Turbine channel inlet temperature rapid prediction method based on Fourier neural operator
Through the prediction method based on Fourier neural operator, a small amount of measurement point data is used to quickly and accurately predict the temperature distribution of turbine inlets, solving the problems of missing data and low computing efficiency in the prior art, and achieving efficient and accurate description of temperature distribution.
Patent Information
- Application Number
- CN202510104809.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-01-23
- Publication Date
- 2025-05-23
AI Technical Summary
The prior art is difficult to quickly and accurately predict the temperature distribution of turbine inlets while ensuring accuracy, especially in the problems of missing data and low computing efficiency.
The prediction method based on Fourier neural operator is adopted to construct a prediction model through a small amount of measurement point data on the downstream cross section of the turbine channel to achieve rapid prediction of the inlet temperature distribution of the turbine channel.
It realizes high efficiency, high precision and high resolution turbine inlet temperature distribution description, with a maximum error of less than 10%, effectively solving the problem of rapid and accurate prediction of inlet temperature.
Smart Images
Figure CN120030937A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of engineering thermophysics, and in particular relates to a method for quickly predicting turbine channel inlet temperature based on a Fourier neural operator. Background Art
[0002] The temperature distribution at the turbine inlet is a key factor affecting the performance of aircraft engines. Accurately understanding the temperature field at the turbine inlet not only helps to optimize the engine aerodynamic design and improve fuel efficiency, but also effectively reduces the heat load and extends the service life of the engine. As a key component, the turbine blades must withstand extreme high temperature and high pressure environments. By accurately predicting the inlet temperature distribution, the thermal protection of the turbine blades can be efficiently carried out, the cooling scheme design can be optimized, and thermal stress analysis can be performed, thereby improving the aerodynamic performance and safety of the turbine. Therefore, the prediction of the turbine inlet temperature distribution is one of the key contents of turbine channel research.
[0003] At present, the acquisition of turbine inlet temperature distribution mainly relies on two methods: experimental measurement and numerical simulation. Experimental measurement methods include infrared thermal imaging, thermocouple measurement, and high-frequency response hot wire anemometer. These methods have limitations in practical applications. For example, although thermocouples can provide high-precision data, they not only cannot fully cover the entire turbine inlet area, but also cannot be monitored dynamically in real time. In addition, the layout of the inlet thermocouples will affect the flow and heat transfer results in the channel, thereby affecting the accuracy of the measurement results; on the other hand, although the numerical simulation method can obtain more comprehensive temperature field information, its calculation cost is high, especially under complex three-dimensional flow fields and high Reynolds numbers. The high-precision simulation process is very time-consuming and resource-intensive, and the accuracy of the low-precision numerical simulation calculation results cannot meet many engineering requirements. Therefore, how to solve the problems of missing data and low calculation efficiency while ensuring accuracy is a difficult problem that needs to be solved for accurate prediction of inlet temperature distribution.
[0004] In recent years, the application of deep learning in various fields has achieved remarkable results, especially in image processing, natural language processing and physical modeling. Deep learning methods have begun to be applied to the prediction of turbine inlet temperature distribution. For example, convolutional neural networks (CNN) and recurrent neural networks (RNN) are used to predict temperature fields based on historical data and sensor data. These methods can achieve good prediction results when the data volume is large and the quality is high. However, these methods also have some disadvantages. First, deep learning models usually require a large amount of high-quality data for training, while in actual engineering, high-precision turbine inlet temperature data is often scarce and difficult to obtain. Second, existing deep learning methods are often difficult to effectively combine physical knowledge, resulting in deficiencies in extrapolation ability and physical interpretability. In addition, the training and reasoning processes of many deep learning models are complex, the computing resource requirements are high, and their generalization ability is limited when dealing with complex fluid dynamics problems. Therefore, how to use less data while maintaining high accuracy has become a key issue in the application of deep learning methods.
