEa, i.e.:
[0053] A=f(t max / C, t / C, LEb / LEa);
[0054] The artificial neural network model is constructed using an artificial neural network model, which has 8 hidden layers, each with 2400 neurons, and the activation function is selected as the PReLU function.
[0055] b. For the term B to be determined, it is assumed that B is a function of the ratio of the airfoil installation angle to the geometric turning angle γ / (β1+β2), the relative airfoil thickness tmax / C, the relative pitch t / C, the leading edge ellipticity LEb / LEa, the incoming flow angle of attack i, the ratio of the leading edge diameter to the pitch LEd / t, the leading edge wedge angle We, and the airfoil convergence ratio cos(β1) / cos(β2), i.e.:
[0056] B=f(t max / C, t / C, i, LEb / LEa);
[0057] When the leading edge small circle is an ellipse, the value of the leading edge diameter LEd is the same as that of LEa;
[0058] The artificial neural network model g is constructed using an artificial neural network model, which has 8 hidden layers, each with 2000 neurons, and the activation function is the PReLU function.
[0059] c. For F(χ), it is considered that F(χ) is a function of the angle of attack variable χ raised to the power of 1 to 8, that is:
[0060]
[0061] Constructing a function using an artificial neural network model The artificial neural network model used has 7 hidden layers, each with 3000 neurons, and the activation function is the PReLu function.
[0062] Step S6: Construct a sample set using the airfoil parameters obtained in Step S1 and the corresponding airfoil loss obtained in Step S3. Then, use the sample set to train the empirical model for predicting turbine airfoil loss and the empirical model that can predict additional losses caused by angle of attack, to obtain the final result, wherein:
[0063] Determine the functional forms and undetermined constant coefficients W of the six functions F1, F2, G1, G2, H1, and H2. P The specific method for determining the exact size is as follows:
[0064] 1) For F1, F2, W P Substituting this into the modified form of the boundary layer loss prediction formula, the formula is as follows:
[0065]
[0066] Among them, Y P,AMDC The value of Re represents the airfoil loss calculated using the AMDC turbine loss model, where Re is the Reynolds number.
[0067] Using airfoil geometry parameters and inlet / outlet aerodynamic parameters as inputs, and boundary layer loss Y as the input... blThe output sample set is used, and 8 / 9 of the sample size is randomly selected for training the model, ultimately determining the specific parameters of F1 and F2 and W. P The specific size;
[0068] 2) For G1, G2, H1, and H2, substitute them into the corrected form of the wake loss prediction formula, as shown below:
[0069]
[0070] Among them, Y TET,AMDCKO The magnitude of the wake loss is calculated using the AMDCKO turbine loss model;
[0071] Using airfoil geometry parameters and inlet / outlet aerodynamic parameters as inputs, and the wake loss Y as the input... m The output sample set is used, and 8 / 9 of the sample size is randomly selected to train the model, and finally the functions G1, G2, H1, and H2 are determined.
[0072] Determine the magnitudes of l, m, and n, and the functions f, g, When considering the specific form, the specific method is as follows:
[0073] l, m, n, f, g, Substituting this into the formula for predicting additional losses due to angle of attack in the corrected form, the formula is as follows:
[0074]
[0075] The model is trained using a sample set with airfoil geometry parameters and inlet / outlet aerodynamic parameters as inputs and additional losses due to angle of attack as outputs, ultimately determining f, g, and The specific parameters.
[0076] When training the empirical model for predicting turbine blade loss at zero angle of attack and the empirical model that can predict additional losses due to angle of attack, the sample set is randomly divided into a training set and a test set at a ratio of 8:1. The model is trained and the generalization ability of the training results is tested. The HuberLoss loss function is selected to evaluate the prediction accuracy of the model.
[0077] In this invention, the model predicts the total pressure loss coefficient generated by the two-dimensional airfoil of the flow turbine; the relative error predicted by the model construction method is within 5%.
[0078] In this invention, the turbine is an axial flow structure, and the working medium of the turbine is air, oxygen, natural gas, etc. The application fields of the turbine include medium and low temperature waste heat power generation devices, renewable energy power generation devices, aero engines, and gas turbines.
[0079] Compared with the prior art, the present invention has the following advantages:
[0080] (1) The requirements for the user's understanding of the impeller mechanical flow are relatively low.
