Multi-dimensional prediction optimization digital twin system for turbine performance of gas turbine

Abstract

The invention discloses a gas turbine turbine performance multi-dimensional prediction optimization digital twin system, and relates to the field of gas turbine turbine performance prediction, and the system comprises a data acquisition and classification module which is used for acquiring initial data and removing and classifying the initial data to obtain sample data; the data processing module is used for identifying, analyzing and processing the data to obtain result data; and the result generation and output module is used for outputting a multi-dimensional prediction report of the turbine performance of the gas turbine. According to the multi-dimensional prediction optimization digital twin system for the turbine performance of the gas turbine, the turbine performance of the gas turbine is predicted by fusing the neural network in a digital twin mode, so that the performance analysis of the gas turbine is comprehensively supported, the calculation time is greatly shortened, and the design iteration efficiency is remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of gas turbine turbine performance prediction, and in particular to a gas turbine turbine performance multi-dimensional prediction and optimization digital twin system. Background Art

[0002] Gas turbines are key equipment in the fields of modern aviation and energy. Their efficient operation depends on accurate prediction and optimized design of internal airflow. Although traditional CFD (computational fluid dynamics) simulation methods are accurate, they are computationally expensive and time-consuming. In recent years, fast prediction models based on deep learning have gradually become a research hotspot. They can significantly reduce computing costs and provide efficient solutions for complex performance prediction.

[0003] This study uses deep learning technology to propose a comprehensive prediction method for gas turbine multi-dimensional performance parameters, including zero-dimensional performance parameter prediction, blade performance test parameter prediction along the blade height, two-dimensional flow field prediction, and blade surface parameter prediction. This method aims to provide more efficient and accurate support for gas turbine design and optimization.

[0004] In order to solve these problems, a digital twin system for multi-dimensional prediction and optimization of gas turbine turbine performance is urgently needed. Summary of the invention

[0005] To solve the above problems, this application proposes a digital twin system for multi-dimensional prediction and optimization of gas turbine turbine performance, including: Data acquisition and classification module: The data acquisition and classification module is used to acquire initial data, eliminate the initial data, and classify it to obtain sample data; Data analysis and prediction module: The data analysis and prediction module is used to identify and analyze the data to obtain result data; Result generation output module: the result generation output module is used to output a multi-dimensional prediction report on gas turbine turbine performance; Preferably, the data acquisition and classification module includes a data extraction module and a data classification module: The data extraction module is used to obtain the overall design information of the gas turbine, and extract the overall design parameters and initial data from the overall design information of the gas turbine; The data classification module is used to analyze the initial data, remove abnormal data and identify missing data, and perform secondary data extraction based on the missing data to obtain sample data; Divide the sample data according to the sample data type and set the identifier to obtain the sample classification group; The sample classification group includes basic modeling data, three-dimensional blade design data, single blade design data, and two-dimensional flow channel coordinates.

[0006] Preferably, the data analysis and prediction module includes a physical twin component, a fusion framework, and a learning prediction network; The physical twin component is used to generate a gas turbine turbine physical model, and the specific contents of generating the gas turbine turbine physical model include: Acquire sample data and identify the identifier to obtain basic modeling data, and obtain the component structure type, installation position, connection relationship and structural parameters of the gas turbine turbine according to the basic modeling data; Generate virtual component blocks according to component structure types and structural parameter ratios, and arrange them according to installation positions and connection relationships to obtain a gas turbine turbine physical model; The fusion framework is used to integrate and coordinate the work of physical twin components and learning prediction networks; The learning prediction network is used to train sample data and perform multi-dimensional prediction on gas turbine turbine performance.