[0005] With the continuous development of deep learning technology, neural operator methods have gradually become a new and effective way to solve physical problems. Fourier neural operator (FNO) is a physical modeling method based on deep learning. By combining convolutional neural networks with Fourier transform, it can achieve efficient solution of complex physical systems. The advantage of the FNO method is that it can process data with periodic boundary conditions and high-dimensional input and output, and can effectively use the prior knowledge of the physical model to improve the accuracy and generalization ability of the model. Compared with traditional neural networks, FNO can process data in the frequency domain through Fourier transform, thereby significantly reducing the amount of calculation and improving the calculation efficiency. In addition, FNO can achieve excellent performance in problems such as fluid dynamics and heat conduction, especially when solving complex nonlinear problems. Its accuracy and calculation efficiency have obvious advantages over traditional numerical simulation methods. It is a feasible and effective technical path to quickly predict the temperature distribution of the turbine inlet through the sparse measurement point data of the downstream or outlet section of the turbine channel combined with the FNO neural operator, which provides a new way to obtain the temperature distribution of the turbine inlet section. Summary of the invention
[0006] In order to solve the problem that the turbine inlet temperature distribution is difficult to obtain, the purpose of the present invention is to provide a method for quickly predicting the turbine channel inlet temperature based on Fourier neural operator, which can quickly predict the turbine inlet temperature distribution through a small amount of measurement point data in the downstream section of the turbine channel.
[0007] In order to achieve the above object, the present invention is implemented by the following technical solutions:
[0008] A method for quickly predicting turbine channel inlet temperature based on Fourier neural operator, characterized by comprising the following steps:
[0009] Step 1: Obtain a data set, including establishing a turbine channel model, obtaining samples of different inlet temperature distributions by designing different inlet hot spot positions, and obtaining a downstream outlet section temperature distribution data set through CFD numerical simulation calculations.
[0010] Step 2: Extract 6*6 sparse lattice data of downstream section temperature data as input, and the inlet section temperature distribution as output.
[0011] Step 3: Set the total number of samples to 50, read the input and output data partition labels of step 2, and divide them into training set, test set and validation set.
[0012] Step 4: Based on the FNO neural network architecture, the input set data in the training set in step 3 is transferred to the training data in the FNO model, and the output set data in the training set in step 3 is transferred to the loss function for training, error calculation and updating of training parameters.
[0013] Step 5: Input the input set data of the validation set into the training model, and use the output set data of the validation set to verify the model to obtain the prediction model.
[0014] Step 6: Input the input set data of the test set into the prediction model to obtain the prediction results of the turbine channel inlet temperature distribution.
[0015] Furthermore, the specific steps of obtaining the data set in step 1 include:
[0016] Step 1-1: Determine the size of the hot spot and the temperature variation range, divide the center position of the hot spot, and perform arithmetic division in the radius and angle direction according to the shape of the inlet fan-shaped section to obtain different turbine inlet temperature distribution conditions.
[0017] Step 1-2: Obtain the inlet temperature distribution matrix based on the size, temperature and center position of the hot spot, which contains the temperature information of different points, and obtain the output data of different samples.
[0018] Step 1-3: Input the set inlet conditions into the numerical simulation calculation model, and obtain the calculation results based on the RANS numerical calculation method and the automatic calculation process of the script.
[0019] Step 1-4: Extract the temperature distribution results of the downstream section and output them as a txt file. Using the interpolation algorithm, the data containing x, y, z and temperature are interpolated into a 50*50 matrix to obtain the input data of different samples.
[0020] Furthermore, the specific steps of training in step 4 include:
[0021] Step 4-1: Build the FNO model framework, which is inspired by the Green function and constructs the following operator space expression:
[0022] υ t (x) = o(W*υ t-1 (x)+∫ D k(x,y,a(x),a(y))*υ x (dy))
[0023] where v t It represents the inlet temperature distribution data of different iteration steps, σ represents the activation function, W represents the linear transformation part of the prediction value between two adjacent iteration steps, k represents the Fourier transform and inverse transform process, and a represents the scattered point data of the downstream turbine channel.
