Turbine temperature distribution prediction method based on deep learning

Abstract

A turbine temperature distribution prediction method based on deep learning is used for evolution prediction of gas turbine transient experiment solid domain temperature distribution along with time, and comprises the following steps: constructing a turbine cascade experiment geometric model, generating a turbine cascade calculation grid, and processing according to the turbine cascade geometric model to obtain a turbine cascade test piece; the method comprises the following steps: constructing a gas turbine transient experiment device, determining transient experiment working conditions and boundary conditions of a turbine cascade, and constructing a recurrent neural network coded by a transient heat conduction equation; a gas turbine transient experiment is carried out, main flow inlet total temperature time sequence data, turbine blade grid test piece surface temperature time sequence data and material thermophysical parameters of a turbine blade grid test piece are collected to form experiment data, the recurrent neural network is trained, and inversion of experiment working conditions is completed; and utilizing the trained recurrent neural network to predict the temperature distribution in the whole stable operation stage of the solid domain of the turbine cascade. The transient experiment time can be shortened, and the transient experiment cost can be remarkably reduced.

Description

Technical Field

[0001] The present invention belongs to the technical field of thermal protection of aircraft engine turbines, and in particular relates to a turbine temperature distribution prediction method based on deep learning, which is used for predicting the evolution of solid domain temperature distribution over time in a gas turbine transient experiment. Background Art

[0002] In gas turbine design, accurate prediction of operating life is a major challenge. With the increase of heat load, thermal stress fatigue has attracted more and more attention. The most important requirement for introducing thermal stress considerations in high-pressure turbine heat transfer design is to accurately predict the metal temperature distribution of the turbine blade grid. However, current high-temperature measurement technology can only measure the surface temperature of the turbine blade grid or the temperature of a limited number of discrete points, and cannot predict the temperature field of the entire solid domain. This undoubtedly makes it difficult to accurately evaluate the life of the turbine. In order to meet this demand and obtain the time-series evolution data of the temperature field of the entire solid domain, researchers have proposed a series of methods, but these methods have the following advantages and disadvantages:

[0003] (1) The heat transfer problem between the turbine blades and the mainstream is a typical third-class boundary condition problem. However, the heat transfer coefficient, recovery temperature and adiabatic temperature during the operation of the turbine are unknown quantities, and the third-class thermal boundary conditions cannot be used directly to achieve the time series prediction of the temperature field of the entire solid domain. For this reason, researchers usually use CFD for three-dimensional simulation, that is, to construct a fluid domain and solid domain calculation model so that it has inlet and outlet boundary conditions and geometric conditions consistent with the experimental conditions, and then simulate and solve the calculation model through the conjugate heat transfer model. However, due to the complex flow structure inside the turbine blade, if a low-precision turbulence model, such as the RANS / URANS model, is used, the flow structure cannot be accurately predicted, which in turn affects the accuracy of the solid domain temperature field prediction structure; and if a high-precision turbulence model, such as the SAS / LES model, is used, although the prediction accuracy can be improved, its high computational cost limits the use of researchers.

[0004] (2) In order to avoid the problem of turbulence model selection in CFD, some researchers calculate the surface heat flux density or surface convective heat transfer coefficient through the surface temperature time series data, and then input the surface heat flux density or surface convective heat transfer coefficient as the boundary condition into the CFD simulation calculation. Among them, the LRM method proposed by Virginia Tech and the IRM method proposed by Oxford University are typical representatives for calculating the surface heat flux density or surface convective heat transfer coefficient and are widely used. However, the disadvantage of the above method is that the time series temperature measurement accuracy of the turbine surface temperature is high. The measurement uncertainty of 2K may even lead to a 50% deviation in the surface heat flux density or surface convective heat transfer coefficient, which will greatly affect the prediction of the solid domain temperature field.

[0005] (3) The LRM method proposed by Virginia Tech and the IRM method proposed by Oxford University have high restrictions on solid thermal conductivity and data processing window period. For this reason, Northwestern Polytechnical University proposed a method for obtaining the convective heat transfer coefficient based on the numerical solution of transient solid thermal conductivity. This method can significantly reduce the restrictions on thermal properties such as solid thermal conductivity. However, in its calculation process, the blade surface temperature is directly used as the first-class boundary condition for calculating the temperature field, which still puts forward high requirements on the blade surface temperature measurement technology, namely the time resolution and measurement accuracy of the surface blade. And because of the use of the first-class thermal boundary condition, this method cannot be extrapolated, that is, if the full-cycle temperature field evolution is required, a full-cycle experiment needs to be conducted and the full-cycle surface temperature needs to be provided, which undoubtedly limits the use of this method. Summary of the invention

[0006] In order to overcome the shortcomings of the above-mentioned prior art, the purpose of the present invention is to provide a turbine temperature distribution prediction method based on deep learning. As a means of processing gas turbine experimental data, this method effectively realizes the prediction of the time series changes of the temperature distribution inside the solid domain of the turbine, greatly reduces the cost of experimental equipment and the difficulty of experimental operation, and obtains more accurate experimental data, which is more in line with the needs of gas turbine experimenters.