[0081] (2) By using deep learning methods, the functional relationship between the two-dimensional turbine airfoil geometric parameters and the coefficients of the existing turbine airfoil loss prediction model can be obtained more accurately and quickly, and a new turbine airfoil loss prediction model can be constructed. Attached Figure Description
[0082] Figure 1 This is a flowchart illustrating the construction of the turbine blade loss model based on deep learning and loss weight analysis in this invention.
[0083] Figure 2 It takes the form of an artificial neural network model. Detailed Implementation
[0084] The technical solution of the present invention will be further described below with reference to the accompanying drawings, but it is not limited thereto. Any modifications or equivalent substitutions to the technical solution of the present invention that do not depart from the spirit and scope of the technical solution of the present invention should be covered within the protection scope of the present invention.
[0085] This invention provides a method for constructing a turbine blade loss model based on deep learning and loss weight analysis. To explore how to modify existing turbine blade loss models, the method first decomposes the various losses in the blade loss and compares them with the predicted values of existing models. This analysis identifies the coefficients and terms that need modification and the correction terms that need to be added, thus forming a modified loss prediction model. Furthermore, by comparing turbine blade losses with different blade parameters, the blade parameter variables that need to be considered are identified. An artificial neural network model is used to establish the functional relationship between the blade parameter variables to be considered and the coefficients (or terms) and correction terms that need to be modified. This relationship is then input into the modified loss prediction model to construct the turbine blade loss prediction model. Figure 1 As shown, it includes the following steps:
[0086] 1) Using CAE software, represent the two-dimensional turbine airfoil parameters and obtain the turbine's two-dimensional airfoil geometric parameters (including at least the inlet geometric angle β1, outlet geometric angle β2, installation angle γ, leading edge wedge angle We, leading edge relative thickness LEd / C, trailing edge relative thickness TEd / C, relative maximum thickness tmax / C, relative pitch t / C, and leading edge ellipticity LEb / LEa). Conduct a blowing test on the turbine airfoil planar blade cascade to obtain the inlet and outlet aerodynamic parameters (including at least the inlet pressure P1, outlet pressure P2, inlet Mach number Ma1, outlet Mach number Ma2, incoming flow angle of attack i, outlet flow angle α2, and the magnitude of the total pressure loss at the outlet section Y). p ).
[0087] 2) The two-dimensional turbine blade profile is calculated using commercial CFD software, and the CFD calculation results are verified by comparing the profile pressure distribution, circumferential distribution of total outlet pressure loss and the magnitude of total outlet pressure loss obtained from CFD calculation and experimental measurement.
[0088] 3) Combining the CFD calculation results and experimental measurement results, the loss is broken down as follows:
[0089] a. For the zero angle-of-attack inflow condition, the loss is decomposed into boundary layer loss and wake loss:
[0090] i> Boundary layer loss — The magnitude of energy loss in the boundary layer is calculated based on the boundary layer energy thickness:
[0091]
[0092] According to the definition of boundary layer energy thickness δ e :
[0093]
[0094] Therefore, the magnitude of energy loss in the boundary layer is:
[0095]
[0096] And the ideal kinetic energy at the outlet is calculated based on the assumption that the flow is isentropic:
[0097]
[0098] The final calculated boundary layer energy loss coefficient is:
[0099]
[0100] The total pressure loss coefficient Y of the boundary layer is converted to the following formula. bl :
[0101]
[0102] ii>Wake loss—Since no shock wave is generated in the studied operating condition, the total loss Y is measured experimentally. p Subtract boundary layer loss Y bl To obtain, that is:
[0103] Y m =Y P -Y bl .
[0104] b. For operating conditions with non-zero angle of attack incoming flow, the additional loss ΔΦ caused by the angle of attack of the incoming flow is calculated by using the difference between the energy loss coefficient under non-zero angle of attack and the energy loss coefficient under zero angle of attack. 2 P .