[0007] Preferably, the learning prediction network includes a zero-dimensional performance prediction unit, a blade performance prediction unit along the blade height direction, a two-dimensional flow field performance prediction unit, and a blade surface performance prediction unit; The zero-dimensional performance prediction unit is used to obtain three-dimensional blade design data and use a tree model and a fully connected neural network NN to predict the zero-dimensional performance parameters of the gas turbine, wherein the zero-dimensional performance parameters of the gas turbine include: power, flow, total static efficiency, total efficiency and reaction degree; The tree model is used to perform significance analysis on zero-dimensional parameters, and the similarity and distribution of three-dimensional blade design data are measured according to the significant parameters to extract the first third of the parameters as the input of the fully connected neural network NN; The fully connected neural network NN includes 3 hidden layers, an activation function ReLU, an epoch of 1000, an optimizer using Adam, and the output of the fully connected neural network NN is the performance parameters of the gas turbine.

[0008] Preferably, the blade performance prediction unit along the blade height direction adopts a neural operator network DeepONet to predict the performance test parameters of the blade along the blade height direction, and the performance test parameters of the blade along the blade height direction include reaction degree, total pressure loss of stationary blades and total pressure loss of moving blades; The blade performance prediction unit along the blade height direction adopts a custom DeepONet model, and the custom DeepONet model includes a branch network and a main network. The branch network input is the design data of the single blade, the main network input is the position x along the blade height direction, and the output is the performance test parameters along the blade height direction.

[0009] Preferably, the two-dimensional flow field performance prediction unit uses a U-Net network to predict the two-dimensional flow field parameters of the gas turbine blade, and the two-dimensional flow field parameters include a pressure field, an axial velocity field, a circumferential velocity field, a Mach number field and a temperature field; The U-Net network model inputs two-dimensional flow channel coordinates and outputs two-dimensional flow field parameters.

[0010] Preferably, the blade surface performance prediction unit adopts a neural operator network DeepONet to predict the parameters of the gas turbine's moving and stationary blade surfaces, and the parameters of the gas turbine's moving and stationary blade surfaces include surface temperature and surface pressure.

[0011] Preferably, the fusion framework adopts the Modulus framework.

[0012] In summary, the digital twin system for multi-dimensional prediction and optimization of gas turbine turbine performance of the present invention greatly reduces the calculation time and significantly improves the design iteration efficiency compared with traditional CFD simulation. By optimizing the network architecture and input feature selection, the error of the prediction result is controlled within an acceptable range. It covers zero-dimensional, two-dimensional and blade surface multi-dimensional performance prediction, and fully supports gas turbine performance analysis.

[0013] The technical method of the present invention is further described in detail below through the drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 It is a digital twin prediction platform for the multi-dimensional performance of the gas turbine turbine of the present invention; Figure 2 This is the error diagram of the total pressure loss of the stationary blades of the present invention; Figure 3 This is a diagram of the total pressure loss error of the moving blades of the present invention; Figure 4 This is a comparison chart of the pressure field results of the present invention. Figure 4 (a) is the real pressure field at the blade root. Figure 4 (b) is the predicted pressure field at the blade root. Figure 4 (c) is the pressure field error distribution at the blade root; Figure 5 This is a comparison diagram of the temperature field results of the present invention. Figure 5 (a) is the real temperature field in the middle of the leaf. Figure 5 (b) is the real temperature field in the middle of the leaf. Figure 5 (c) in the figure is the temperature field error distribution at the middle of the leaf; Figure 6 This is a comparison chart of the Mach number field results of the present invention. Figure 6 (a) is the real Mach number field at the blade tip. Figure 6 (b) is the predicted Mach number field at the blade tip. Figure 6 (c) is the Mach number error distribution at the blade tip; Figure 7 This is the streamline diagram of the stator blade of the present invention. Figure 7 (a) is the real streamline diagram at the root of the stationary blade. Figure 7 (b) is the predicted streamline diagram at the root of the stationary blade; Figure 8 This is the streamline diagram of the moving blade of the present invention. Figure 8 (a) is the real streamline diagram at the middle of the moving blade. Figure 8 (b) is the predicted streamline diagram at the middle of the moving blade; Fig. 9 is the prediction result of the surface temperature of the stationary blade of the present invention, Fig. 9 (a) is the actual temperature distribution of the three-dimensional moving blade. Fig. 9 (b) is the predicted temperature distribution of the three-dimensional moving blade. Fig. 9 (c) is the three-dimensional moving blade temperature error distribution; Fig.10 is the prediction result of the stationary blade surface pressure of the present invention, Fig.10 (a) is the actual pressure distribution of the three-dimensional stationary blade. Fig.10 (b) is the predicted pressure distribution of the three-dimensional stationary blade. Fig.10 (c) is the three-dimensional static blade pressure error distribution; Fig.11 It is a schematic diagram of a physical information neural network of a digital twin prediction platform for multi-dimensional performance of a gas turbine turbine according to the present invention; Fig.12 It is a schematic diagram of the autoencoder convolutional neural network of the digital twin prediction platform for the multi-dimensional performance of the gas turbine turbine of the present invention; Fig.13 Schematic diagram of the deep operator neural network of the digital twin prediction platform for the multi-dimensional performance of the gas turbine turbine of the present invention. DETAILED DESCRIPTION