[0007] In order to achieve the above object, the technical solution adopted by the present invention is:

[0008] A method for predicting turbine temperature distribution based on deep learning, comprising the following steps:

[0009] Step 1, constructing a turbine blade cascade experimental geometric model, generating a turbine blade cascade computational grid, and processing a turbine blade cascade test piece according to the turbine blade cascade geometric model;

[0010] Step 2, constructing a gas turbine transient test device, wherein a turbine blade cascade test piece is installed in a wind tunnel test section of the gas turbine transient test device, and the gas turbine transient test device is used to simulate the cooling characteristics and heat transfer characteristics of the turbine blade cascade when the upstream incoming flow instantaneously reaches a certain operating condition;

[0011] Step 3, determining the transient experimental working conditions and boundary conditions of the turbine blade grid, and constructing a recurrent neural network encoding the transient heat conduction equation in combination with the turbine blade grid calculation grid;

[0012] Step 4, conduct a gas turbine transient experiment, collect the mainstream inlet total temperature time series data, the turbine blade test piece surface temperature time series data and the material thermophysical property parameters of the turbine blade test piece, and form the experimental data;

[0013] Step 5, using the experimental data to train the recurrent neural network to complete the inversion of the experimental conditions;

[0014] Step 6: Use the trained recurrent neural network to predict the temperature distribution of the solid domain of the turbine blade grid during all stages of stable operation.

[0015] Compared with the prior art, the present invention has the following beneficial effects:

[0016] (1) The third type of thermal boundary conditions are directly inverted by the turbine blade surface temperature and the mainstream inlet boundary conditions. It is applicable to any turbine blade, avoiding the influence of complex flow prediction accuracy in turbine blade experiments.

[0017] (2) The present invention couples the concept of full three-dimensional solid thermal conductivity, making it applicable to blade test pieces processed from any material, eliminating restrictions on the thermal properties of the processed materials.

[0018] (3) The present invention integrates the concept of linear regression, which greatly reduces the requirements for the time resolution and measurement accuracy of temperature measurement technology, reduces the experimental cost, and improves the experimental accuracy.

[0019] (4) The present invention can effectively realize the inversion of the third type of thermal boundary conditions, which not only reduces the limitation of the window period during experimental operation, but also can realize the extrapolation prediction of the full stable operation cycle, thus reducing the operation time of the experiment.

[0020] (5) The present invention can achieve the inversion of experimental conditions within a few minutes and quickly predict the evolution within a set prediction period. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 It is a flow chart of a turbine temperature distribution prediction method based on deep learning of the present invention.

[0022] Figure 2 The invention is a gas turbine transient test device according to a preferred embodiment of the present invention.

[0023] Figure 3 The present invention is a method for operating an experimental device according to a preferred embodiment of the present invention.

[0024] Figure 4 It is a calculation model of a preferred embodiment of the present invention.

[0025] Figure 5 It is a computing grid of a preferred embodiment of the present invention.

[0026] Figure 6 It is a schematic diagram of the hard coding method adopted by the present invention.

[0027] Figure 7 It is a schematic diagram of a recurrent neural network encoding the transient heat conduction equation of the present invention.

[0028] Figure 8It is a heat transfer coefficient prediction deviation of a preferred embodiment of the present invention.

[0029] Fig. 9 It is a solid domain temperature distribution prediction effect diagram of a preferred embodiment of the present invention. DETAILED DESCRIPTION

[0030] The embodiments of the present invention are described in detail below with reference to the accompanying drawings and examples.

[0031] The present invention provides a method for predicting turbine temperature distribution based on deep learning, which has the function of inverting turbine operating condition parameters and can be used as an effective data processing tool for turbine blade row experiments; and the ability to predict the temperature distribution of the solid domain of the turbine blade row can shorten the transient experiment time and significantly reduce the transient experiment cost. Figure 1 As shown, it mainly includes the following steps:

[0032] Step 1: construct a turbine blade cascade experimental geometric model and generate a turbine blade cascade calculation grid, and obtain a turbine blade cascade test piece based on the turbine blade cascade geometric model.