[0105] 4) For the existing empirical model for predicting turbine blade loss (which can predict the loss of each component in the blade loss) (AMDCKO model), compare the predicted values of the loss of each component with the corresponding loss magnitude extracted in step 3), and analyze the coefficients that need to be corrected and the correction terms that need to be added, including:
[0106] a. Correct the exponent in the expression for the Mach number correction factor K2 for blade boundary layer loss in the AMDCKO turbine loss prediction model, changing the original K2 = (Ma1 / Ma2) to (Ma1 / Ma2). 2 Replace with K2 = (Ma1 / Ma2) a , where a is the term to be corrected;
[0107] b. Correct the exponent in the expression for the Mach number correction factor K1 for blade boundary layer loss in the AMDCKO turbine loss prediction model. The original K1 = 1 - (1.25(Ma2 - 0.2)) 1 Replace with K1 = 1 - (1.25(Ma2 - 0.2)) b , where b is the term to be corrected;
[0108] c. In the revised AMDCKO turbine loss prediction model, the coefficient 2 / 3, which is only related to the magnitude of boundary layer loss, is replaced with coefficient W. P W P These are coefficients to be determined;
[0109] d. The predicted value Y' of the wake loss in the AMDCKO turbine loss prediction model. TET Multiply by the compressible correction term K tet Its expression is:
[0110]
[0111] Among them, c, d, e, and f are all items to be determined.
[0112] For the empirical model predicting the additional losses caused by the angle of attack, its predicted values are compared with the magnitude of the additional losses extracted in step 3), and the coefficients that need to be corrected and the correction terms that need to be added are analyzed, including:
[0113] a. The expression for the angle of attack variable χ in the prediction model:
[0114]
[0115] Become
[0116]
[0117] Among them, A and B are items to be determined;
[0118] b. The piecewise functions in the original model (Benner model) that used the angle of attack variable χ to calculate the additional loss caused by the angle of attack were merged into a single function ΔΦ. 2 P =F(χ).
[0119] 5) Construct the functional form of the coefficients to be corrected and the terms to be added using an artificial neural network model:
[0120] a. For the existing empirical model for predicting turbine blade losses (which can predict the losses of each component in the blade loss) (AMDCKO model): the terms to be corrected in step 4), a, b, c, d, e, and f, are all functions of the blade relative thickness tmax / C, the ratio of the blade installation angle to the airflow deflection angle under zero angle of attack condition γ / (β1+α2), and the relative pitch t / C, that is:
[0121]
[0122] Six functions, F1, F2, G1, G2, H1, and H2, were constructed using an artificial neural network model. This model has three hidden layers, each with 400 neurons, and uses the PReLU activation function. The neural network structure is as follows: Figure 2 As shown in (a), the specific parameters in the model will be obtained later using experimental measurement results and model training.
[0123] b. Empirical models for predicting additional losses due to angle of attack:
[0124] For the term A to be determined, it is assumed that A is a function of the ratio of the airfoil installation angle to the geometric turning angle γ / (β1+β2), the relative airfoil thickness tmax / C, the relative pitch t / C, and the leading edge ellipticity LEb / LEa, i.e.:
[0125] A=f(t max / C, t / C, LEb / LEa);
[0126] The artificial neural network model f is constructed using an artificial neural network model. This model has 8 hidden layers, each with 2400 neurons. The activation function is PReLU. The neural network takes the form shown below. Figure 2 As shown in (b), the specific parameters in the model will be obtained during subsequent model training;
[0127] ii> For the term B to be determined, it is assumed that B is a function of the ratio of the airfoil installation angle to the geometric turning angle γ / (β1+β2), the relative airfoil thickness tmax / C, the relative pitch t / C, the leading edge ellipticity LEb / LEa, the incoming flow angle of attack i, the ratio of the leading edge diameter to the pitch d / t, the leading edge wedge angle We, and the airfoil convergence ratio cos(β1) / cos(β2), that is:
[0128] B=g(γ / (β1+β2), t max / C, f / C, LEb / LEa, i, LEd / t, We, cos(β1) / cos(β2));
[0129] When the leading edge small circle is an ellipse, the value of the leading edge diameter LEd is the same as that of LEa;
[0130] Constructor g using an artificial neural network model, wherein the artificial neural network model has 8 hidden layers, each with 2000 neurons, and the activation function is the PReLU function. The neural network structure is as follows. Figure 2 As shown in (c), the specific parameters in the model will be obtained during subsequent model training;
[0131] iii> For F(χ), it is considered that F(χ) is a function of the angle of attack variable χ raised to the power of 1 to 8, that is:
[0132]
[0133] Constructing a function using an artificial neural network model The artificial neural network model used has 7 hidden layers, each with 3000 neurons. The activation function is the PReLU function. The neural network structure is as follows: Figure 2 As shown in (d), the specific parameters in the model will be obtained during subsequent model training.