[0015] The technical method of the present invention is further described below by means of the accompanying drawings and embodiments. It should be noted that unless otherwise specifically stated, the relative arrangement of the components and steps, numerical expressions and numerical values ​​described in these embodiments do not limit the scope of the present application.

[0016] The following description of at least one exemplary embodiment is merely illustrative in nature and is in no way intended to limit the present application, its application, or uses.

[0017] Technologies, systems, and devices known to ordinary technicians in the relevant art may not be discussed in detail, but where appropriate, the technologies, systems, and devices should be considered part of the specification.

[0018] In all examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limiting. Therefore, other examples of the exemplary embodiments may have different values.

[0019] Unless otherwise defined, technical or scientific terms used in the present invention shall have the common meanings understood by one having ordinary skills in the field to which the present invention belongs.

[0020] like Figure 1 , Fig.11 , Fig.12 , Fig.13 As shown, the digital twin system for multi-dimensional prediction and optimization of gas turbine turbine performance proposed in this application includes: Data acquisition and classification module: The data acquisition and classification module is used to acquire initial data, eliminate the initial data, and classify it to obtain sample data; Preferably, the data acquisition and classification module includes a data extraction module and a data classification module: The data extraction module is used to obtain the overall design information of the gas turbine, and extract the overall design parameters and initial data from the overall design information of the gas turbine; The data classification module is used to analyze the initial data, remove abnormal data and identify missing data, and perform secondary data extraction based on the missing data to obtain sample data; Divide the sample data according to the sample data type and set the identifier to obtain the sample classification group; The sample classification group includes basic modeling data, three-dimensional blade design data, single blade design data, and two-dimensional flow channel coordinates.

[0021] Data analysis and prediction module: The data analysis and prediction module is used to identify and analyze the data to obtain result data; Preferably, the data analysis and prediction module includes a physical twin component, a fusion framework, and a learning prediction network; The physical twin component is used to generate a gas turbine turbine physical model, and the specific contents of generating the gas turbine turbine physical model include: Acquire sample data and identify the identifier to obtain basic modeling data, and obtain the component structure type, installation position, connection relationship and structural parameters of the gas turbine turbine according to the basic modeling data; Generate virtual component blocks according to component structure types and structural parameter ratios, and arrange them according to installation positions and connection relationships to obtain a gas turbine turbine physical model; The fusion framework is used to integrate and coordinate the work of physical twin components and learning prediction networks; The learning prediction network is used to train sample data and perform multi-dimensional prediction on gas turbine turbine performance.