[0033] In this step, a three-dimensional modeling software is used to construct a turbine blade cascade experimental geometric model. The specific operations are as follows:

[0034] A turbine blade cascade geometric model is generated through 3D modeling instructions based on the turbine blade surface data, a connection structure geometric model is generated based on the assembly structure required by the wind tunnel test section of the gas turbine transient test device, and the turbine blade cascade geometric model and the connection structure geometric model are merged through combined instructions of 3D modeling software to obtain a turbine blade cascade experimental geometric model.

[0035] In this step, the turbine blade cascade manufacturing material (such as metal material) is processed by additive manufacturing or casting to obtain a turbine blade cascade test piece.

[0036] Among them, the turbine blade manufacturing material needs to have complete thermophysical parameters, that is, its specific heat capacity Cp, density ρ and thermal conductivity λ are known. If it is a normal temperature experiment (i.e. the mainstream temperature is 0℃-100℃), the effect of temperature on thermophysical properties can be ignored and the above thermophysical parameters are set to constant values. If it is a medium-high temperature experiment, the relationship between the above physical parameters and temperature is still required.

[0037] On this basis, the turbine blade cascade experimental geometric model is meshed using 3D computational mesh generation software to obtain the turbine blade cascade computational mesh, which simulates the role of the turbine blade cascade test piece in the gas turbine transient experiment, namely the solid domain part.

[0038] Step 2, construct a gas turbine transient test device, wherein the turbine blade test piece is installed in the wind tunnel test section of the gas turbine transient test device. The gas turbine transient test device is used to simulate the cooling characteristics and heat transfer characteristics of the turbine blade when the upstream flow instantly reaches a certain operating condition, and can measure the surface temperature of the turbine blade test piece and the total temperature and total pressure of the mainstream during the process.

[0039] The gas turbine transient experimental device and the experimental device operation method of the present invention can refer to patents CN112414739A, CN112683943A and CN118443350A, etc. The device can complete the gas turbine transient experiment and measure, collect and record the time series data of the mainstream total temperature and pressure and the turbine surface temperature.

[0040] Figure 2 A preferred structure of a gas turbine transient test device is shown, which mainly includes a wind tunnel, a cold air supply system, a total temperature probe, a total pressure probe, a transient infrared thermal imager, an infrared window and a collection system, etc.; the wind tunnel includes an air intake section, a test section and an exhaust section, wherein the test section is the core part of the wind tunnel, and the turbine blade test piece is installed in the wind tunnel test section; the cold air supply system is only installed when simulating the cooling characteristics experiment of the turbine blade (i.e., cooling experiment), and does not need to be installed when simulating the heat transfer characteristics experiment of the turbine blade (i.e., heat transfer experiment), and the cold air supply system is connected to the test section at the bottom interface of the turbine blade test piece; the total temperature probe and the total pressure probe are installed in the test section and located upstream of the turbine blade test piece, and are used to measure the total temperature and total pressure of the mainstream; the infrared window is installed in the wind tunnel test section at a position directly opposite to the turbine blade test piece; the transient infrared thermal imager measures the surface temperature of the turbine blade test piece through the infrared window; the collection system collects and records the data measured by the total temperature probe, the total pressure probe and the transient infrared thermal imager.

[0041] Step 3: Determine the transient experimental conditions and boundary conditions of the turbine blade grid, and construct a recurrent neural network encoding the transient heat conduction equation in combination with the turbine blade grid calculation grid.

[0042] The transient experimental conditions of the turbine blade grid described in this step mainly include: the inlet and outlet Mach number, total temperature, total pressure, cold air inlet mass flow rate and cold air inlet total temperature of the turbine blade grid. This experimental condition is used to set the mainstream flow parameters, cold air flow parameters of the gas turbine transient experimental device, and the construction of mainstream, cold air and boundary information in the recurrent neural network. It should be noted that the cold air flow parameters (inlet mass flow rate and inlet total temperature) are only required when the gas turbine experimental device is used for cooling experiments, and do not need to be set for heat transfer experiments.

[0043] In this step, the turbine blade grid calculation grid is used to construct geometric information in the recurrent neural network, that is, the calculation domain of the recurrent neural network is set as the solid domain part through the turbine blade grid calculation grid, so that the recurrent neural network calculation domain accurately simulates the turbine blade grid test piece, thereby achieving the purpose of predicting the temperature distribution inside the turbine blade grid test piece, that is, the solid domain; the mainstream and boundary information of the recurrent neural network are set by a hard coding method to make it consistent with the experimental working conditions and boundary conditions, thereby achieving the purpose of accurately inverting the experimental working conditions.