[0134] 6) Construct a sample set using the leaf shape parameters obtained in step 1) and the corresponding leaf shape loss obtained in step 3). Randomly divide the sample set into a training set and a test set at an 8:1 ratio. Train the model and test the generalization ability of the training results. Select the HuberLoss loss function to evaluate the model's prediction accuracy. The construction of the sample set differs for different correction terms.
[0135] a. For F1, F2, W P Substitute these parameters into the modified boundary layer loss prediction formula, using the airfoil geometry parameters and inlet / outlet aerodynamic parameters described in step 1) as inputs, and the boundary layer loss Y extracted in step 3) as input. bl The model is trained using the output sample set, ultimately determining the specific functions of F1, F2, and W. PThe specific size;
[0136] b. For G1, G2, H1, and H2, substitute them into the modified wake loss prediction formula, using the airfoil geometry parameters described in step 1) and the inlet / outlet aerodynamic parameters described in step 1) as inputs, and the wake loss Y extracted in step 3). m The model is trained using the output sample set, and the specific functions of G1, G2, H1, and H2 are finally determined.
[0137] c. Regarding f, g, Substituting this into the modified form of the formula for predicting additional losses due to angle of attack, and using the airfoil geometry parameters described in step 1) and the inlet / outlet aerodynamic parameters described in step 1) as inputs, and the sample set of additional losses due to angle of attack extracted in step 3) as outputs, the model is trained to finally determine f, g, The specific function.
Claims
1. A method for constructing a turbine loss model based on deep learning and loss weight analysis, characterized in that... The method includes the following steps: Step S1: Obtain the geometric parameters of the turbine two-dimensional airfoil, and conduct a planar blade cascade test on the turbine two-dimensional airfoil according to the selected working conditions to obtain the inlet and outlet aerodynamic parameters of the turbine airfoil. Step S2: Perform CFD calculations on the two-dimensional turbine blade profile. By comparing the surface pressure distribution, circumferential distribution of total outlet pressure loss, and magnitude of total outlet pressure loss obtained from the CFD calculations and the experimental measurements in Step S1, the CFD calculation results are verified. Step S3: Combining the CFD calculation results obtained in step S2 and the experimental measurement results obtained in step S1, extract the magnitude of the loss of each component in the blade shape loss and the magnitude of the additional loss caused by the angle of attack of the incoming flow. Step S4: Decompose each loss component of the existing zero angle-of-attack turbine blade loss prediction empirical model and compare it with the corresponding loss magnitude extracted in Step S3 to analyze the coefficients that need to be corrected and the correction terms that need to be added in the zero angle-of-attack turbine blade loss prediction empirical model; For the empirical model that can predict the additional loss caused by the angle of attack, compare it with the additional loss magnitude extracted in Step S3 to analyze the coefficients that need to be corrected and the correction terms that need to be added in the model. Step S5: Construct the empirical model for predicting turbine blade loss using a neural network model, and the functional form of the coefficients that need to be corrected and the correction terms that need to be added in the empirical model that can predict the additional loss caused by the angle of attack. Step S6: Construct a sample set using the blade parameters obtained in step S1 and the corresponding blade loss obtained in step S3. Then, use the sample set to train the empirical model for predicting turbine blade loss and the empirical model that can predict additional losses caused by angle of attack, and obtain the final result.
2. The method for constructing a turbine loss model based on deep learning and loss weight analysis according to claim 1, characterized in that... In step S1, the geometric parameters include inlet geometric angle β1, outlet geometric angle β2, mounting angle γ, leading edge wedge angle We, leading edge relative thickness LEd / C, trailing edge relative thickness TEd / C, relative maximum thickness tmax / C, relative pitch t / C, and leading edge ellipticity LEb / LEa. Inlet geometric angle β1, outlet geometric angle β2, and mounting angle γ are all angles with the axial direction. LEa is the diameter of the ellipse perpendicular to the incoming flow direction in the leading edge ellipse under zero angle of attack. LEb is the diameter of the ellipse in the leading edge ellipse with the same direction as the incoming flow under zero angle of attack. LEd is the diameter of the blade leading edge circle. C is the blade chord length. TEd is the diameter of the blade trailing edge circle. tmax is the maximum thickness of the blade, and t is the pitch.