[0024] The representation of the output data mainly consists of two parts, the linear transformation part and the nonlinear transformation part. The linear transformation part is controlled by the parameter W; the nonlinear transformation part first transforms the time domain data into the frequency domain data by Fourier transform. The specific expression is as follows:
[0025] F[f j (K)]=∫ D f j (K)e -2iπ<x，K> dx
[0026] Where F represents the Fourier transform function, f represents the value to be transformed, K represents the convolution kernel, and e -2iπ<x，K> Represents a complex function. Since the input data is a 50*50 matrix, 50 frequency domain modes are obtained after Fourier transform. The first five high-frequency modes and the last five low-frequency modes are saved as effective modes, and other modes are ignored. The effective modes can effectively transform the integral and differential transformation of the operator space into a product calculation and fit after linear transformation. The modes after linear transformation are returned to the time domain space through inverse Fourier transform. The specific formula is as follows:
[0027] F -1 [f j (K)]=∫ D f j (K)e -2iπ<x,K> dx
[0028] where F -1 represents the inverse Fourier transform function, f represents the value to be transformed, K represents the convolution kernel, e-2iπ<xz，K> Represents a complex function. Repeating the above process multiple times is the iterative calculation of the Fourier operator. In this paper, the number of iterations is set to 3.
[0029] Step 4-2: Import input and output data, standardize them, and randomly shuffle them into the input model framework for calculation.
[0030] Step 4-3: Use the mini-batch stochastic gradient descent algorithm and set the batch size to 4. Calculate the model error through forward propagation; update the model parameters through back propagation until the training reaches the specified number of cycles or the error is less than the set value and then stop updating.
[0031] Beneficial effects of the present invention:
[0032] 1. The present invention can construct a prediction model based on the Fourier neural operator according to the mapping relationship between the sparse lattice data of the turbine channel downstream and the channel inlet temperature distribution, predict the temperature distribution of the turbine channel inlet section, achieve high-efficiency, high-precision, and high-resolution description of the turbine inlet temperature distribution, and complete the function of restoring the inlet temperature distribution according to different downstream lattice data distribution conditions.
[0033] 2. The present invention uses pixel difference as an indicator: in the comparison of single pixel results, the maximum error of the training result is less than 5%, and the maximum error of the test result is less than 10%; in the comparison of the average results of all pixel points in a sample, the maximum error of the training result is less than 2%, and the maximum error of the test result is less than 5%; using the hot spot center coordinates as an indicator: the maximum errors of the training set and the test set are both less than 1%, so the inlet temperature can be effectively predicted quickly and accurately. BRIEF DESCRIPTION OF THE DRAWINGS
[0034] Figure 1 Turbine channel structure diagram.
[0035] Figure 2 Schematic diagram of import condition settings.
[0036] Figure 3 Calculation results of the stator blade outlet section.
[0037] Figure 4 Framework diagram of the Fourier neural operator model.
[0038] Figure 5 Loss change graph during training.
[0039] Figure 6 Comparison chart of prediction results for the first test sample.
[0040] Figure 7 Comparison chart of prediction results for the second test sample.
[0041] Figure 8 Comparison chart of prediction results for the third test sample.
[0042] Fig. 9 Comparison chart of the prediction results of the 4th test sample.
[0043] Fig.10 Comparison chart of prediction results for the 5th test sample.
[0044] Fig.11 Comparison chart of prediction results for the 6th test sample.
[0045] Fig.12 Comparison chart of prediction results for the 7th test sample.
[0046] Fig.13 Comparison chart of prediction results for the 8th test sample. DETAILED DESCRIPTION
[0047] The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present invention, but should not be construed as limiting the present invention.
[0048] The embodiments of the present invention are described in detail below with reference to the accompanying drawings.