[0044] The recurrent neural network constructed in this step includes multiple recurrent units, and the recurrent units of different time steps are connected by recurrent operators to realize the recurrent advancement of the temperature distribution time series prediction inside the turbine blade test piece, i.e., the solid domain. Each recurrent unit includes an exponential operator module and a CNe operator module, or includes an exponential operator module, a CNe operator module, and an LNe operator module.

[0045] Among them, the exponential operator module is used to calculate the temperature change of any spatial position x in the solid domain within the time step h; the CNe operator module and the LNe operator module are used to calculate the influence of the temperature of the spatial position x' on the temperature change at the spatial position x. The difference is that when the LNe operator module is further used on the basis of the CNe operator module, higher accuracy can be achieved.

[0046] The construction of the recurrent neural network in this step only considers the turbine blade test piece (i.e., solid domain) as the calculation domain, that is, the inversion of the boundary conditions of the turbine blade test piece (i.e., solid domain) under the experimental conditions is realized through the collected mainstream inlet total temperature, cold air inlet total temperature (only during cooling experiment) and turbine blade test piece surface temperature. Therefore, the transient heat conduction equation is encoded into the recurrent neural network construction, where the transient heat conduction equation can be selected according to the three-dimensional effect of the turbine blade test piece in the gas turbine transient experiment, that is, one-dimensional transient heat conduction, two-dimensional transient heat conduction and three-dimensional transient heat conduction.

[0047] The transient heat conduction equation is expressed as follows:

[0048]

[0049] Where T represents the solid domain temperature of the turbine blade, t represents time, and α represents the diffusion coefficient;

[0050] The CNe operator module is used to discretize formula (1) in time and space, and r is introduced i and A i Expressed as a simplified equation, the resulting equation is as follows:

[0051]

[0052] Where, T in+1 represents the temperature at the i-th grid point at the n+1th time, T i n represents the temperature at the ith grid point at the nth time, represents the temperature of the adjacent grid points of the ith grid point at the nth time (excluding the ith grid point), h is the time step, r i is a commonly used definition of grid ratio, A i is a compact form of information about neighboring grid points.

[0053] According to the calculation accuracy requirement, the result calculated by the CNe operator module is used as the calculation result or the predicted value of the calculation result;

[0054] When there is a higher requirement for calculation accuracy, the LNe operator module is used to discretize formula (1) in time and space to obtain:

[0055]

[0056] Where, T i pred It is the predicted value of the calculation result obtained by the CNe operator module or the LNe operator module last time. It is a compact form of prediction information of adjacent grid points, that is, prediction-correction is realized through the LNe operator module, thereby achieving higher calculation accuracy.

[0057] Both CNe and LNe algorithms are in explicit time format, which is conducive to parallel computing. For the convenience of programming, the present invention decomposes CNe and LNe algorithms into exponential operator modules (exp Conv), CNe operator modules (CNe Block) and LNe operator modules (LNe Block) according to their working methods. There can be multiple CNe operator modules and LNe operator modules in each loop unit. More CNe operator modules and LNe operator modules mean higher calculation accuracy and greater calculation cost (slower calculation speed and memory / video memory space required for calculation).

[0058] The hard coding method of the present invention refers to processing the boundary conditions by introducing virtual nodes (i.e., ghost nodes). The process utilizes the padding function of the convolution kernel of the convolutional neural network to fill a corresponding amount of grid padding outside the boundary of the calculation domain according to the spatial discrete format requirements. The filled value is the result calculated by the boundary node algebraic equation method or the virtual boundary method according to the actual physical boundary conditions and the virtual node spatial discrete format.

[0059] Among them, the third type of boundary conditions (Robin boundary conditions) are used for the fluid-solid interface where the mainstream contacts the turbine blade test piece, and the first type (isothermal) or second type (adiabatic) boundary conditions are used for the surfaces of other turbine blade test pieces; when the mainstream is a stable flow, the heat transfer coefficient of its fluid-solid interface is a specific constant, and the linear regression idea is introduced into the third type of boundary conditions, which is rewritten as:

[0060]

[0061] Where q represents the heat flux density of fluid-solid interface convection heat transfer, T t represents the mainstream temperature, T w represents the wall temperature, T r represents the recovery temperature, h represents the convective heat transfer coefficient of the fluid-solid interface, which is an unknown variable and is obtained through neural network training; due to the introduction of the linear regression idea, formula (4) can be regarded as T t -T w The linear equation with x as the independent variable and y as the dependent variable is obtained by linear regression. The slope a of the linear equation is the convective heat transfer coefficient h, and the intercept b is h(T r -T t ).