3. The method for constructing a turbine loss model based on deep learning and loss weight analysis according to claim 1, characterized in that... In step S1, the inlet and outlet pneumatic parameters are the inlet and outlet pneumatic parameters under different operating conditions, including at least the total inlet pressure. Inlet static pressure P1, outlet static pressure P2, inlet Mach number Ma1, outlet Mach number Ma2, incoming flow angle of attack i, outlet flow angle α2, total pressure loss at the outlet section Y p The surface pressure distribution of the blade and the circumferential distribution of the total pressure loss at the outlet.
4. The method for constructing a turbine loss model based on deep learning and loss weight analysis according to claim 1, characterized in that... In step S3, the calculation method for the magnitude of each component loss in the airfoil loss is as follows: 1) Regarding the total pressure loss coefficient of the boundary layer The specific calculation method is as follows: The energy loss in the boundary layer is calculated based on the boundary layer energy thickness. ； in, The magnitude of the fluid velocity within the boundary layer. For mainstream speed size, The density of the fluid within the boundary layer. The nominal thickness of the boundary layer; According to the definition of boundary layer energy thickness : ； in, Given the mainstream density, the energy loss in the boundary layer is: ； The ideal kinetic energy at the outlet is calculated based on the assumption that the flow is isentropic: ； in, For quality flow, The total enthalpy of imports. For the outlet static enthalpy, For the specific heat of a gas at constant pressure, For the total import temperature, For specific heat ratio, The isentropic Mach number for the export. The final calculated boundary layer energy loss coefficient is: ； The total pressure loss coefficient Y of the boundary layer is converted to the following formula. bl : ； 2) Regarding wake loss Utilizing the total pressure loss at the outlet section Subtract boundary layer loss To obtain, that is: ； The additional loss ΔΦ caused by the angle of attack of the incoming flow 2 P The calculation method is: the difference between the energy loss coefficient at a non-zero angle of attack and the energy loss coefficient at a zero angle of attack.
5. The method for constructing a turbine loss model based on deep learning and loss weight analysis according to claim 1, characterized in that... In step S4, the existing empirical model for predicting zero angle-of-attack turbine blade loss is the AMDCKO turbine loss prediction model. The coefficients that need to be corrected and the correction terms that need to be added include: 1) The exponentiation of the Mach number correction factor K2 in the expression for the leaf-shaped boundary layer loss is K2=(Ma1 / Ma2). 2 Replace with K2=(Ma1 / Ma2) a , where a is the term to be corrected, Ma1 is the Mach number for imports, and Ma2 is the Mach number for exports; 2) The exponent in the expression for the Mach number correction factor K1 for the leaf-shaped boundary layer loss is K1 = 1 - (1.25(Ma2 - 0.2)). 1 Replace with K1=1-(1.25(Ma2-0.2)) b , where b is the term to be corrected; 3) The weight of boundary layer loss is determined by replacing the coefficient 2 / 3, which is only related to the magnitude of boundary layer loss, with the coefficient W in the original formula. P W P These are coefficients to be determined; 4) Predicted value Y of wake loss ’ TET Multiply by the compressible correction term K tet Its expression is: ； Among them, c, d, e, and f are all items to be determined.
6. The method for constructing a turbine loss model based on deep learning and loss weight analysis according to claim 5, characterized in that... In step S4, the empirical model that can predict the additional losses caused by the angle of attack is the Benner turbine loss prediction model. The coefficients that need to be corrected and the correction terms that need to be added include: 1) Angle of attack variable The expression: Become ； Where LEd is the diameter of the leading edge circle, We is the leading edge wedge angle, β1 is the inlet geometric angle, β2 is the outlet geometric angle, l, m, and n are undetermined coefficients, and A and B are terms to be determined. 2) The angle of attack variable will be utilized The piecewise functions that calculate the additional loss due to the angle of attack are combined into a single function. .