[0022] Preferably, the learning prediction network includes a zero-dimensional performance prediction unit, a blade performance prediction unit along the blade height direction, a two-dimensional flow field performance prediction unit, and a blade surface performance prediction unit; The zero-dimensional performance prediction unit is used to obtain three-dimensional blade design data and use a tree model and a fully connected neural network NN to predict the zero-dimensional performance parameters of the gas turbine, wherein the zero-dimensional performance parameters of the gas turbine include: power, flow, total static efficiency, total efficiency and reaction degree; The tree model is used to perform significance analysis on zero-dimensional parameters, and the similarity and distribution of three-dimensional blade design data are measured according to the significant parameters to extract the first third of the parameters as the input of the fully connected neural network NN; In this implementation, 104 three-dimensional blade design data are used. However, since 104 three-dimensional blade design data are too much, a sensitivity analysis is first performed, and a tree model is used to perform significance analysis on zero-dimensional parameters. The similarity and distribution of samples are measured based on significant parameters, and the first 35 parameters are extracted as the input of the NN neural network model.

[0023] The fully connected neural network NN includes 3 hidden layers, an activation function ReLU, an epoch of 1000, an optimizer using Adam, and the output of the fully connected neural network NN is the performance parameters of the gas turbine.

[0024] This embodiment tests the Deepxde library and the custom DeepONet based on Pytorch respectively, but the training result of the Deepxde library is not as good as that of the custom DeepONet based on Pytorch. Therefore, the present invention uses a custom DeepONet model and establishes a DeepONet model based on pytorch. The working conditions and geometric parameters are input into the branch network, and the prediction point coordinate information is input into the trunk network to predict the aerodynamic information.

[0025] Preferably, the blade performance prediction unit along the blade height direction adopts a neural operator network DeepONet to predict the performance test parameters of the blade along the blade height direction, and the performance test parameters of the blade along the blade height direction include reaction degree, total pressure loss of stationary blades and total pressure loss of moving blades; The specific implementation steps of this embodiment are as follows: 1. The dataset contains 1500 sets of data, divided into 1200 training sets and 300 test sets.

[0026] 2. Optimize feature selection through error analysis, taking the 35 most significant factors as input.

[0027] 3. Train the prediction model separately for each performance parameter to ensure prediction accuracy.

[0028] 4. To avoid the impact of magnitude differences between different parameters, the input data is regularized.

[0029] As shown in Table 1, the average relative errors of power, flow rate, etc. are all controlled within 6%.

[0030] Table 1 Detailed data of power, flow, etc. in zero dimension, including the number of neurons, learning rate, average relative error, and removal of values ​​greater than 100 ;

[0031] The blade performance prediction unit along the blade height direction adopts a custom DeepONet model, and the custom DeepONet model includes a branch network and a main network. The branch network input is the design data of the single blade, the main network input is the position x along the blade height direction, and the output is the performance test parameters along the blade height direction.

[0032] In this example, the single leaf design data is set to 104. One-dimensional (trunk and branch layers have the same number of neurons) PLCR has 6 hidden layers with 1600 neurons per layer, PLCS has 4 hidden layers with 1200 neurons per layer, Omega has 5 hidden layers with 1200 neurons per layer, the activation function is ReLU, all learning rates are the same at 0.0005epoch5000, and the optimizer uses Adam.

[0033] DeepONet combines a branch network (for input parameters, such as design parameters, angle of attack, and Mach number) with a stem network (for position or spatial coordinates). In this way, the branch network can learn the relationship between input parameters and physical quantities, and the stem network can learn the influence of spatial distribution. This separation structure can more effectively handle complex input-output relationships, while traditional neural networks (such as multi-layer perceptrons and convolutional neural networks) often find it difficult to model complex physical relationships in this multi-dimensional space.

[0034] Traditional neural networks may encounter problems such as unstable training or overfitting when the input dimensions are too large. DeepONet avoids this problem by separately training the branch network and the main network, and can generalize better, so it has strong processing capabilities for high-dimensional inputs.

[0035] The prediction noise level of DeepONet is significantly lower than that of the traditional CNN network and Fourier neural operator network. In terms of detailed prediction, DeepONet has higher prediction accuracy.

[0036] DeepONet can achieve good generalization performance with a small amount of training data, which is especially important for practical engineering problems.