[0062] Step 4: Conduct a gas turbine transient experiment to collect the mainstream inlet total temperature time series data, the turbine blade test piece surface temperature time series data, and the material thermophysical property parameters of the turbine blade test piece to form the experimental data.

[0063] For the specific operation of this step, please refer to Figure 3 , the mainstream inlet total pressure is adjusted by the flow control device of the wind tunnel, and the mainstream inlet total temperature is adjusted by the temperature control device of the wind tunnel to achieve the predetermined inlet conditions respectively. If it is a cooling experiment, the flow meter and temperature control device of the cold air supply system are also required to adjust the cold air inlet mass flow and the cold air inlet temperature respectively to achieve the predetermined cold air flow conditions respectively; during the experiment, the mainstream inlet total pressure is guaranteed, that is, the mainstream reaches a stable flow state. If it is a cooling experiment, the stability of the above-mentioned cold air flow parameters must also be guaranteed; the acquisition system collects and records the data measured by the total temperature probe, the total pressure probe and the transient infrared thermal imager, among which the inlet total pressure is used to timely adjust the mainstream or cold air inlet flow through negative feedback to ensure the flow stability requirements; the mainstream inlet total temperature (cold air inlet total temperature, if it is a cooling experiment) and the turbine blade surface temperature time series data measured by the transient infrared thermal imager are used for training the constructed transient heat conduction equation encoded recurrent neural network.

[0064] Step 5, using the obtained experimental data to train the recurrent neural network to complete the inversion of the experimental conditions.

[0065] The fluid-solid interface where the mainstream of the present invention contacts the turbine blade test piece adopts the third type of boundary conditions (Robin boundary conditions), and the surfaces of other turbine blade test pieces can adopt the first type (isothermal) or second type (adiabatic) boundary conditions. The first type (isothermal) and second type (adiabatic) boundary conditions are both known boundary conditions, while the convective heat transfer coefficient h of the fluid-solid interface in the third type of boundary conditions is an unknown number, that is, it is set as the training parameter of the recurrent neural network encoded by the transient heat conduction equation. The surface temperature time series data of the turbine blade test piece obtained in step 4 is taken as the true value, and the surface temperature result predicted based on the guessed value h' is taken as the predicted value, and the mean square error (MSE, Mean Square Error) is used to measure the deviation between the predicted value and the true value.

[0066]

[0067] Where N is the number of measurement points of the fluid-solid interface surface temperature, T s,i is the surface temperature of the ith measurement point, f(h i ') is a recurrent neural network encoding the transient heat conduction equation based on the guessed value h i 'Predicted surface temperature results.

[0068] The Adam optimizer is used to update the guessed value h' of the convective heat transfer coefficient to make it close to the true value h of the convective heat transfer coefficient. The learning rate during training is set to 2, and the equal-interval learning rate adjustment strategy StepLR is adopted. In StepLR, the adjustment interval is set to 100, the adjustment factor is set to 0.9, and the training cycle is 1000 epochs.

[0069] Through the above operations, the training of the recurrent neural network encoded by the transient heat conduction equation can be completed. At this time, the obtained recurrent neural network can accurately represent the boundary conditions of the blade test piece (i.e., the solid domain) in the gas turbine transient experiment, that is, the experimental working condition inversion of the gas turbine transient experiment is completed.

[0070] Step 6: Use the trained recurrent neural network to predict the temperature distribution of the solid domain of the turbine blade grid during all stages of stable operation.

[0071] By accurately describing the boundary conditions of the blade test piece in step 5, the prediction of the temperature distribution in the solid domain of the turbine blade at any time and the evolution of the temperature distribution in the solid domain of the turbine blade during the stable operation cycle can be obtained by using the recurrent unit of the recurrent neural network encoded by the transient heat conduction equation.

[0072] In a preferred embodiment of the present invention, in order to demonstrate the accuracy of the method, a CFD example is used to analogize the actual experimental operation process, and the high-precision prediction results of CFD are used as true values ​​to illustrate the present invention.

[0073] In the CFD software, select the CFD turbulence model to simulate the turbine heat transfer / cooling experiment, set the CFD data recording format, data recording range and data recording duration. The above operations are similar to the construction of the gas turbine transient experimental device and the settings of data measurement, collection and recording.

[0074] The calculation model and calculation grid used in this embodiment are as follows Figure 4 , Figure 5 As shown. The calculation model is an outward-swept flat plate fluid-solid coupling model, and the length of the fluid-solid coupling area is 50 mm and the thickness is 5 mm. The solid domain material of this model is Haynes 230 high-temperature alloy. The grid is divided by computational grid generation software, and a uniform orthogonal structured grid scheme is adopted, and its grid unit size is: Δx=Δy=Δz=0.25mm. The initial state of this embodiment is the inlet Mach number Ma=0.2, and the inlet total temperature is room temperature 300K. Then the flow condition changes transiently, that is, the inlet total temperature jumps to 800K.