7. The method for constructing a turbine loss model based on deep learning and loss weight analysis according to claim 6, characterized in that... In step S5, when constructing the functional forms of the coefficients that need to be corrected and the correction terms that need to be added in the empirical model for predicting turbine blade loss, it is assumed that a, b, c, d, e, and f are all functions of the relative thickness tmax / C of the blade, the ratio γ / (β1+α2) of the airflow turning angle under the condition of blade installation angle and zero angle of attack, and the relative pitch t / C. Six functions, F1, F2, G1, G2, H1, and H2, were constructed using an artificial neural network model. The artificial neural network model used has three hidden layers, each with 400 neurons, and the activation function is the PReLu function model.
8. The method for constructing a turbine loss model based on deep learning and loss weight analysis according to claim 7, characterized in that... In step S5, when constructing the functional form of the coefficients that need to be corrected and the correction terms that need to be added in the empirical model that can predict the additional losses caused by the angle of attack, it is assumed that A, B, It takes the following function form: a. For the term A to be determined, it is assumed that A is a function of the ratio of the airfoil installation angle to the geometric turning angle γ / (β1+β2), the relative airfoil thickness tmax / C, the relative pitch t / C, and the leading edge ellipticity LEb / LEa, i.e.: ； The artificial neural network model is constructed using an artificial neural network model, which has 8 hidden layers, each with 2400 neurons, and the activation function is selected as the PReLU function. b. For the term B to be determined, it is assumed that B is a function of the ratio of the airfoil installation angle to the geometric turning angle γ / (β1+β2), the relative airfoil thickness tmax / C, the relative pitch t / C, the leading edge ellipticity LEb / LEa, the incoming flow angle of attack i, the ratio of the leading edge diameter to the pitch LEd / t, the leading edge wedge angle We, and the airfoil convergence ratio cos(β1) / cos(β2), i.e.: ； When the leading edge small circle is an ellipse, the value of the leading edge diameter LEd is the same as that of LEa; The artificial neural network model g is constructed using an artificial neural network model, which has 8 hidden layers, each with 2000 neurons, and the activation function is the PReLU function. c. For ,think It is the angle of attack variable The first to eighth powers of the variable are functions of the variable, that is: ； Constructing a function using an artificial neural network model The artificial neural network model used has 7 hidden layers, each with 3000 neurons, and the activation function is the PReLu function. Determine the magnitudes of l, m, and n, and the functions f, g, When considering the specific form, the specific method is as follows: Substituting l, m, n, f, g, and φ into the modified form of the formula for predicting additional losses due to angle of attack, the formula is as follows: ； The model was trained using a sample set with blade geometry parameters and inlet / outlet aerodynamic parameters as inputs and the additional loss due to angle of attack as output, ultimately determining f, g, and The specific parameters.
9. The method for constructing a turbine loss model based on deep learning and loss weight analysis according to claim 8, characterized in that... In step S6, the functional forms and undetermined constant coefficients W of the six functions F1, F2, G1, G2, H1, and H2 are determined. P The specific method for determining the exact size is as follows: 1) For F1, F2, W P Substituting this into the modified form of the boundary layer loss prediction formula, the formula is as follows: ； Among them, Y P,AMDC The value of Re represents the airfoil loss calculated using the AMDC turbine loss model, where Re is the Reynolds number. Using airfoil geometry parameters and inlet / outlet aerodynamic parameters as inputs, and boundary layer loss Y as the input... bl The output sample set is used, and 8 / 9 of the sample size is randomly selected for training the model, ultimately determining the specific parameters of F1 and F2 and W. P The specific size; 2) For G1, G2, H1, and H2, substitute them into the corrected form of the wake loss prediction formula, as shown below: ； Among them, Y TET,AMDCKO The magnitude of the wake loss is calculated using the AMDCKO turbine loss model; Using airfoil geometry parameters and inlet / outlet aerodynamic parameters as inputs, and the wake loss Y as the input... m The output sample set is used, and 8 / 9 of the sample size is randomly selected to train the model, and finally the functions G1, G2, H1, and H2 are determined.
10. The method for constructing a turbine loss model based on deep learning and loss weight analysis according to claim 1, characterized in that... In step S6, when training the zero angle-of-attack turbine blade loss prediction empirical model and the empirical model that can predict additional losses caused by the angle of attack, the sample set is randomly divided into a training set and a test set in a ratio of 8:
1. The model is trained and the generalization ability of the training results is tested. The HuberLoss loss function is selected to evaluate the prediction accuracy of the model.