[0037] In gas turbine and aircraft airfoil design, CFD simulation requires a lot of computing resources and time to generate enough training data, while DeepONet can be trained well with limited data and reduce the need for a large number of samples. DeepONet has strong learning ability with small samples. Other neural network methods usually require a large training data set to ensure good prediction performance, while DeepONet can effectively learn from limited data and provide accurate predictions through its unique structure.

[0038] The blade performance prediction unit data set of this embodiment includes 1500 sets of data, which are divided into 1200 sets of training sets and 300 sets of test sets. Figure 2 and Figure 3 As shown, the average absolute error of the reaction degree is 0.055, the average absolute error of the total pressure loss of the stationary blades is 0.052, and the average absolute error of the total pressure loss of the moving blades is 15.389.

[0039] Preferably, the two-dimensional flow field performance prediction unit uses a U-Net network to predict the two-dimensional flow field parameters of the gas turbine blade, and the two-dimensional flow field parameters include a pressure field, an axial velocity field, a circumferential velocity field, a Mach number field and a temperature field; The U-Net network model inputs two-dimensional flow channel coordinates and outputs two-dimensional flow field parameters.

[0040] The number of input channels of the U-Net network model is 1 (grayscale image), the number of output channels is 2 (for segmentation), the convolution kernel size is 4x4 for downsampling, 3x3 for upsampling, and 4x4 for the last layer. The stride is 2 for downsampling and upsampling, 1 for most layers, ReLU for activation, batch normalization after each convolution, bicubic interpolation for upsampling, 2 for magnification, ReduceLROnPlateau for learning rate scheduler, 0.5 for factor, 20 for patience, 10 for cooldown, all initial learning rates are the same for 0.5epoch10000, and Adam is used as optimizer.

[0041] In this embodiment, the U-Net network model randomly samples 3,000 coordinate points as reference coordinates; or 1,500 coordinate points for each of the moving and stationary blades.

[0042] Specific implementation method of this embodiment: 1. The number of data sampling points increased from 3000 to 6000 to improve accuracy.

[0043] 2. Network parameters such as the number of neurons and learning rate were optimized for different positions (root, middle, and top).

[0044] like Figure 4-Figure 6 As shown, the average relative error of the pressure field is about 5%, the average relative error of the temperature field is about 1.7%, and the average absolute error of the Mach number field is 0.07.

[0045] Preferably, the blade surface performance prediction unit adopts a neural operator network DeepONet to predict the parameters of the gas turbine's moving and stationary blade surfaces, and the parameters of the gas turbine's moving and stationary blade surfaces include surface temperature and surface pressure.

[0046] In this embodiment, the blade surface performance prediction unit is one-dimensional (the number of neurons in the trunk and branch layers is the same).

[0047] Among them, the temperature prediction unit has 3 hidden layers, 20 neurons in the first 2 layers, and 240 neurons in the third layer; Pressure prediction unit, 20 neurons in the first 2 layers, 240 neurons in the third layer. Activation function: ReLU, all learning rates are the same at 0.001epoch30000, and the optimizer uses Adam.

[0048] like Figure 7-10 As shown, the average relative error of the stationary blade surface temperature is 1.73%, and the average relative error of the stationary blade surface pressure is 2.8%.

[0049] Preferably, the fusion framework adopts the Modulus framework.

[0050] Result generation output module: The result generation output module is used to output a multi-dimensional prediction report on gas turbine turbine performance.

[0051] Finally, it should be noted that the above embodiments are only used to illustrate the technical method of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical method of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical method to deviate from the spirit and scope of the technical method of the present invention.