[0075] The uniform orthogonal structured grid divided in the above operation is used to construct the computational domain of the recurrent neural network encoded by the transient heat conduction equation. The structured grid divided represents the internal area and geometric shape of the computational domain.

[0076] Through the convolution kernel padding function of the convolutional neural network, virtual nodes (i.e., ghost nodes) are generated for the structured grid. The third type of boundary conditions are used for the fluid-solid interface with the mainstream, and the mainstream total temperature information is encoded into the boundary conditions. The schematic diagram is shown in the figure. Figure 6 It should be noted that when the mainstream is a stable flow, the heat transfer coefficient of the fluid-solid interface is a specific constant. For this reason, the linear regression idea is introduced into the third type of boundary conditions and rewritten:

[0077]

[0078] Where q represents the heat flux density of fluid-solid interface convection heat transfer, h represents the fluid-solid interface convection heat transfer coefficient, T t represents the mainstream temperature, T w represents the wall temperature, T r represents the recovery temperature. h is an unknown variable and needs to be obtained through neural network training. In this embodiment, the total temperature of the mainstream inlet is 800K and is substituted into the above formula.

[0079] For other solid domain surfaces, the second type of boundary conditions is used to encode the adiabatic information into the boundary conditions, that is, the filling value is the temperature at the center of the first layer of grid on the boundary, as shown in the schematic diagram. Figure 6 As shown in (a).

[0080] The recurrent neural network encoded by the transient heat conduction equation connects the recurrent units through the recurrent operator to realize the recurrent advancement of the time series prediction of the temperature distribution in the solid domain of the turbine blade. The convolution values ​​of the index module, CNe module and LNe module are calculated based on the divided uniform orthogonal structured grid to construct the recurrent unit. After testing, the use of 1 index operator, 1 CNe operator module and 4 LNe operator modules in this embodiment can achieve the best calculation accuracy and efficiency. The schematic diagram is shown in FIG. Figure 7 shown.

[0081] Haysen 230 high temperature alloy is used as the solid domain material of the model in this embodiment. The initial operating conditions are the inlet Mach number of 0.2 and the inlet total temperature of 300K. When the conditions of flow stability and solid domain thermal balance are met, the inlet total temperature is adjusted to 800K in a stepwise manner, and other flow conditions are kept unchanged. The surface temperature of the solid domain at the fluid-solid interface is measured, collected and recorded with a time step of h = 0.01s.

[0082] As mentioned above, h is an unknown variable and needs to be obtained through neural network training. In this embodiment, according to the high-precision calculation results of CFD, the mean value of the heat transfer coefficient can be obtained. To this end, the initial value of h is assigned to U(400,600), and the guessed value h' is obtained. The experimental operation interval is predicted with h = 0.01s as the time step to obtain the predicted solid domain surface temperature time series data T w,pred The collected solid domain surface temperature time series data T w As the true value, the mean square error (MSE) is used to measure the predicted value T w,pred With the true value T w deviation.

[0083]

[0084] Where N is the number of measurement points of the fluid-solid interface surface temperature, T w,i is the surface temperature of the ith measurement point, T w,pred,i The recurrent neural network encoding the transient heat conduction equation is based on the guess value h i 'Predicted surface temperature results.

[0085] The Adam optimizer is used to update the guessed value h' of the convective heat transfer coefficient to make it close to the true value h of the convective heat transfer coefficient. The learning rate during training is set to 2, and the equal-interval learning rate adjustment strategy StepLR is adopted. In StepLR, the adjustment interval is set to 100, the adjustment factor is set to 0.9, and the training cycle is 1000 epochs.

[0086] Through the above operations, the training of the recurrent neural network encoded by the transient heat conduction equation can be completed. At this time, the trained recurrent neural network encoded by the transient heat conduction equation can accurately represent the boundary conditions of the blade cascade test piece (i.e., the solid domain) in the conducted gas turbine transient experiment, that is, the inversion of the experimental conditions of the conducted gas turbine transient experiment is completed. Figure 8 The deviation distribution between the predicted heat transfer coefficient h' and the true value h is shown after the training of the recurrent neural network encoding the transient heat conduction equation is completed through the above operation. Figure 8 It can be seen that the heat transfer coefficient deviation inverted by the recurrent neural network encoded by the trained transient heat conduction equation is in the range of (-5,5), which can accurately invert the actual operating conditions.