[0049] A method for quickly predicting turbine passage inlet temperature based on Fourier neural operator comprises the following steps:
[0050] Step 1: Obtain the data set and first establish the turbine channel model Figure 1 As shown, it contains one stationary blade and two moving blades, the rotation angle is 15 degrees, and the downstream section is selected at the stationary blade channel outlet. The specific channel geometric parameters include: the circumferential spacing of the stationary blades is 55.48mm, the axial chord length is 34.56mm, the blade height is 34,64mm, and the number of blades per circle is 26. Samples with different inlet temperature distributions are obtained by designing different inlet hot spot positions. In order to determine the size of the hot spot and the range of temperature variation, the center position of the hot spot is divided, and the radius and angle directions are divided according to the shape of the inlet fan-shaped section to obtain different turbine inlet temperature distribution conditions. Among them, the center positions of the hot spots of the 40 samples in the training set are obtained by arithmetic division according to the radius and angle, and the 10 samples of the validation set and the test set are randomly generated. The inlet hot spot position is as shown in Figure 2 As shown in the figure, the inlet temperature distribution matrix is obtained according to the size, temperature and center position of the hot spot, which contains the temperature information of different points and the output data of different samples. The set inlet conditions are input into the numerical simulation calculation model. Based on the RANS numerical calculation method and the automatic calculation process of the script, the downstream section temperature calculation results are obtained as shown in the figure. Figure 3 shown.
[0051] Step 2: If Figure 6-Figure 13 The first subgraph extracts the 6*6 scattered sparse lattice data of the downstream cross-section temperature data uniformly distributed as input, and the inlet cross-section temperature distribution as output.
[0052] Step 3: Set the total number of samples to 50, read the input and output data partition labels of step 2, and divide them into training set, test set and validation set. The number of training sets is 40 randomly shuffled, the number of test sets is 8, and the number of validation sets is 2.
[0053] Step 4: Figure 4 The FNO model framework is shown in the figure. Inspired by the Green function, it constructs the following operator space expression:
[0054] υ t (x) = σ(W*υ t-1 (x)+∫ D k(x,y,a(x),a(y))*υ x (dy))
[0055] where v t It represents the inlet temperature distribution data of different iteration steps, σ represents the activation function, W represents the linear transformation part of the prediction value between two adjacent iteration steps, k represents the Fourier transform and inverse transform process, and a represents the scattered point data of the downstream turbine channel.
[0056] The representation of the output data mainly consists of two parts, the linear transformation part and the nonlinear transformation part. The linear transformation part is controlled by the parameter W; the nonlinear transformation part first transforms the time domain data into the frequency domain data by Fourier transform. The specific expression is as follows:
[0057] F[f j (K)]=∫ D f j (K)e -2iπ<x，K> dx
[0058] Where F represents the Fourier transform function, f represents the value to be transformed, K represents the convolution kernel, and e -2iπ<x，K> Represents a complex function. Since the input data is a 50*50 matrix, 50 frequency domain modes are obtained after Fourier transform. The first five high-frequency modes and the last five low-frequency modes are saved as effective modes, and other modes are ignored. The effective modes can effectively transform the integral and differential transformation of the operator space into a product calculation and fit after linear transformation. The modes after linear transformation are returned to the time domain space through inverse Fourier transform. The specific formula is as follows:
[0059] F -1 [f j (K)]=∫ D f j (K)e -2iπ<x,K> dx
[0060] where F -1 represents the inverse Fourier transform function, f represents the value to be transformed, K represents the convolution kernel, e-2iπ<xz，K> Represents a complex function. Repeating the above process multiple times is to perform Fourier operator iterative calculation. This paper sets the number of iterations to 3. Import input and output data, standardize them, and randomly shuffle them into the input model framework for calculation. Use a small batch stochastic gradient descent algorithm and set the batch to 4. Calculate the model error through forward propagation; update the model parameters through back propagation until the training reaches the specified number of cycles or the error is less than the set value and then stop updating. The single training time is about 0.12 seconds, and the model basically converges after about 100 cycles. Figure 5 It is the loss graph of the training set and the validation set during the training process.
[0061] Step 5: Input the input set data of the validation set into the training model, and use the output set data of the validation set to verify the model to obtain the prediction model.
[0062] Step 6: Input the input data of the test set into the prediction model to obtain the prediction results of the turbine channel inlet temperature distribution. Figure 6-13 As shown. The first sub-graph is the input scatter points, the second sub-graph is the model prediction result, the third sub-graph is the actual temperature distribution, and the fourth result is the absolute error between the predicted result and the actual result.
[0063] The present invention rapidly predicts the turbine inlet temperature distribution through a small amount of measurement point data in the downstream section of the turbine passage, thereby achieving a high-efficiency, high-precision, and high-resolution description of the turbine inlet temperature distribution.