Claims

1. A digital twin system for multi-dimensional prediction and optimization of gas turbine turbine performance, characterized by: include: Data acquisition and classification module: The data acquisition and classification module is used to acquire initial data, eliminate the initial data, and classify it to obtain sample data; Data analysis and prediction module: The data analysis and prediction module is used to identify and analyze the data to obtain result data. The data analysis and prediction module includes a physical twin component, a fusion framework, and a learning prediction network; Result generation output module: the result generation output module is used to output a multi-dimensional prediction report on gas turbine turbine performance; The learning prediction network includes a zero-dimensional performance prediction unit, a blade performance prediction unit along the blade height direction, a two-dimensional flow field performance prediction unit, and a blade surface performance prediction unit; The zero-dimensional performance prediction unit is used to obtain three-dimensional blade design data and use a tree model and a fully connected neural network NN to predict the zero-dimensional performance parameters of the gas turbine, wherein the zero-dimensional performance parameters of the gas turbine include: power, flow, total static efficiency, total efficiency and reaction degree; The tree model is used to perform significance analysis on zero-dimensional parameters, and the similarity and distribution of three-dimensional blade design data are measured according to the significant parameters to extract the first third of the parameters as the input of the fully connected neural network NN; The fully connected neural network NN includes 3 hidden layers, an activation function ReLU, an epoch of 1000, an optimizer using Adam, and the output of the fully connected neural network NN is the performance parameters of the gas turbine; The blade performance prediction unit along the blade height direction adopts the neural operator network DeepONet to predict the performance test parameters of the blade along the blade height direction. The performance test parameters of the blade along the blade height direction include reaction degree, total pressure loss of stationary blades and total pressure loss of moving blades.

2. The gas turbine turbine performance multi-dimensional prediction and optimization digital twin system according to claim 1 is characterized in that: The data acquisition and classification module includes a data extraction module and a data classification module: The data extraction module is used to obtain the overall design information of the gas turbine, and extract the overall design information of the gas turbine to obtain the overall design parameters and initial data; The data classification module is used to analyze the initial data, remove abnormal data and identify missing data, and perform secondary data extraction based on the missing data to obtain sample data; Divide the sample data according to the sample data type and set the identifier to obtain the sample classification group; The sample classification group includes basic modeling data, three-dimensional blade design data, single blade design data, and two-dimensional flow channel coordinates.

3. The gas turbine turbine performance multi-dimensional prediction and optimization digital twin system according to claim 2 is characterized in that: The physical twin component is used to generate a gas turbine turbine physical model, and the specific contents of generating the gas turbine turbine physical model include: Acquire sample data and identify the identifier to obtain basic modeling data, and obtain the component structure type, installation position, connection relationship and structural parameters of the gas turbine turbine according to the basic modeling data; Generate virtual component blocks according to component structure types and structural parameter ratios, and arrange them according to installation positions and connection relationships to obtain a gas turbine turbine physical model; The fusion framework is used to integrate and coordinate the work of physical twin components and learning prediction networks; The learning prediction network is used to train sample data and perform multi-dimensional prediction on gas turbine turbine performance.

4. The gas turbine turbine performance multi-dimensional prediction and optimization digital twin system according to claim 1 is characterized in that: The blade performance prediction unit along the blade height direction adopts a custom DeepONet model, and the custom DeepONet model includes a branch network and a main network. The branch network input is the design data of the single blade, the main network input is the position x along the blade height direction, and the output is the performance test parameters along the blade height direction.

5. The gas turbine turbine performance multi-dimensional prediction and optimization digital twin system according to claim 1 is characterized in that: The two-dimensional flow field performance prediction unit uses a U-Net network to predict the two-dimensional flow field parameters of the gas turbine blade, and the two-dimensional flow field parameters include a pressure field, an axial velocity field, a circumferential velocity field, a Mach number field, and a temperature field; The U-Net network model inputs two-dimensional flow channel coordinates and outputs two-dimensional flow field parameters.

6. The gas turbine turbine performance multi-dimensional prediction and optimization digital twin system according to claim 1 is characterized in that: The blade surface performance prediction unit adopts the neural operator network DeepONet to predict the parameters of the gas turbine's moving and stationary blade surfaces, and the parameters of the gas turbine's moving and stationary blade surfaces include surface temperature and surface pressure.

7. The gas turbine turbine performance multi-dimensional prediction and optimization digital twin system according to claim 3 is characterized in that: The fusion framework adopts the Modulus framework.

Images