[0087] The recurrent neural network encoded by the trained transient heat conduction equation is used to predict the solid domain temperature distribution at t = 10s with a time step of h = 0.01s. Fig. 9 The figure shows the comparison between the CFD high-precision calculation results and the trained recurrent neural network encoded by the transient heat conduction equation. Fig. 9 It can be seen that the recurrent neural network encoded by the trained transient heat conduction equation can accurately predict the temperature distribution inside the solid domain, and the maximum deviation is only 0.07K, which proves the accuracy of the present invention in predicting the temperature distribution inside the solid domain.

Claims

1. A turbine temperature distribution prediction method based on deep learning, characterized in that: The following steps are involved: Step 1, constructing a turbine blade cascade experimental geometric model, generating a turbine blade cascade computational grid, and processing a turbine blade cascade test piece according to the turbine blade cascade geometric model; Step 2, constructing a gas turbine transient test device, wherein a turbine blade cascade test piece is installed in a wind tunnel test section of the gas turbine transient test device, and the gas turbine transient test device is used to simulate the cooling characteristics and heat transfer characteristics of the turbine blade cascade when the upstream incoming flow instantaneously reaches a certain operating condition; Step 3, determining the transient experimental working conditions and boundary conditions of the turbine blade grid, and constructing a recurrent neural network encoding the transient heat conduction equation in combination with the turbine blade grid calculation grid; Step 4, conduct a gas turbine transient experiment, collect the mainstream inlet total temperature time series data, the turbine blade test piece surface temperature time series data and the material thermophysical property parameters of the turbine blade test piece, and form the experimental data; Step 5, using the experimental data to train the recurrent neural network to complete the inversion of the experimental conditions; Step 6: Use the trained recurrent neural network to predict the temperature distribution of the solid domain of the turbine blade grid during all stages of stable operation.

2. The turbine temperature distribution prediction method based on deep learning according to claim 1 is characterized in that: The step 1 is to construct a turbine blade cascade experimental geometric model using a three-dimensional modeling software, and to obtain a turbine blade cascade test piece by processing the turbine blade cascade manufacturing material through additive manufacturing or casting; The turbine blade cascade experimental geometric model is meshed using three-dimensional computational mesh generation software to obtain the turbine blade cascade computational mesh, which simulates the solid domain part of the turbine blade cascade test piece in the gas turbine transient experiment.

3. The turbine temperature distribution prediction method based on deep learning according to claim 1 is characterized in that: The gas turbine transient test device includes a wind tunnel, a cold air supply system, a total temperature probe, a total pressure probe, a transient infrared thermal imager, an infrared window and a collection system; the wind tunnel includes an air intake section, a test section and an exhaust section, and the turbine blade test piece is installed in the wind tunnel test section; the cold air supply system is only installed when simulating the cooling characteristics experiment of the turbine blade, and does not need to be installed when simulating the heat transfer characteristics experiment of the turbine blade, and the cold air supply system is connected to the test section at the bottom interface of the turbine blade test piece; the total temperature probe and the total pressure probe are installed in the test section and located upstream of the turbine blade test piece, and are used to measure the total temperature and total pressure of the mainstream; the infrared window is installed in the wind tunnel test section at a position directly opposite to the turbine blade test piece; the transient infrared thermal imager measures the surface temperature of the turbine blade test piece through the infrared window; the collection system collects and records the data measured by the total temperature probe, the total pressure probe and the transient infrared thermal imager.

4. The turbine temperature distribution prediction method based on deep learning according to claim 1 is characterized in that: The transient experimental working conditions of the turbine blade grid include: the turbine blade grid inlet and outlet Mach number, total temperature, total pressure, cold air inlet mass flow rate and cold air inlet total temperature; The experimental conditions are used to set mainstream flow parameters, cold air flow parameters of a gas turbine transient experimental device and to construct mainstream, cold air and boundary information in a recurrent neural network.

5. The turbine temperature distribution prediction method based on deep learning according to claim 1 is characterized in that: In the step 3, the computational domain of the recurrent neural network is set as the solid domain part through the turbine blade grid computational grid, so that the computational domain of the recurrent neural network accurately simulates the turbine blade grid test piece, thereby achieving the purpose of predicting the temperature distribution inside the turbine blade grid test piece, i.e., the solid domain; the mainstream and boundary information of the recurrent neural network are set by a hard coding method to make them consistent with the experimental conditions and boundary conditions, thereby achieving the purpose of accurately inverting the experimental conditions.