Claims
1. A method for rapid prediction of turbine channel inlet temperature based on Fourier neural operator, characterized in that: The following steps are involved: Step 1: Obtaining a data set, including establishing a turbine channel model, obtaining samples of different inlet temperature distributions by designing different inlet hot spot positions, and obtaining a downstream outlet section temperature distribution data set by CFD numerical simulation calculation; Step 2: Extract 6*6 scattered lattice data under the downstream section temperature data as input, and the inlet section temperature distribution as output; Step 3: Set the total number of samples, read the input and output data partition labels of step 2, and divide them into training set, test set and validation set; Step 4: Based on the FNO neural network architecture, the input set data in the training set in step 3 is transferred to the training data in the FNO model, and the output set data in the training set in step 3 is transferred to the loss function for training, error calculation and updating of training parameters; Step 5: Input the input set data of the validation set into the training model, and use the output set data of the validation set to verify the model to obtain the prediction model; Step 6: Input the input set data of the test set into the prediction model to obtain the prediction results of the turbine channel inlet temperature distribution.
2. The method for rapid prediction of turbine channel inlet temperature based on Fourier neural operator according to claim 1 is characterized in that: The specific steps for obtaining the data set in step 1 include: Step 1-1: Determine the size of the hot spot and the temperature variation range, divide the center position of the hot spot, and perform arithmetic division in the radius and angle direction according to the shape of the inlet sector section to obtain different turbine inlet temperature distribution conditions; Step 1-2: Obtain the inlet temperature distribution matrix based on the size, temperature and center position of the hot spot, including the temperature information of different points, and obtain the output data of different samples; Step 1-3: Input the set inlet conditions into the numerical simulation calculation model, and obtain the calculation results based on the RANS numerical calculation method and the automatic calculation process using the script; Step 1-4: Extract the temperature distribution results of the downstream section and output them as a txt file. Use the interpolation algorithm to interpolate the data containing x, y, z and temperature into a 50*50 matrix to obtain the input data of different samples.
3. The method for rapid prediction of turbine channel inlet temperature based on Fourier neural operator according to claim 1 is characterized in that: The specific steps of training in step 4 include: Step 4-1: Build the FNO model framework and construct the following operator space expression: v t (x)=σ(W * v t-1 (x)+∫ D k(x,y,a(x),a(y))*v x (of)) Where vt represents the inlet temperature distribution data of different iteration steps, σ represents the activation function, W represents the linear transformation part of the prediction value between two adjacent iteration steps, k represents the Fourier transform and inverse transform process, and a represents the scattered point data of the downstream turbine channel; The representation of the output data consists of two parts, a linear transformation part and a nonlinear transformation part; the linear transformation part is controlled by the parameter W; the nonlinear transformation part first transforms the time domain data into frequency domain data by Fourier transform, and the specific expression is as follows: F[f j (K)]=∫ D f j (K)e -2iπ＜x，K＞ dx Where F represents the Fourier transform function, f represents the value to be transformed, K represents the convolution kernel, and e -2iπ＜x，K＞ Represents a complex function; since the input data is a 50*50 matrix, 50 frequency domain modes are obtained after Fourier transform, and the first five high-frequency modes and the last five low-frequency modes are saved as effective modes, and other modes are ignored; the effective modes are transformed from the integral and differential transformation of the operator space into product calculations through linear transformation, and then fitted; the modes after linear transformation are returned to the time domain space through inverse Fourier transform, and the specific formula is as follows: F -1 [f j (K)]=∫ D f j (K)e -2iπ＜x，K> dx Among them, F -1 represents the inverse Fourier transform function, f represents the value to be transformed, K represents the convolution kernel, e -2iπ＜x，K> represents a complex function; the above process is repeated multiple times to perform Fourier operator iterative calculation. In this paper, the number of iterations is set to 3; Step 4-2: Import input and output data, standardize them, and randomly shuffle them into the model framework for calculation; Step 4-3: Use the mini-batch stochastic gradient descent algorithm, set the batch size to 4, calculate the model error through forward propagation, and update the model parameters through back propagation until the training reaches the specified number of cycles or the error is less than the set value.