6. The turbine temperature distribution prediction method based on deep learning according to claim 1 or 5, characterized in that: In the recurrent neural network, recurrent units of different time steps are connected by recurrent operators to realize the recurrent advancement of the time series prediction of the temperature distribution inside the turbine blade test piece, i.e., the solid domain; the recurrent unit includes an exponential operator module and a CNe operator module, or includes an exponential operator module, a CNe operator module and an LNe operator module; the exponential operator module is used to calculate the change of the temperature at any spatial position x in the solid domain within the time step h; the CNe operator module and the LNe operator module are used to calculate the influence of the temperature at the spatial position x' on the temperature change at the spatial position x.

7. The turbine temperature distribution prediction method based on deep learning according to claim 6 is characterized in that: The transient heat conduction equation is as follows: Where T represents the solid domain temperature of the turbine blade, t represents time, and α represents the diffusion coefficient; The CNe operator module is used to discretize formula (1) in time and space, and r is introduced i and A i Expressed as a simplified equation, the resulting equation is as follows: Where, T i n+1 represents the temperature at the i-th grid point at the n+1th time, T i n represents the temperature at the ith grid point at the nth time, represents the temperature of the adjacent grid points of the ith grid point at the nth time, h is the time step, r i is the grid ratio, A i It is a compact form of information about adjacent grid points; According to the calculation accuracy requirement, the result calculated by the CNe operator module is used as the calculation result or the predicted value of the calculation result; When there is a higher requirement for calculation accuracy, the LNe operator module is used to discretize formula (1) in time and space to obtain: Where, T i pred It is the predicted value of the calculation result obtained by the CNe operator module or the LNe operator module last time. A compact form of prediction information for neighboring grid points.

8. The turbine temperature distribution prediction method based on deep learning according to claim 5 is characterized in that: The hard coding method processes the boundary conditions by introducing virtual nodes (i.e., ghost nodes). The process utilizes the padding function of the convolution kernel of the convolutional neural network to fill a corresponding amount of grid padding outside the boundary of the computational domain according to the requirements of the spatial discrete format. The filled value is the result calculated by the boundary node algebraic equation method or the virtual boundary method according to the actual physical boundary conditions and the virtual node spatial discrete format. Among them, the third type of boundary conditions (Robin boundary conditions) are used for the fluid-solid interface where the mainstream contacts the turbine blade test piece, and the first type (isothermal) or second type (adiabatic) boundary conditions are used for the surfaces of other turbine blade test pieces; when the mainstream is a stable flow, the heat transfer coefficient of its fluid-solid interface is a specific constant, and the linear regression idea is introduced into the third type of boundary conditions, which is rewritten as: Where q represents the heat flux density of fluid-solid interface convection heat transfer, T t represents the mainstream temperature, T w represents the wall temperature, T r represents the recovery temperature, h represents the convective heat transfer coefficient of the fluid-solid interface, which is an unknown variable and is obtained through neural network training; Formula (4) is regarded as T t -T w The linear equation with x as the independent variable and y as the dependent variable is obtained by linear regression. The slope a of the linear equation is the convective heat transfer coefficient h, and the intercept b is h(T r -T t ).

9. The turbine temperature distribution prediction method based on deep learning according to claim 1 is characterized in that: In step 4, the inlet total pressure of the mainstream is adjusted by the flow control device of the wind tunnel, and the inlet total temperature of the mainstream is adjusted by the temperature control device of the wind tunnel, so that they respectively reach the predetermined inlet conditions. If it is a cooling experiment, the flow meter and temperature control device of the cold air supply system are also required to adjust the cold air inlet mass flow rate and the cold air inlet temperature respectively, so that they respectively reach the predetermined cold air flow conditions; during the experiment, the inlet total pressure of the mainstream is guaranteed, that is, the mainstream reaches a stable flow state. If it is a cooling experiment, the stability of the above-mentioned cold air flow parameters must also be guaranteed; the acquisition system collects and records the data measured by the total temperature probe, the total pressure probe and the transient infrared thermal imager, wherein the inlet total pressure is used to timely adjust the mainstream or cold air inlet flow through negative feedback to ensure the flow stability requirements; The mainstream inlet total temperature and the turbine blade surface temperature time series data measured by the transient infrared thermal imager are used to train the recurrent neural network encoded by the constructed transient heat conduction equation.

10. The turbine temperature distribution prediction method based on deep learning according to claim 1, characterized in that: In the step 5, the surface temperature time series data of the turbine blade test piece obtained in step 4 is taken as the true value, the surface temperature result predicted based on the guessed value h' is taken as the predicted value, and the Adam optimizer is used to update the guessed value h' of the convective heat transfer coefficient to make it close to the true value h of the convective heat transfer coefficient, thereby completing the training of the recurrent neural network.

Images